August 22, 2026

Four things this week. Money moved, prices moved, and a lot of copper moved between warehouses. None of it changed how much of anything actually exists.

$500 Million, Seven Projects, and a Capability That Doesn’t Exist Yet

The money is real. What it’s paying for is years out.

The Department of Energy awarded $500 million across seven projects on August 20, expanding domestic critical-mineral processing, battery manufacturing and recycling. Lilac Solutions, Jervois and Nth Cycle received $100 million each. Princeton NuEnergy, Arcanum Ventures, Elevated Materials and Coreshell Technologies received $50 million each. Energy Secretary Chris Wright’s framing was direct: America has depended on foreign actors for materials that underpin its economy and national security.

Worth being precise about what the money buys and when. Lilac’s facility on the Great Salt Lake targets 5,000 tonnes a year of battery-grade lithium carbonate by 2028, which would double current US lithium production, with a second phase taking it toward 20,000 tonnes. Jervois would build the only commercial cobalt refinery in the country, at an industrial site that has not yet been selected.

Worth knowing who’s building it. Jervois controls a large cobalt deposit in Idaho and was taken private last year through a pre-packaged bankruptcy caused by low cobalt prices. That isn’t disqualifying, and cobalt’s price collapse was an industry event rather than a company failure. It does mean the balance sheet behind the country’s only planned domestic cobalt refinery is twelve months out of restructuring.

The money exists today. The capability it’s paying for still has to be sited, built, commissioned and qualified before it counts, the same clock this note has tracked across every domestic buildout all summer.

For investors: this is not a near-term revenue catalyst for any of the seven names, publicly traded or not. The signal worth tracking is siting and construction pace against the 2028 targets, not the announcement. A funding commitment is not a production timeline, and the two have diverged before in this exact sector.

China’s Index Has Fallen for Three Weeks. Western Prices Move on Something Else.

268.4 to 259.1 on China’s index. The prices Western buyers pay are assessed separately, and the gap has been widening.

China’s Rare Earth Price Index has declined steadily since late July, from 268.4 on July 29 to 259.1 as of August 18, according to the Association of China Rare Earth Industry. That’s a drop of roughly 3.5%.

Read alone, that looks like the whole market easing. It isn’t, and the reason is structural rather than cyclical.

The index is a composite of domestic reference quotations collected by a Chinese industry association, not a market-clearing price. On its August 12 sheet, dysprosium oxide sat at 1,365 to 1,405 yuan per kilogram, about $205 to $211, and terbium oxide near $990 to $998.

Western prices exist, but they are assessed rather than exchange-cleared, thinly traded, and mostly behind subscription. The public anchor is a ratio rather than a level: Benchmark Mineral Intelligence calculated that dysprosium oxide delivered into North America ran 4.4 times the Chinese ex-works price across 2025, and forecasts that multiple reaching 8.3 times by 2027.

That last number is the one that matters. The two markets aren’t converging as Western capacity comes online. On Benchmark’s own forecast they diverge further, because the constraint outside China is qualified separation capacity rather than mined material, and that takes longer to build than a mine.

For investors: don’t let a softening Chinese domestic index read as an easing supply picture for Western-facing names. The relevant price for MP, USAR, Lynas and Energy Fuels is the assessed ex-China number, which is set by different constraints and has been widening against China, not closing. A Chinese domestic move is not, on its own, a reason to change a Western supply chain thesis in either direction.

When the chokepoint is in the conversion chemistry, not the mine, domestic Chinese price movements don’t travel West.

Copper’s Squeeze Broke This Week. The Deficit Didn’t.

The tightest spread since 2021 collapsed in four days. Not one tonne of new metal was mined.

Through Friday the 14th, copper inventories in London Metal Exchange warehouses had fallen for 42 consecutive sessions, down to 204,975 tonnes, the lowest since February. By Monday the premium for metal delivered immediately over metal delivered in three months reached about $545 a tonne, the widest since late 2021.

Then it came apart. Trafigura and other traders delivered more than 20,000 tonnes into LME warehouses on Tuesday, followed by a larger jump on Wednesday, the biggest single-day rise in readily available stock since 2024. By Thursday total inventories stood at 239,925 tonnes, up 17% from the prior Friday’s low. By Friday that $545 premium was roughly $76, and LME three-month copper closed at $14,111.50 a tonne, down 0.30% on the week and snapping a seven-week winning streak.

None of that metal was newly produced. It already existed, sitting in the wrong place. Traders had been shipping copper to the United States ahead of a tariff ruling that still hasn’t arrived, draining the warehouses where LME contracts settle, and it came back only when the premium for delivering to London finally beat the premium for holding in America. What broke was an arbitrage, not a shortage.

The supply picture never moved. Jefferies’ tracker covering miners responsible for roughly 55% of global supply shows second-quarter output down 3.9% year over year, with Ivanhoe down 43% after the Kamoa-Kakula disruption and Newmont down 53% on declining ore grades. Chile expects national production to fall 2.6% this year. Jefferies models a 442,000-tonne global deficit in 2026, widening to 782,000 by 2030. Even after this week’s rebuild, available stock is down roughly 75% from its April peak.

For investors: the week’s spread collapse says nothing about copper supply, and treating it as easing would be a mistake. It says the metal is mobile and the arbitrage is powerful. What determines where the metal sits next is a decision that still hasn’t come. Commerce recommended a phased tariff on refined copper imports of 15% from January 1, 2027, rising to 30% a year later, and the President has yet to rule on it. Societe Generale reads the current COMEX premium as implying only about a 14.6% chance the 15% duty lands on schedule. The market is barely pricing the single biggest catalyst in front of it. Watch the ruling and LME stock levels, not the spread.

The Stocks Swung Twice This Week. Neither Move Was About Rare Earths.

Down 3-4% Tuesday on unrelated AI sentiment. Up again Friday on a rotation that had nothing to do with the sector either.

MP Materials and USA Rare Earth fell Tuesday, MP down about 3% to $56.67 and USAR down about 4% to $18.51, dragged lower by a broad AI-capex sentiment reset that had nothing to do with either company. Reporting at the time explicitly noted no company-specific catalyst.

Friday ran the other way, and not because of anything in this note either. The entire resource complex caught a bid as thirty-year Treasury yields hit their highest since 2007 and capital rotated into hard assets. The rare earth ETF was up around 6% intraday, Critical Metals Corp around 10%.

By Friday’s close MP stood at $60.05 and USAR at $19.26, up about 6% and 4% respectively from Tuesday, and both finished within 1.2% of their session highs. The bid held into the bell. Neither name recovered because of anything that happened in rare earths.

For investors: two moves in opposite directions inside four sessions, one driven by an AI sentiment reset and one by a macro rotation into hard assets. Neither had any connection to the Stillwater ramp, the DOE awards, or next Friday’s Serra Verde vote. In a sector this thin, that is closer to the normal state than the exception, which is why the calendar items are worth more of your attention than the tape.

The Common Thread

Whether it’s half a billion dollars funding capability still years from existing, a Chinese index falling while the Western price moves on entirely different constraints, a copper squeeze that broke without a single tonne being mined, or two stock moves that had nothing to do with rare earths, the signal and the physical reality kept running on separate clocks this week.

Three Things I’ll Be Watching Next Week

USA Rare Earth’s shareholder vote on the Serra Verde acquisition, August 28, and whether the combined company says anything about yttrium, which sits outside the offtake entirely.

Whether LME copper inventories keep rebuilding or the spread widens again, which will tell you how much of this week was shorts covering and how much was a market genuinely short of deliverable metal.

Whether Jervois names a site for the cobalt refinery, which is the first real test of the 2028 timeline.

The Clock

November 10 - REE extractant controls resume.

November 27 - gallium truce expiry.

January 1 - DFARS 252.225-7052.

80 days to extractant controls. 97 days to the gallium truce. 132 days to DFARS.

Live critical mineral prices at scarceearth.io.

The full archive is at williamdavid.substack.com. The Week in Minerals publishes every Saturday.

Share

Financial data sourced from company earnings reports and public market data as of August 21, 2026. DOE awards sourced via Reuters (Ernest Scheyder and Jarrett Renshaw) and Metal Tech News, confirmed August 20-21, 2026. Jervois bankruptcy and private ownership sourced via Reuters, confirmed August 20, 2026. USA Rare Earth special meeting details sourced via the company’s definitive proxy statement filed with the SEC, confirmed August 22, 2026. China Rare Earth Price Index and domestic reference prices sourced via the Association of China Rare Earth Industry as reported by Rare Earth Exchanges, index readings July 29 through August 18, 2026 and oxide reference prices dated August 12, 2026; ACREI figures are industry reference data, not independently verified market-clearing prices. Ex-China price ratio sourced via Benchmark Mineral Intelligence Rare Earths Service, 2025 actual and 2027 forecast. LME copper inventories, spreads and closing prices sourced via Reuters, Bloomberg as reported by Mining.com, and Benchmark Mineral Intelligence, August 14-21, 2026; Section 232 refined copper tariff recommendation sourced via the White House proclamation of July 30, 2025 and the Federal Register; Societe Generale tariff probability via CNBC, August 14, 2026; Jefferies mine production tracker dated August 16, 2026; Chile production forecast via Cochilco. MP Materials and USA Rare Earth closing prices (Tuesday August 18: MP $56.67, USAR $18.51; Friday August 21: MP $60.05, USAR $19.26) confirmed against live terminal data; Friday sector move sourced via 24/7 Wall St.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

Before “computer” meant a machine, it meant a person.

In the early twentieth century, if a company or a government needed a hard calculation done - the trajectory of an artillery shell, the orbit of a planet - they didn’t reach for a machine. They hired a room full of people, often women, and gave each of them one small piece of a much larger sum. One person multiplied two numbers, wrote the result on a slip of paper, and passed it down the line. The next person did the next step. Hour after hour, a room full of clerks turned an impossibly large calculation into a thousand small, boring ones.

That’s compute. It’s arithmetic, repeated, at whatever scale the problem demands. It never stopped being that. What changed is we got extraordinarily better at building the room.

Why this should already feel like your problem

If you’ve never thought about compute before today, that’s a reasonable place to have been. It happens somewhere else, on machines you’ll never see, and it’s easy to assume it stays there.

It doesn’t anymore. The room full of clerks scaled into the single largest capital allocation decision happening anywhere in the world right now - trillions of dollars, and the decision about how much of it gets built runs through boardrooms, not physics labs. If you hold an index fund, a meaningful share of it is a bet on how much arithmetic capacity a handful of companies can build, finance, and keep running. If your electricity bill has moved this year, some of that is this too. You don’t have to work in AI, or even like it, for this to already be your problem. It already is - and understanding what compute actually is comes first.

From a room of people to a room of switches

Modern compute happens inside a chip: a piece of silicon covered in billions of tiny switches (transistors) that flip on and off, incredibly fast, to do arithmetic. One switch flipping is one tiny step in a calculation - not so different, in principle, from one clerk doing one multiplication. The difference is scale and speed. A modern chip can perform billions, and in many cases trillions, of these steps every second. And where the clerks’ room topped out at maybe a hundred people working side by side, a modern data center runs thousands of these chips together, wired so tightly that they behave as one enormous room of arithmetic rather than thousands of separate ones.

That rate - how many operations a chip can do in a second - is usually measured in something called FLOPS (floating-point operations per second; “floating-point” just means the kind of number the arithmetic is done with, and you can safely ignore the detail). You’ll see chips advertised by their FLOPS the way a car is advertised by its horsepower. It’s not the whole story of what a chip can do, but it’s the number everyone reaches for first, because it answers the most basic question: how much arithmetic can this thing do per second if you point it at a problem?

Why AI needs so much of it

Training an AI model - teaching it through a huge number of examples about how language works, or how to recognize an image - is millions of these small calculations adjusted and repeated, over and over, until the model gets better at the task. Running the model afterward (what’s called inference, generating an answer to something you’ve asked it) is the same kind of arithmetic, just less of it per question.

Here’s the part that explains why this accelerated so fast. Each new generation of AI model has needed dramatically more of this arithmetic than the one before it, largely because bigger models, trained on more examples, have kept getting meaningfully better at the task - with no clear ceiling yet in sight. Nobody fully predicted (even a few years ago) how steep that curve would turn out to be. That’s how a technology built on simple, repeated arithmetic became the reason companies are now spending more on infrastructure than almost anything else in the economy.

Two companion pieces to this one, the AI Hardware Primer and the AI Memory Primer, go deep on the chip itself and on the memory that has to sit beside it feeding it data fast enough to keep up. This piece isn’t trying to repeat that ground. What matters here is simpler: AI is, underneath everything, an enormous, sustained demand for exactly the kind of repeated arithmetic a room full of clerks used to do by hand - just at a scale no room, and until recently no machine, could actually deliver.

The turn: compute becoming a commodity in itself

Here’s where this gets interesting, and where most explanations of AI infrastructure stop short.

Once you can measure a rate of arithmetic reliably, and once a company can promise to deliver that rate to a customer for months or years at a time, that promise starts to behave like anything else you can meter, sell, and contract in advance: electricity, natural gas, freight capacity. Nobody’s shipping you the electrons; they’re selling you a guaranteed rate, sustained over time, and you pay for it whether or not you use every unit.

Compute has started working the same way. A company like CoreWeave doesn’t sell chips - it sells access to a rate of computation, under contract, for a set period. Think of it less like buying a car and more like reserving GPU-hours: a customer locks in a certain amount of computing capacity, at a certain level of performance, for a certain stretch of time, the way a factory might reserve electricity in advance, or a retailer might reserve cargo space on a ship. And once that contracted rate is reliable enough, it becomes something else again: an asset a lender will finance against, the same way a bank will lend against a signed, long-term power contract before a single kilowatt has been delivered.

That’s not theoretical. In 2023, CoreWeave borrowed against its GPUs at a floating rate around 15 percent - the kind of rate lenders charge when the collateral is unfamiliar and the risk is hard to price. By early 2026, a comparable facility, this time for $8.5 billion, secured directly against its own fleet of chips and a long-term customer contract, priced at roughly a third of that and carried an investment-grade rating. Same basic collateral. Three years apart. That compression is the market deciding this asset class is real. It’s part of a fast-growing category: JPMorgan thinks this kind of financing could reach $30 to $40 billion a year through 2027, up from about $27 billion in 2025. Worth being precise about what that number actually covers, though - it's not GPU-backed debt specifically, it's the broader category that debt sits inside.

None of that works, though, if the capacity sits unused. A reserved GPU-hour that nobody runs a workload on costs exactly as much to finance as one running at full tilt, and earns nothing. That’s utilization - how much of a company’s installed compute is actually doing work at any given moment - and it’s one of the quiet numbers everyone financing this pays close attention to. Installed capacity and productive capacity aren’t the same thing, and the whole commodity argument only holds up as long as that gap stays small. Owning the hardware and profiting from it turn out to be two different skills.

One catch is worth being upfront about. Compute isn’t a perfectly interchangeable commodity yet, not the way a barrel of oil is. Two hours on one company’s chips can produce meaningfully more, or less, useful work than two hours on another’s, depending on the memory attached, the network connecting the chips together, and how reliably the whole system stays running. The commodity is emerging underneath those differences. It isn’t fully standardized yet, and that gap is exactly why financing it is still being figured out in real time rather than settled.

None of that changes the direction of travel. Compute isn’t just running AI anymore. It’s beginning to behave like an asset class of its own, with its own financing, its own contracts, and, increasingly, its own market price.

One honest complication

All of this rests on an assumption worth naming plainly: that a chip generating value today will still be worth something close to that a year or two from now. That assumption isn’t settled, and serious, well-resourced people disagree about it.

It helps to separate two different clocks here, because coverage of this tends to blend them together. One is accounting useful life: how many years a company spreads a chip’s cost over on its own books, for accounting purposes. Meta uses five and a half years. Microsoft, Alphabet, and CoreWeave still use six. Amazon uses five - and its case is particularly interesting, because it’s the one hyperscaler that moved the opposite direction. It ran six years too, until 2025, when it shortened its own estimate and said why in its own filing: “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” That’s not outside commentary. That’s one company’s own accounting team formally taking a side in the argument this section is about. The other clock is economic life: how long that chip actually keeps earning real money before a newer, faster chip makes it uncompetitive. Those two clocks don’t have to agree, and the gap between them is exactly where the disagreement lives.

An older chip doesn’t become useless the moment a faster one ships, though. What determines whether it’s still worth running is the workload, not its age. A model that needs the newest chip to train doesn’t necessarily need it to run afterward, and a chip too slow for one job can be perfectly adequate for another. What actually decides this: whether the software still supports the older chip, whether it has enough memory for the job at hand, how much electricity it burns per unit of useful work, and how heavily it’s actually being used. That’s why older racks tend to migrate downward through the workload stack rather than getting scrapped, the way an old company car goes from the sales team racing between meetings to a less demanding job. Still useful, just not for the job it started with.

If a rack can stay useful for two or three more years this way, an operator can meet more of its growing compute needs without buying new hardware immediately - the installed base doesn’t vanish every time the next generation ships. But that only holds while the economics still make sense. Once electricity cost per unit of work gets too high, memory can’t fit the job anymore, the software stops being maintained, or the cost of keeping it running simply exceeds what its output is worth, it becomes a genuine write-off rather than a hand-me-down. Whether an old chip finds a home depends on the same kind of arithmetic the whole piece has been building toward: is this thing still doing useful work for what it costs to keep it running.

That tension is the real argument underneath the financing. Lenders put the resale value of a three-year-old chip anywhere from 10 percent to 60 percent of what it originally cost - an enormous range to build a loan around. Tiering pushes toward the higher end: a chip still doing useful work somewhere in the stack is worth something real. The gap between accounting life and economic life pushes toward the lower end: a chip can look fine on the books while the market has already decided nobody wants to pay full price for its output. Some of today’s strong demand for older chips may also be a temporary bridge rather than a permanent floor - the newest generation of chip can’t physically fit into most of the older data centers built to hold the last one, so the old chips are earning their keep by default, not necessarily because anyone still wants them for their own sake. Untangling how much of today’s resale value is genuine tiering demand versus that kind of bridge is exactly what lenders disagree about.

None of that means the reframing is wrong. It means the financing question isn’t really how long the books say a chip lasts. It’s what someone will still pay for that chip’s computing capacity once a newer generation exists - and right now, even the people financing this at scale don’t fully agree on the answer.

Why this matters, again

This has a practical edge, if you’re reading a company’s results with any of this in mind. “Exposure to AI” isn’t really a category worth much on its own. Two companies can both call themselves AI companies and mean completely different things by it - one owns compute that depreciates, competes on price, and needs constant reinvestment just to stay useful; another sells something scarce that goes into everyone else’s compute, with a backlog and no need to chase the next generation just to keep up. Only one of those companies is necessarily well positioned in what’s actually being built. And it’s worth remembering that even once compute genuinely does become commonplace, the scarcity behind it doesn’t disappear. It just moves - to memory, to power, to whatever the next layer down turns out to be the one nobody can build fast enough.

Compute used to be a cost of doing business: you bought the machine, you ran your calculation, the machine slowly became worth less. What’s happening now is different. A contracted, reliable rate of arithmetic has started to look, to the people who finance things for a living, like something closer to a commodity: metered, sold forward, borrowed against.

That shift is the thing sitting underneath almost everything written about AI infrastructure spending, financing, and the companies caught in the middle of it - even when it’s never said out loud. Understanding it doesn’t require an engineering background, or a finance background. It just requires seeing compute for what it actually is: a room full of arithmetic, scaled up past anything a room of people could ever do, that the market has started treating like something you can own a piece of.

Two companion pieces to this one, the AI Hardware Primer and the AI Memory Primer, go deep on the chip itself and on the memory that has to sit beside it feeding it data fast enough to keep up:

Financial data sourced from company disclosures and JPMorgan research as of August 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

The full archive is at williamdavid.substack.com.

Share

August 15, 2026

Five stories this week, one thread underneath: rules built to control other people end up binding the people who wrote them.

Malaysia and Indonesia Learn Their Own Rules Bind Them Too

Resource nationalism is not only a China problem, and this week made that plain twice.

Malaysia is assessing whether to ease its 2024 moratorium on raw rare earth exports. Deputy Natural Resources Minister Syed Ibrahim Syed Noh confirmed to Bloomberg that the government is under pressure from investors in the US, Australia, France, and India. “They are knocking at the doors already,” he said. But any resumption would carry conditions tied to inbound investment and technology transfer - the same instrument China has used against everyone else, deployed by a jurisdiction Western buyers have been quietly treating as the neutral alternative. No timeline has been given.

Indonesia found the same wall from a different angle. A regulatory gap over how much rare earth content is permissible in other mineral exports - alumina and nickel derivatives among them - left cargoes stuck at port. Jakarta resumed exports August 3 while it revises the underlying 2026 Trade Ministry regulation that created the gap in the first place.

Neither country intended to restrict trade. Both did, briefly, because a rule written to capture value at home didn’t anticipate how it would bind the people who wrote it. Positions built on Malaysian or Indonesian midstream capacity as the politically safe alternative to China carry more sovereign risk than the geography suggests. The instrument is the same. Only the flag on the building has changed.

Sigma Lithium’s Best Quarter Ever, and the Plant Isn’t Running

Record profitability, and the operation has been paused since mid-July. Both are true at once.

Sigma Lithium posted a 47% EBITDA margin this quarter, the highest in the company’s history, up from 39% in Q1. Net revenue hit a quarterly record of $55 million on 24,400 tonnes sold, with realized price up 17% to $2,089 per tonne. Production volume climbed 50% to 35,400 tonnes, and costs fell hard alongside it: plant gate cost down 36%, CIF down 33%. Net debt is down 25% year over year.

None of that is happening on the ground right now. Mining and plant operations at Sigma’s Grota do Cirilo complex in Brazil have been paused since mid-July, pending a TAC Agreement with the Minas Gerais state government that is still under negotiation. Tailings reprocessing continues; the mine and plant do not.

A position in Sigma right now is a bet on that agreement, not on lithium pricing or the margin print. The next move in the stock is more likely to be triggered by a TAC headline out of Minas Gerais than by anything in a future earnings release. Watch the state government, not the tape.

One Company, Four Structural Moves, and a New CEO Who Runs the Target

A company doesn’t install the acquisition target’s own CEO as its incoming chief unless it expects the deal to close.

USA Rare Earth moved on several fronts in a single week. Definitive funding documentation was signed with the US Department of Commerce. The company announced intent to acquire Serra Verde, the only scaled producer of all four magnetic rare earths outside Asia. It announced intent to invest in Carester, a French separation business. And Blacksburg, South Carolina was named as the company’s second US metal and magnet site.

The detail that says more than the rest combined: CEO Barbara Humpton will retire October 1, and her successor is Thras Moraitis, currently CEO of Serra Verde itself. Installing the target company’s own chief executive as your incoming CEO only makes sense if the deal is expected to close, and the timing points at the same window as the funding documentation.

For anyone weighing USAR, that leadership signal carries more information about deal-close probability than the “intent to acquire” language in the press release does on its own. Structure, not a print.

Five Tonnes

Against a 10,000-tonne ambition, five tonnes is not volume. It’s proof the route exists.

Evolution Metals took delivery of its first non-China NdPr metal shipment this month: five metric tonnes, sourced from SRE Vietnam via Japan’s Senri Trading, bound for the company’s South Korean magnet facilities. The company is targeting 10,000 tonnes a year of magnet output at those facilities, and it explicitly linked this shipment to the January 2027 DFARS sourcing deadline.

Five tonnes against a 10,000-tonne target is a rounding error stated honestly rather than dressed up. This isn’t a supply chain that’s arrived, it’s one that’s been proven to exist, once, at a scale too small to matter on its own.

What it de-risks is narrow but real: any thesis premised on the Vietnam-to-Korea route reaching DFARS-compliant scale by 2027 no longer has to assume the logistics work. It still has to assume the volume shows up. Sizing a position on this shipment alone would be sizing on the wrong variable.

The US Reserves Its Own Scrap, and Its Own Domestic Price Just Broke From the World’s

Starting August 27, American sellers of their own tungsten scrap can no longer sell it abroad. The tungsten market just showed why that might not be enough.

On August 6, the Bureau of Industry and Security published a rule under the Defense Production Act reserving all domestic sales of shredded tungsten scrap and lithium battery black mass for US buyers, effective August 27. The mechanism runs in reverse from the one this note usually tracks: the US has been exporting tungsten scrap to the Philippines, Taiwan, Vietnam, and South Korea for processing because it has no domestic tungsten mining and thin recycling capacity of its own. This rule locks the material inside the country before there’s the processing capacity built to use it. Commitment ahead of capability, just running the other way.

The tungsten market delivered a second, independent version of the same story this week. China's domestic APT price has fallen roughly 24.6% since a July peak, while the Western benchmark, assessed CIF Rotterdam, has stayed essentially flat near $3,075 per mtu. Fastmarkets responded by launching a dedicated China-domestic APT assessment on August 5, citing an arbitrage mechanism between domestic and international prices that, in its own words, “is becoming less effective.”

Treasury made a smaller, related move on August 7, welcoming new S&P Global reference prices for gallium, germanium, tungsten, antimony, NdPr, and praseodymium - a first step toward a Western benchmark that doesn’t run through China’s domestic market at all. Three institutions, ten days, one underlying fact: the pricing relationship that used to link Chinese and Western tungsten markets has stopped transmitting signals reliably in both directions.

The domestic-reservation rule protects US buyers from a supply squeeze, but it doesn’t create the processing capacity to use the scrap it locks in. Names with actual domestic tungsten recycling or recovery capacity are the ones positioned to benefit before 2027; scrap generators without a processing partner just have a captive buyer they didn’t have before.

The Common Thread

Whether it’s Malaysia and Indonesia, Sigma Lithium’s closed regulatory gate, or the US reserving its own scrap, the bottleneck this week wasn’t extraction. It was who controls the door - and increasingly, whether that’s even the same country running the mine.

Three Things I’ll Be Watching Next Week

Whether Malaysia’s export moratorium review produces an actual timeline, or stays open-ended pressure with no decision attached.

Terms of Sigma Lithium’s TAC Agreement with Minas Gerais, and whether operations resume before the next earnings cycle.

China’s MIIT H2 mining and smelting quota, still unpublished as of the most recent assessment. It was already the largest open catalyst carried over from June and July. It still is.

The Clock

November 10 - REE extractant controls resume.

November 27 - gallium truce expiry.

January 1 - DFARS 252.225-7052.

87 days to extractant controls. 104 days to the gallium truce. 139 days to DFARS.

Live critical mineral prices at scarceearth.io.

The full archive is at williamdavid.substack.com. The Week in Minerals publishes every Saturday.

Malaysia and Indonesia export policy sourced via Bloomberg, Reuters, and rare-earth-mining.com, confirmed August 14, 2026. Sigma Lithium Q2 2026 results sourced from company press release via TradingView, StockTitan, and Junior Mining Network, confirmed August 14, 2026. USA Rare Earth Q2 2026 results and leadership transition sourced from company press release via GlobeNewswire and SEC filings (Form DEFA14A), confirmed July 20 and August 10, 2026. Evolution Metals shipment sourced via rare-earth-mining.com, confirmed August 2026. BIS tungsten scrap and black mass rule sourced via Federal Register publication and Strategic Metals Weekly News Review, confirmed August 6, 2026. Tungsten APT pricing via SMM (metal.com) and Fastmarkets, benchmark dates July 30 and August 3, 2026. Treasury reference price announcement sourced via Asia Times, confirmed August 7, 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

On July 28th, Meta signed a data center deal in El Paso structured almost exactly like one it had signed nine months earlier in Louisiana. An outside investor owns 80%. Meta owns 20, occupies the whole campus, and guarantees the property’s value on a declining schedule running sixteen years. Debt behind the venture runs to 2048. Four-year leases, sixteen-year guarantees, debt to 2048, three different clocks, on one building.

A week later, Arista Networks reported its first-ever $3 billion quarter. Revenue up 37.7%. No debt to speak of. No vendor holding equity in exchange for a purchase commitment. No special purpose vehicle - a legal entity created just to hold one deal’s debt and assets separately from the parent company - backstopping anyone’s obligations. Just a company that sells switches AI infrastructure needs, at a price the market will pay, to customers who have nowhere else to go.

Both of these are AI infrastructure stories. Only one of them is a circular financing story. The interesting question isn’t which one is riskier. It’s why the distinction exists at all, and where, exactly, it stops applying.

Here’s the plain version of what that term actually means. A company sells something expensive. That same company also provides, guarantees, or arranges the financing the buyer uses to pay for it. If the buyer could never have afforded the purchase on its own, the seller’s own money is what’s actually holding the deal together - not the buyer’s underlying business. That’s circular financing. It doesn’t automatically mean the product is fake or the demand behind it is fake. It means the money proving the demand exists and the money paying for the product can trace back to the same source.

The Bank for International Settlements (BIS) named the sustainability of AI-related investment as one of four pressure points threatening the global economy this year, and inside that risk, singled out circular financing by name: equity stakes and purchase commitments between chipmakers, hyperscalers, and AI labs, tangled enough that the same asset can end up pledged more than once. Bloomberg has mapped the web: Nvidia, Microsoft, OpenAI, AMD, Anthropic, all invested in each other, all buying from each other, in arrangements that get more tangled the closer you look. Most of what’s been written treats this as one undifferentiated phenomenon - a sign of the whole AI buildout resting on money that doesn’t actually exist.

It isn’t one phenomenon. It’s concentrated almost entirely in the top layers of the hardware stack, and it thins out, sharply, the deeper you go. By the time you reach advanced packaging, foundry capacity, equipment, and the raw materials underneath all of it, it disappears completely. That boundary is the actual finding here, and it’s more useful than another map of who’s connected to whom.

What circular financing actually looks like

Google’s own filings are the cleanest place to see the mechanism. Alphabet’s guarantee obligations tied to data center backstops were zero at the end of 2024. By the end of 2025 it was $16.9 billion. By the end of March 2026 it was $28.4 billion. It’s now roughly $43.8 billion, after another $15.3 billion in backstop agreements disclosed in April. That’s not a typo, and it’s not organic growth in the ordinary sense - it’s Google agreeing to guarantee the leases that smaller data center developers sign, so those developers can borrow more cheaply than their own balance sheets would allow. In exchange, Google took warrants (the right to buy shares later at today’s price) that could convert into roughly 14% of one of those developers (TeraWulf) and 5.4% of another (Cipher Digital). Google isn’t just a customer here. It’s the reason the debt gets priced the way it does.

That structure has since repeated twice more, each time with a different guarantor doing the same job. Bitdeer’s obligations to a neocloud intermediary are backed by a $1.3 billion bank letter of credit instead of a hyperscaler’s own balance sheet - same function, different institution. Anthropic’s newest infrastructure deal stacks two guarantees on one project: Google on the property, Broadcom on the chip costs, with Google expected to walk away holding roughly a fifth of it.

Three deals, three different guarantors, same underlying move: decouple the customer’s credit risk from the customer’s own balance sheet. Three instances in five days is a template, not a coincidence.

None of this makes the underlying compute demand fake. What it means is that some of these companies’ ability to pay is, in part, backstopped by the same ecosystem selling to them. If the assumptions behind the long-term commitments stop holding, decoupling the credit risk doesn’t make the risk disappear. It just moves who’s actually holding it.

Where the pattern gets weaker

High bandwidth memory looks, at first glance, like it should belong in the same category. OpenAI, Samsung, and SK Hynix signed a letter of intent for eventual supply of 900,000 DRAM wafers a month for Stargate. In the same window, Samsung bought 50,000 Nvidia GPUs for its own AI factory, and SK Group announced an Nvidia-powered facility for semiconductor research. Nvidia and SK Hynix formalized a multiyear co-development agreement in June. Everyone is buying from everyone.

But no equity changes hands anywhere in it. No vendor invests in a customer’s ability to purchase from them. No special purpose vehicle sits underneath any of it. What’s actually happening is reciprocal purchasing and deep technical co-development, with each party becoming supplier and customer to the other across the ecosystem. That’s real, and it’s consequential, but it isn’t the financing mechanism running underneath the Nexus Data Centers deal. SK Hynix holds an estimated 56 to 58% of the HBM market and 60 to 70% of the HBM4 volume allocated to Nvidia’s Vera Rubin platform. Micron’s HBM capacity is sold out through the end of 2026. Neither company needed anyone’s equity to get there. They got there on contracts and yield, the same two things that have driven this entire memory cycle all year.

Where it stops entirely

Several layers checked clean, and each one is worth its own look because they fail the same test for slightly different reasons.

Networking is Arista’s story specifically. Debt-to-equity near zero. Microsoft and Meta each account for over 10% of revenue (the disclosure threshold) and neither holds equity in the company. No vendor financing anywhere in the relationship. The only sole-source dependency runs the other direction - Arista depends on Broadcom for switching silicon, not the other way around. The data center Ethernet switch market hit $15.4 billion in the first quarter of this year, up 39.8%, with Nvidia’s own share in that market rising from under 4% to 21.5% in a single year. That’s a fast, real share shift, and it happened entirely through selling better products faster - not through financial engineering. One caveat worth keeping distinct: this finding is specifically scoped to switching. Nvidia made a $4 billion strategic investment in Lumentum and Coherent ahead of a major optics conference this year, a genuine instance of vendor equity at the optics layer, sitting right next to the clean networking finding rather than inside it. The two shouldn’t be collapsed into one connectivity story.

Advanced packaging is a queue, not a cap table. Total CoWoS demand for this year is estimated near 1.0 million wafers, against roughly 370,000 in 2024. Nvidia is estimated to hold 60% of 2026 output and has reportedly booked more than half of the expansion capacity through 2027. TSMC’s own chairman told shareholders in June that capacity remains sold out through 2026. Nobody bought equity to get to the front of that line. They booked capacity and waited, the same as everyone else, with purchase orders and lead time as the only currency that works.

Foundry runs on the identical logic, one layer over. Leading-edge capacity is booked into 2028, allocated the same way packaging is, by priority and lead time rather than by who owns a piece of whom. Nobody buys their way to the front of a two-year queue with equity. They buy their way there by ordering early and paying for volume.

Equipment is the one worth slowing down on because for a moment it seemed like it might break the pattern, but it does the opposite once you actually look. There’s exactly one arrangement anywhere below the memory layer that resembles vendor equity at all, and it involves the company that makes the machines the entire leading-edge industry depends on to make anything. What actually happened there, and why it ended the way it did, turns out to be the clearest single piece of evidence for everything this whole boundary argument is built on. It gets its own section below, because it deserves one.

Materials are the floor, and they don’t run on financing logic at all - they run on politics and chemistry. Rare earths and critical minerals, the inputs sitting underneath everything above them, are gated by Chinese export licensing, qualification timelines running 12 to 18 months, and shipment-by-shipment approval from Beijing. No equity stake gets a company past that. No guarantee speeds it up. The constraint here isn’t financial in any sense. It’s geopolitical and physical - a government deciding whether to issue a license, not a bank deciding whether to extend credit. That’s the same wall this entire publication has been mapping all year in a different context, and it’s worth naming directly: the deepest layer of the AI stack and the deepest layer of the critical minerals story are, it turns out, the same layer.

The exception that confirms the rule

In 2012, ASML needed its customers’ money to keep EUV lithography development alive. Intel, TSMC, and Samsung together took a 23% minority stake for 3.85 billion euros in cash, plus R&D funding commitments. The shares were non-voting except in extraordinary circumstances, and the cash proceeds were returned to non-participating shareholders through a synthetic buyback so no one got diluted. The funding wasn’t restricted to the participants - it went toward advancing the technology itself, not toward locking in individual customers.

Intel started unwinding its stake in 2017, and by 2024 ASML held the entire EUV market and no longer needed anyone’s equity to fund its own research. That’s the tell: customer equity showed up only while EUV was unproven, and once scarcity value was established beyond doubt, the financing structure became unnecessary and simply went away. Deep layers of this stack don’t require capital engineering because the scarcity underneath them is real enough not to need dressing up.

Nvidia’s own answer to this question

That still leaves the top of the stack, where the actual question everyone’s been asking is loudest. This week, the industry answered it directly.

On August 10, Nvidia announced financing partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, aimed at mobilizing more than $500 billion in third-party capital for AI infrastructure over time. Jensen Huang addressed the obvious question in his own post: “Is this circular financing? This initiative is designed to address that concern.”

Here’s what was actually signed: memorandums of understanding, a signal of intent rather than a funded commitment, confirmed by Nvidia itself and every outlet that covered it. The $500 billion is a mobilization target, not money raised. It isn’t Nvidia’s revenue, a single fund, or a promise to any one customer, and every individual project still needs its own final agreement before a dollar moves. Nvidia’s own disclosed exposure is a residual-value guarantee - a promise to cover part of the loss if the equipment is worth less than expected later - of up to 25% on any one deal, decided project by project. That’s the same idea as Google’s guarantees earlier in this piece, structured with a stated ceiling and disclosed up front, rather than the way Google’s grew, with no announced cap, and became visible only in a filing months after the fact.

Huang’s defense answers one question well and doesn’t touch a second one. His case rests on independent underwriting - six firms deciding for themselves, not Nvidia’s own balance sheet deciding for them - and on that narrow point, he’s right: this isn’t the same mechanism as Google guaranteeing a lease. What it doesn’t address is how independent the six underwriters actually are of Nvidia, and of each other. The Kuwait Investment Authority is the clearest tell. It’s a co-founding partner in Brookfield’s own $100 billion fund with Nvidia, running since November 2025, nine months before this week’s announcement. It’s also a co-investor alongside Nvidia in KKR’s infrastructure platform, Helix. And it’s the anchor investor in the BlackRock and Microsoft data center vehicle Nvidia already sits inside. Three of the six firms in this week’s consortium - Brookfield, KKR, and BlackRock - already share the same sovereign investor sitting next to Nvidia. None of that makes the underwriting fake. It means the pool of firms actually writing these checks is smaller and more repeat than six new independent partners implies.

Huang’s rebuttal and this finding aren’t actually in conflict. Both can be true. Independent underwriting is real. So is a narrower, more familiar pool of underwriters than the announcement’s own framing suggests. Nvidia answered the question it wanted to answer. It didn’t answer the one this piece is asking.

There’s a second, separate defense worth testing too: whether the chips themselves justify any of this. CoreWeave’s own August 12 earnings answered it. A renewed Nvidia A100 contract, six-year-old silicon, extends through 2029 at full pricing. CoreWeave also raised prices roughly 25% across its GPU lineup in July - not a company discounting aging inventory. Worth holding against it: one analyst projects CoreWeave running roughly $39 billion of negative free cash flow this year. Pricing power on old hardware and severe cash burn are both true at once.

The same week, Intel raised its own $20 billion the plain way. A public stock sale, a same-day 4% drop nobody hid from, and still more than $100 billion in demand for the shares. No guarantee, no consortium, no residual floor. One company found $20 billion by asking anyone who wanted in to buy shares. The other is building a syndicate to backstop chips it sells to itself.

Nvidia reports earnings on August 26. That’s the first time any of this gets checked against a balance sheet instead of a press release.

What the boundary is actually telling you

Circular financing isn’t a verdict on the whole AI buildout. It’s a symptom that shows up specifically where a company’s ability to pay depends on continuing to raise money rather than on selling something the market would pay for anyway. That’s a real risk, concentrated at a specific altitude in the stack - model labs, some neoclouds, some data center financing vehicles - not evenly distributed across everything with “AI” in its description.

The companies sitting below that altitude aren’t insulated from a broader market correction. If the top of the stack unwinds hard enough, credit tightens for everyone and equity multiples compress across names that never touched a circular structure. In a real panic, everything gets sold at once regardless of individual quality, because everyone needs cash at the same time. Expecting full immunity from that repricing anywhere near this space isn’t realistic.

But medium-term, the distinction that actually decides who comes through intact isn’t whether a name got marked down in the panic. It’s whether the underlying cash flow was ever real in the first place. Arista’s $3 billion quarter didn’t happen because a hyperscaler needed Arista to look bigger to its own investors. It happened because Arista sells the switches that infrastructure needs and nobody else makes at that scale. Micron’s sold-out HBM capacity isn’t a favor from Nvidia. It’s a company that can’t make enough of something everyone needs. If the circular arrangements at the top of the stack come under real stress, the question worth asking isn’t which logo appears in the most financing headlines. It’s which businesses were getting paid for something real all along, and which ones were getting paid to keep a story going.

If the top of the stack corrects hard, watch which companies keep beating guidance on the strength of bookings and lead times rather than on the strength of a new financing structure. That’s the tell before the earnings call confirms it either way. Nvidia’s own August 26 report is the nearest one worth watching.

This isn't new ground. Part 4 of the Semiconductor Series raised this same problem earlier this year, in a section called 'The Circular Demand Problem,' using CoreWeave's own S-1 debt-against-GPU-collateral disclosure as the founding example. This piece runs the same test across the whole stack.

Financial data sourced from company earnings reports, SEC filings, company statements, wire and financial press reporting, and third-party analyst research, as of August 13, 2026.

Standard disclaimer: This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

Full archive at williamdavid.substack.com.

August 8, 2026

Correction (added August 9): The MP Materials section originally stated the Texas magnetics facility "contributed meaningfully to revenue for the first time this quarter." That was wrong. Magnetics revenue was actually down, both sequentially (from $21.1M in Q1 to $16.5M in Q2) and year-over-year, as production ramp-up pulled material away from precursor sales. The section has been corrected below with accurate figures. Apologies for the error.

Five critical mineral stories broke this week. None of the headline numbers told the real one.

Lynas Doubled Its Heavy Rare Earth Output. The Money Followed.

Dysprosium and terbium production more than doubled this quarter. That’s the story. The revenue print is the consequence.

Lynas’s dysprosium and terbium output jumped from 8 to 19 tonnes this quarter, more than doubling. These are two of the elements at the center of this year’s export-control fight, the same two Japan has received zero of from China for eight straight months. Lynas produced its first samarium oxide in March, with gadolinium, yttrium, and lutetium next in the pipeline, all four sit among the seven elements China placed under export licensing in April 2025.

The financial print reflects that mix shift more than it reflects raw growth. Q4 FY26 gross sales revenue hit A$288.9 million, up 70% year over year, on a record average selling price of A$98.2 per kilogram, driven by a bigger share of heavy rare earth sales and stronger premiums over the market index, not a volume story. NdPr output actually fell, from 1,996 to 1,857 tonnes, worth naming plainly rather than glossing over: the heavy rare earth ramp appears to be drawing focus and capacity away from light rare earth volume, at least this quarter.

The Malaysia heavy rare earth expansion tells the same processing-difficulty story from the cost side. Confirmed at A$294 million, the increase traces specifically to additional equipment needed to meet customer-specific purity requirements, plus the higher cost of sourcing that equipment outside China. That’s not a generic cost overrun, it’s the qualification bar itself getting more expensive to clear. And the Kuantan facility’s operating license, the one the entire expansion depends on, was due for renewal March 2, applied for in January, with terms “not yet received” as of the most recent report available.

A production story, a mix-driven financial beat, a processing-cost story, and an unresolved regulatory gate, four different things happening inside one quarter, and only the second one made the headline.

MP Materials: MP Materials: A Gadolinium Deal, and a Magnet Line Still Waiting on GM

A new customer for an export-licensed element, and a downstream facility still funding the wait. The earnings beat is the evidence, not the news.

MP signed a new long-term gadolinium supply agreement with a US aerospace and defense customer this quarter, described on the earnings call as a nine-figure deal running multiple years, for gadolinium oxide specifically. Gadolinium sits in the same China-export-licensed category as three of the four elements Lynas is racing to bring online, worth flagging as a genuine parallel: two Western rare earth companies, same week, moving toward the same restricted-element category from different starting points.

The Texas magnetics facility generated $16.5 million in revenue and $7.5 million in adjusted profit this quarter, and it’s worth being precise about what that actually is: precursor product sales, not finished magnets, running at margins above 40%. That revenue was down both sequentially (from $21.1 million in Q1) and year over year, as production start-up at the Independence facility pulled material away from precursor sales. Finished magnets have been delivered to General Motors for in-car qualification, with initial commercial deliveries still guided to the fourth quarter. So the downstream business isn’t yet a magnet business. It’s a precursor business at good margins, funding the wait for one, and the real story here is that MP also completed its first heavy rare earth separation circuit this quarter, with terbium and dysprosium output expected later this year.

The financial context around both of those: Q2 revenue of $108.5 million, up 89% year over year and beating estimates by roughly 13%, with adjusted EBITDA turning positive for the first time, $28.5 million, a $41 million swing from a year ago. Worth knowing before reading that as pure market strength: $17.6 million of it is price protection agreement income, contractual price support rather than unsupported market pricing. GAAP net loss narrowed to $0.11 per share from $0.19, still negative on depreciation and start-up costs, the same buildout-phase economics working against Lynas’s own margins.

Same sector, same week, same regulatory backdrop, but MP’s version carries a more precise asterisk than Lynas’s: not whether the story holds, but whether GM qualification clears on schedule. If it slips past Q4, the precursor business carries the wait another two quarters.

Indium Just Showed Its Third Different Face in Three Months

China, the US, and Europe are now pricing the same kilogram of indium within four cents of each other. Nobody called this shape.

For most of indium’s recent history, Western buyers paid a premium over China’s domestic price. In July, that flipped entirely, China’s own domestic benchmark moved above both US and European prices for the first time on record. This week’s SMM data shows a third state neither pattern predicted: all three benchmarks have converged to within four cents of each other, $804.96 per kilogram in China, $805.00 in the US, $805.00 in Europe.

The mechanism matters as much as the number. This didn’t happen because China’s premium unwound. It happened because Western prices jumped to meet it, US indium up 12.2% on the month, Europe up 13.4%, while China’s own domestic price rose a comparatively modest 3.8%. Western buyers are now paying more to match China’s level, not benefiting from China’s advantage disappearing. Why is genuinely unclear from the available sourcing, and whether this convergence holds into September or breaks into a fourth pattern entirely is an open question this note isn’t going to pretend to answer.

Same week, same 2023 China export-control tranche, opposite behavior next door: gallium fell 18.8% to $234.62/kg with its Western premium roughly flat, while germanium fell a smaller 9.2% to $3,104.31/kg even as its own Western premium widened from 82.9% to 104.6%, domestic giving back gains while Western warehouse pricing barely moved. Three metals from the same regulatory tranche, three different relationships between domestic and Western pricing, in the same week. Indium is the one that changed its entire market structure twice in ninety days. It’s also part of a broader pullback, thirteen of the nineteen elements this note tracks pulled back this week, reversing July’s near-unanimous rally.

Fifty Percent More Ore, Zero Percent More Rare Earth

China just approved a major expansion at the world’s largest rare earth deposit. It has almost nothing to do with rare earths.

Baogang Group will spend $74 million expanding the Bayan Obo mine in Inner Mongolia, raising total capacity 50%, from 10 to 15 million tonnes a year. Read only the tonnage and this looks like Beijing loosening the tap. It isn’t.

Li Yang, a geologist at Peking University, said it plainly: “Rare earth production is subject to quotas, so there is no need to mine more ore. The main purpose is to mine iron, while [rare earths are] recovered along the way.” Bayan Obo is an iron mine first. Rare earths are the by-product, and the by-product’s output is capped by quota regardless of how much rock comes out of the ground. Fifty percent more ore changes iron supply. It changes rare earth supply not at all.

This is worth sitting with precisely because of what it isn’t: it’s not evidence of loosening, and it’s not evidence of tightening either. It’s evidence that the actual constraint was never the ore in the first place - the same lesson this week’s two Western earnings prints just delivered from the opposite direction, at a mine Beijing doesn’t control instead of one it does.

The Acid Question Finally Has a Chart

A consulting study just put a number on the copper risk nobody’s pricing yet: zero at 30 days, 111,000 tonnes at 90, 947,000 tonnes at a year.

Copper closed at $6.56 per pound on COMEX as of August 7, up from $6.44 the prior week, and has been making fresh highs on two separate supply threads: the DR Congo’s concentrate export restrictions and Codelco’s disruption at El Teniente. Neither of those is the acid story. GEM Mining Consulting’s new stress-test model is a third, quieter thread underneath both.

The model doesn’t forecast a disruption. It measures what one would cost if it happened. More than 15% of global primary copper, roughly 3.6 million tonnes a year, depends on sulfuric acid to convert ore into metal. Run a 30-day acid disruption through the model and almost nothing is exposed. Run a year-long one and just under a million tonnes is.

The reason this isn’t hypothetical: roughly half of all seaborne sulfur trade passes through the Strait of Hormuz, and China halted its own sulfuric acid exports back in May. Two separate acid chokepoints, one physical and one political, sitting on top of the same 3.6 million tonnes. Goldman Sachs raised its own ex-US 2026 copper deficit forecast tenfold in June, to 640,000 tonnes - within the range this model’s 90-to-365-day window would predict, for reasons that have nothing to do with how much copper ore exists or what’s happening at any single mine.

When the chokepoint is in the conversion chemistry, not the mine, domestic Chinese price movements don’t travel West.

The common thread:

Whether it’s Lynas’s heavy rare earth output doubling before the revenue print caught up, MP’s gadolinium deal and magnet line doing the real work behind an EBITDA inflection, indium rewriting its own market structure for the third time in three months, Bayan Obo’s ore expansion changing nothing about rare earth supply, or a copper price climbing on threads that don’t include the one this note actually tracks, the bottleneck isn’t the rock - it’s everything that happens after it leaves the mine.

Three things I’ll be watching next week:

Terms on MP’s gadolinium supply deal and qualification detail on what’s actually moving through the Texas magnetics line - both open as of this edition.

Whether indium’s convergence holds, reverts toward July’s inversion, or breaks into a fourth pattern entirely - three distinct market structures in three months has no precedent in this note’s tracking.

Whether Lynas’s Kuantan facility license renewal terms are actually published - a live regulatory gate sitting underneath the whole Malaysia expansion story.

The clock:

November 10 - REE extractant controls resume.

November 27 - gallium truce expiry.

January 1 - DFARS 252.225-7052.

94 days to extractant controls. 111 days to the gallium truce. 146 days to DFARS.

Live critical mineral prices at scarceearth.io.

The full archive is at williamdavid.substack.com. The Week in Minerals publishes every Saturday.

Lynas Q4 FY26 results sourced from company filings via Quartr, Discovery Alert, and rare-earth-mining.com, confirmed August 6, 2026. Kuantan facility license status via rare-earth-mining.com. MP Materials Q2 2026 results sourced from company press release via StockTitan, Investing.com, and NAI 500, confirmed August 6-7, 2026. Indium, gallium, and germanium pricing via SMM (metal.com) and rare-earth-mining.com, benchmark date August 3, 2026, cross-referenced against ScarceEarth.io admin values. Bayan Obo expansion sourced from South China Morning Post, confirmed August 7, 2026. GEM Mining Consulting study sourced via Mining.com (Cecilia Jamasmie), August 6-7, 2026. Copper price via TradingEconomics (COMEX), August 7, 2026. Copper deficit figures via Goldman Sachs, Morgan Stanley, JPMorgan, and ICSG estimates, June 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

Update, August 10, 2026: Iran’s Supreme Leader Mojtaba Khamenei has restructured the country’s senior military and IRGC command chain, confirmed by Iran military analyst Jason Brodsky. Most relevant to the thesis breaker above: Ahmad Vahidi, the IRGC commander this piece names as the faction most likely to reject any arrangement that meaningfully reduces Iranian control, has been promoted to major general. The appointments elsewhere follow the same direction, including a member of Khamenei’s own inner circle named to command the Basij. The pattern is consolidation of hardline control, not dilution of it. If anything, the odds against episode three ending differently just got longer.

Maersk, MSC, CMA CGM, and Hapag-Lloyd have now watched the Strait of Hormuz get “resolved” twice this year and are currently watching a third attempt. None of the four carriers has rebuilt its Gulf schedule around any of them.

That’s not stubbornness. It’s pattern recognition. The first resolution, a two-week ceasefire agreed April 7-8, never delivered the relief it promised. Iran extended it indefinitely rather than letting it lapse, but transit continued at a trickle for two months, still contested, until the arrangement was folded into a new agreement rather than broken by any single event. The second, the Islamabad Memorandum signed June 17, held longer. Long enough that all four carriers used the window to pull vessels stranded in the Gulf back out. Then Iranian forces attacked three tankers on July 6 and 7, the United States revoked its own sanctions relief the same day, and the ceasefire the whole arrangement depended on was declared dead by July 8. Nothing has crossed the strait’s traditional shipping route with its tracking transponder switched on since.

Now there’s a third attempt. Iran and Oman say their talks are “on the way to being finalized.” The question worth asking isn’t whether this one holds. It’s whether anyone has actually fixed the thing that broke the first two.

The clause nobody read the same way

The Islamabad Memorandum’s fifth clause committed Iran to “dialogue” with Oman on the strait’s future administration, and separately guaranteed free transit for sixty days. Iran read the free-transit language as its own temporary concession, with control reverting to Tehran once the window closed. The United States read it as a permanent settlement of the navigation question. Both governments signed the same document. Neither signed the same agreement.

That gap didn’t stay theoretical. Iran’s foreign ministry spokesman, Esmaeil Baghaei, has said plainly that the strait will not “return to the status quo” that existed before the war, and has been equally direct that an agreement with Oman on a new transit route “has no connection to whether the Strait of Hormuz is opened or remains closed.” Two statements, six weeks apart, the same refusal underneath both.

The same split is already running through the agreement nobody has signed yet, and it is worth watching for, because it is the tell that nothing structural has changed.

What “resolved” actually looks like on the water

Start with something that shouldn’t exist if any of these resolutions had genuinely held: naval mines still sit in the traditional shipping route through the strait’s middle channel, forcing vessels onto alternate paths regardless of what any memorandum says. The IMO has formally advised ships to avoid the route it adopted, and Tehran is now reportedly weighing whether to let European countries clear the mines. The result is a strait split into three routes, not one: a southern route near Omani waters, a middle route now mine-affected, and a northern route that’s Iranian-controlled, fee-gated, and permission-based. Every routing decision is a political decision wearing a logistics decision’s clothes.

Aramco’s chief executive, Amin Nasser, told analysts on the company’s second-quarter call that the conflict “continues to aggravate the biggest supply shock in history,” with more than 2.6 billion barrels removed from global supply.

Then there’s the insurance layer. In late July, Treasury and State designated two entities backed by Iran’s Revolutionary Guard Corps (IRGC, the branch of Iran’s military that runs Hormuz enforcement): Persian Gulf Marine Insurance Company and Hormuz Safe Marine Services Authority.

The mechanism is a protection racket at sea. Manufacture the risk, then sell coverage against the danger the same regime created. Shanaka Anslem Perera has called this the insurance weapon. Treasury just made it a named enforcement action with two specific defendants.

The clock that broke, and the one that replaced it

The Islamabad Memorandum’s toll-free window was real. Sixty days from the June 17 signing, landing August 16, it would have been the actual deadline if the agreement had lasted that long. It didn’t. Iranian forces struck three tankers on July 6 and 7. Treasury revoked its own sanctions relief the same day as the attacks, not sixty days later. Trump declared the ceasefire over on July 8, more than five weeks before the date that was supposed to matter.

Then something stranger happened. The United States, which had spent the prior month insisting the strait stay toll-free, proposed its own toll. Trump floated a 20 percent charge on all cargo transiting the strait, styling the US as the “Guardian of the Hormuz Strait.” Iran’s foreign minister, Abbas Araghchi, replied that Trump was “absolutely right,” and said Iran would charge less. “We will be fair,” he wrote. Washington abandoned the proposal about two weeks later and pivoted toward trade and investment deals with Gulf states. Tehran never abandoned the principle. The service fee now under discussion with Oman is the same idea with a different name on the invoice.

China’s crude imports over this same period fell, and it’s tempting to read that as strait-driven shortage. It isn’t. EIA data shows the pullback is inventory drawdown, not supply failure. China stockpiled at record volumes while oil was cheap earlier in the year and is now drawing down the reserve it built for exactly this scenario. A real number, just not evidence of what it looks like at first glance.

The system, not the episode

Osamah Almokdad’s framing keeps proving itself right, and it just proved itself right a third time. His argument: an agreement resolves the episode, not the system that reproduces it. The April ceasefire ended a specific standoff without touching who actually controls the strait. The Islamabad Memorandum did the same thing two months later, on a longer timeline, with more structure, and still didn’t touch it.

The Iran-Oman talks now underway aren’t really a new negotiation. Per Iran’s own account, Oman and Iran started discussing the strait’s future management the day after the April ceasefire, months before Islamabad, running the entire time in parallel with everything else in this piece. This third attempt isn’t a sequel. It’s the same unresolved question, still open, finally circling back toward a decision point.

What the fleet already knows

Two failed resolutions is a pattern a fleet only has to learn once. Reversing a global shipping network, repositioning vessels, rewriting insurance terms, restoring port call patterns, costs far more than diverting one in the first place, and nobody wants to be the carrier that reverses first into a corridor that breaks a third time. The four carriers watching this aren’t betting the strait stays contested forever. They’re betting that the cost of moving early, twice burned, outweighs the cost of waiting to see if this attempt is actually different.

The Oman track is real and it’s moving. Araghchi told the cabinet in early August that talks are “on the way to being finalized.” He said this the same week a parliamentary committee began compiling documentation to impeach him, over these talks specifically. But what’s being negotiated isn’t a reopening. It’s an administration. Per Reuters reporting, which Marjorie Nadal read closely, Iran would control clearance for vessels entering the Gulf, Oman would issue clearance for vessels leaving after notifying Iran, and Iran would keep a stated right to intervene on the way out. Nadal’s read on the asymmetry is the one worth carrying: a vessel refused entry never reaches the terminal it was sailing for, whether it arrives empty to load crude or loaded with containers. Iran doesn’t need control over what leaves if it controls what gets in. Notification is not consent.

Which raises a question about what Oman is actually getting. On the terms as reported, not much. It would issue clearance for the direction that matters less, under notification to a partner who reserves the right to overrule it. An even split of a service fee is not an even split of authority.

What Oman brings to this is its name. The shipping lanes that have governed the strait since 1968 were proposed jointly by Iran and Oman, and the corridor now under discussion would replace them, on Iran’s own description, as an interim arrangement pending a permanent bilateral successor. Oman’s presence is what allows that to read as continuity with 1968 rather than as one country rewriting an international scheme by itself.

That is inference rather than reporting. Nobody has published Muscat’s internal calculus. But it fits the terms as described better than co-administration does.

If the Oman-Iran talks conclude and hold through a real test, a route both governments actually operate under rather than announce, the three-route structure this piece has tracked collapses into something genuinely new, and the fleet-reversal calculus changes with it. If the talks conclude but don’t hold, or don’t conclude at all, this becomes episode three of the same failure, and the physical clock this publication has followed since June keeps running on whatever the next flashpoint turns out to be.

Who actually owns the lanes

There’s a harder obstacle than any of the parties’ objections, and it has been on the record since July. The IMO is the UN agency that writes the rules for commercial shipping, and the lanes running through Hormuz are its lanes: a traffic separation scheme, meaning designated one-way corridors that keep opposing traffic apart, adopted in 1968 on a joint proposal from Iran and Oman.

On July 13, at the close of its 137th session, the IMO Council said three things that bear directly on what is now being negotiated. Transit through straits used for international navigation cannot be threatened, impeded, denied, hampered, impaired or suspended. Any arrangement between the coastal states must guarantee non-discriminatory and unimpeded transit through that 1968 scheme. And passage through the strait should remain free of any tolls and charges, in accordance with international law.

Three weeks later, Iran and Oman were negotiating a corridor that replaces the 1968 scheme, gated on Iranian clearance, with a service fee split between them.

This isn’t a fringe position at the IMO. Back in March, at an extraordinary session attended by more than 120 member states, the Council adopted a declaration tabled by the UAE and co-sponsored by more than 115 members, the most in the organization’s history, demanding Iran stop interfering with navigation through the strait. Iran kept going.

The legal argument underneath is older than this war. UNCLOS, the UN treaty governing who may do what at sea, prohibits coastal states from charging for transit passage through an international strait. That is why Oman’s own earlier proposal was modeled on the Malacca Strait, where the coastal states levy no transit fee at all. Iran has never ratified UNCLOS. It argues that only the narrower doctrine of innocent passage applies, which would give it considerably more say over who transits and on what terms. The service fee isn’t a pricing detail. It’s the legal dispute wearing a commercial disguise.

What each side is actually saying

Asked directly on August 4 whether Iran would be permitted to charge a toll, Treasury Secretary Scott Bessent answered, “It would be freedom of movement,” and gave no detail on how that would be enforced. The question was about payment. The answer was about passage. Meanwhile Iranian officials describe a service fee split evenly with Oman, covering security and staffing, while a US official quoted in the same New York Times report described the same arrangement as involving no tolls and no Iranian permissions. Same document, two readings, nobody reconciling them. The fifth clause didn’t get fixed. It got reissued.

Brent fell 5.3 percent that day, closing at $79.36, on the strength of a deal that has not been signed and whose parties are describing different documents.

Qatar, which is mediating, has not moved. Its published position is that Hormuz is an international passage that cannot be controlled by any single party. On August 4 its foreign ministry spokesman, Majed Al Ansari, went further: any future system for managing the strait, including any financial arrangement, has to comply with international law and be agreed across the region, and Qatar’s minimum demand is a return to the strait’s previous status. That is the precise thing Baghaei has ruled out. The mediator’s floor and Iran’s ceiling are the same line, viewed from opposite sides.

The thesis breaker is specific, and it comes from the same source that’s been right about everything else in this piece. ISW’s assessment names the reason to doubt this attempt holds any better than the first two: the IRGC faction under Commander Ahmad Vahidi currently dominates Iran’s decision-making on this question, and that faction is likely to reject any arrangement that meaningfully reduces Iranian control, even if pragmatist negotiators like Araghchi are willing to compromise. Watch who signs off on the final route, not who announces it’s close. If Vahidi’s faction blocks or waters down whatever Araghchi brings back, this is episode three. If it doesn’t, that’s the real story: the first time in three tries someone found a version of this agreement that survives contact with the people who’d have to enforce it.

None of this means the strait is closed in the way “closed” usually means. Kpler’s tracking shows more than 9.2 million barrels of Iranian crude have moved through Hormuz since the ceasefire collapsed, running on sanctioned tankers with their transponders switched off. The strait isn’t shut. It’s dark. Those aren’t the same thing, and the gap between them is exactly what the next round of this argument will be fought over.

Markets price narratives. Supply chains price physics. The narrative has had three clean stories this year: a ceasefire, a memorandum, and now a finalized route. The physics has had one unanswered question the whole time, and it is still unanswered.

This is the standalone Hormuz synthesis this publication’s notes on the corridor have been building toward since June, and it reads differently from most of what runs here. Every other series this publication has built assumes a working global economy underneath it. Semiconductors assume shipping lanes stay open. Defense assumes materials move on schedule. Critical minerals assume export routes hold. None of that is guaranteed right now, and Hormuz is where the assumption is currently being tested. This piece is that test, in full.

The next series moves one layer over, from the strait itself to the full energy supply chain sitting downstream of it: extraction, refinement, equipment, distribution. The same question that’s been open here all year, who actually controls the physical infrastructure everything else depends on, doesn’t stop at the strait.

Financial data sourced from company earnings reports and public market data as of August 6, 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

The full archive is at williamdavid.substack.com.

Share

August 1, 2026

For eight months, the story was Japan. This week, the customs data showed China ran the exact same play on a second country, in the same window.

Yttrium Isn’t Just a Japan Story Anymore

The US got the same treatment, in the same window. That’s not a coincidence. It’s a policy.

Chinese exports of yttrium to the United States fell to zero for two consecutive months in mid-2026 - the first time that has happened since the licensing regime began. This wasn’t a Japan-specific measure that happened to coincide. It was the same mechanism, applied simultaneously to a second major economy. Silverado Policy Accelerator’s trade tracking puts the scale in context: China shipped just 17 tons of yttrium to the US across the eight months following the April 2025 controls, against 333 tons in the same window a year earlier. Aerospace suppliers, who depend on yttrium for turbine thermal-barrier coatings, have already had to pause some production.

Dysprosium is the tell. It has stayed close to pre-control export levels this year - almost entirely routed to South Korea. Beijing isn’t short on rare earths. It’s choosing, element by element and country by country, who keeps receiving them. Japan and the US both landed on the restricted side of that choice, in the same stretch of months. EU and Korea did not.

The price gap confirms it. SMM’s domestic China benchmark puts yttrium oxide at $38.22 per kilogram as of July 1. The comparable European price sits in the $270-320 per kilogram range - an 8-to-10x spread for the same material, split by geography and by policy, not by cost.

For investors: that spread is the licensing regime priced in dollars per kilogram, the same mechanism the site has flagged with dysprosium and terbium all summer - not a market inefficiency waiting to close.

Japan Isn’t Waiting Anymore

Eight months of zero rare earth exports just bought Namibia a Japanese government check.

JOGMEC, Japan’s state metals and energy security agency, is investing up to roughly $34 million in TJ Namibia Rare Earths Corporation, a special-purpose vehicle set up by Toyota Tsusho to advance the Lofdal heavy rare earth project in Namibia’s Kunene region. The initial investment landed July 23. The target elements are dysprosium and terbium - the same two Japan has been unable to source from China since November. A feasibility study is underway, with a final commercialization decision targeted for the fiscal year ending March 2027. It is Japan’s first rare earth mine development project on the African continent.

Namibia Critical Metals Inc., JOGMEC's original joint-venture partner on Lofdal since 2020, is the direct investable exposure to the project itself.

This is a government-backed entity moving because waiting stopped being viable, not because a diversification slide finally got funded. But a feasibility study is not production, and a commercialization decision is not qualified supply. This is the same trajectory-not-arrival signal the Alcoa gallium plant sent two weeks ago, showing up again with a different mineral and a different continent.

The Circuit Won’t Beat the Deadline

Construction started this week. The law that requires the output starts five weeks before the circuits do.

Energy Fuels began construction on a $104 million expansion of its White Mesa Mill in Utah, targeting commercial heavy rare earth oxide production. The terbium and dysprosium circuits are slated to be complete by the end of 2027; samarium, europium, and gadolinium follow by the end of 2028. NdPr capacity is already online at 1,000 tonnes per year. Feedstock is secured through the Donald Project joint venture in Australia, and the buildout carries a conditional $725 million Department of War loan commitment. CEO Ross Bhappu put it plainly: heavy rare earth production is a severe pinch point in North American and European magnet supply chains. China currently refines an estimated 98 percent of the world’s heavy rare earth oxides.

DFARS 252.225-7052 takes effect January 1, 2027. Energy Fuels’ own terbium and dysprosium circuits aren’t targeted to be complete until eleven months after that. The law arrives first. The supply chain it requires arrives after it.

The common thread

Whether it’s yttrium’s parallel freeze-out of Japan and the US, Japan writing a check to Namibia because waiting stopped making sense, or Energy Fuels racing a deadline its own circuits won’t beat until the following year, the bottleneck isn’t the rock - it’s everything that happens after it leaves the mine.

Three things I’ll be watching next week

MP Materials Q2 earnings, now confirmed for August 6 (not July 30, as some trackers still show) - NdPr pricing against the $110/kg DoD floor is the number that matters.
Whether the MOFCOM whistleblower system (Announcement No. 26) produces any publicly reported enforcement action - the first real test of whether the chilling effect translates into actual cases.
Any schedule commentary from Energy Fuels on the White Mesa terbium/dysprosium circuit timeline - slippage in either direction changes how the DFARS collision actually plays out.

The clock

November 10 - REE extractant controls resume.
November 27 - gallium truce expiry.
January 1 - DFARS 252.225-7052.
101 days to extractant controls. 118 days to the gallium truce. 153 days to DFARS.

Live critical mineral prices at scarceearth.io.

The full archive is at williamdavid.substack.com. The Week in Minerals publishes every Saturday.

Yttrium export and dysprosium routing data via Silverado Policy Accelerator trade tracking, CSIS analysis, and S&P Global. Yttrium pricing via SMM domestic benchmark (rare-earth-mining.com), verified August 1, 2026, and European/Western pricing range via Argus-sourced data (November 2025) and corroborating Q1 2026 market reports. JOGMEC/Toyota Tsusho Namibia investment via Mining Weekly and Toyota Tsusho press release, July 30, 2026. Energy Fuels White Mesa Mill expansion via company press release, Mining.com, Mining Technology, and Rare Earth Exchanges, July 29, 2026. MP Materials Q2 earnings date confirmed via company IR.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

Share

Picture a computer’s memory chip as a single-story warehouse. Data comes in one door, gets shelved, and comes back out the same door when needed. That’s DRAM (Dynamic Random-Access Memory) - the fast, temporary workspace every processor uses while it’s actively computing. It’s been the same basic design for decades.

Now picture stacking twelve of those warehouses directly on top of each other, and instead of trucks driving in and out through one door on the ground floor, you drill a shaft straight down through all twelve floors and run elevators through it, thousands of them, side by side. That’s HBM (High-Bandwidth Memory). It isn’t a different kind of building. It’s the same warehouse, stacked, with a much faster way to move things in and out.

This distinction matters more than it sounds like it should, because most coverage this week is going to get it backwards.

The shafts are called TSVs - through-silicon vias, tiny vertical copper connections drilled through each layer of the stack so power and data can travel straight up instead of sideways across a circuit board. A twelve-layer HBM stack contains millions of them. One bad connection, out of millions, and the layer above it is dead.

To build the stack, each layer starts life as part of a wafer, a thin, round slice of silicon (roughly the size of a dinner plate) that hundreds of individual chips get cut from once manufacturing is finished. For HBM, those slices have to be ground down to an extreme thinness before they’re stacked - around 50 micrometers for a twelve-layer stack, dropping to roughly 30 micrometers for sixteen layers. That’s less than half the width of a human hair. At that thinness, a wafer cracks easily and bows under its own weight, and a bowed wafer can’t be bonded (fused precisely to the layer beneath it) with the accuracy the stack requires.

Then there’s the power problem, which is less visible but arguably harder. DRAM is “dynamic” because the tiny electrical charges that hold each bit of data leak away on their own and have to be topped back up thousands of times per second - that top-up is called a refresh. Every refresh needs a quick burst of power delivered through those same vertical shafts, all at once, to every layer in the stack. The floors farthest from where the power comes in, the top of the stack, are the hardest to reach cleanly, and during a refresh burst the voltage reaching them can sag right when it’s needed most. Engineers call this voltage droop, and solving it well enough to keep the top layers stable is one of the most closely guarded pieces of know-how in the industry. Nobody publishes exactly how they’ve solved it - it’s one of the few things that meaningfully separates one company’s success rate from another’s.

All of that complexity eventually gets judged by one number: yield. The math behind it is simple enough to hold in your head. If each individual layer-to-layer bond succeeds 99 times out of 100 (already a very good rate), an eight-layer stack’s overall success rate compounds down to about 92%. A twelve-layer stack drops to roughly 87%. Push to sixteen layers and you’re near 86%.

Every additional floor in the warehouse is one more chance for a shaft to fail somewhere along its length, and the failures stack up along with the floors.

This is why only three companies in the world currently manufacture HBM at real volume - SK Hynix, Samsung, and Micron. Not because of a patent wall, either - the real moat is something slower: getting approved.

Before any HBM ships in real volume, it has to pass the customer’s own testing, usually the GPU (the processor chip doing the heavy computational lifting) maker, most often Nvidia. A standard DRAM chip clears that testing in three to six months. HBM takes six to nine months at a minimum - and can stretch past three years when the customer wants custom features built in. Whatever gets approved today is what ships two to three years from now.

Samsung is the clearest proof this isn’t a formality you can rush through with enough money. Its HBM3E chips failed Nvidia’s thermal and power testing in the spring of 2024. A second attempt stumbled again the following year. Samsung, one of the most sophisticated chip manufacturers on the planet, didn’t pass until September 2025 - about eighteen months after that first failure, a delay that swallowed almost an entire GPU generation. That's what "hard to manufacture" actually means (it isn't something thrown around as marketing language). Think about it: a company with nearly unlimited engineering resources needed a year and a half to solve a problem its two rivals had already solved.

Samsung didn’t lose that time because it lacked engineers. It lost it because eighteen months was what the physics demanded.

Micron’s story runs the opposite direction, and it’s just as telling. Micron actually holds nearly twice as many core HBM patents as SK Hynix does, yet controls only about a fifth of the market to SK Hynix’s more than half. If raw engineering depth decided this business, that gap shouldn’t exist. What actually happened: Micron came to HBM later than its two rivals, then quietly cleared the same testing gauntlet that took Samsung a year and a half to pass (production-capable stacks certified for Nvidia’s H200 in early 2024, then certified again for the newer GB300 roughly a year later) without any of the public stumbles Samsung went through. A late start didn’t cost Micron a slow approval - it cost Micron its place in line for orders once approval finally came. Passing the test and winning the volume turn out to be two different races, and a company can win one without winning the other.

Once I understood that approval mattered more than patents, I started reading memory earnings differently: less as a scorecard of who’s most advanced, and more as a scorecard of who’s been in line the longest.

So here’s the part that actually matters for how you read this week’s news:

DRAM and HBM aren’t two different products competing for the same customers. They’re the same underlying chip, split into two very different tiers by how hard each one is to build.

The regular tier - ordinary DRAM and its mobile cousin, LPDDR - is decades-old, well-understood, and increasingly commoditized. This is where Chinese producers like CXMT are closing the gap fastest, because it’s the more standardized, more achievable version of the technology.

The stacked tier (HBM) is where the difficulty lives, and it’s the tier every AI accelerator actually needs. CXMT is reportedly working on its own HBM, using wider shafts that trade away some density for a better success rate - a real, serious effort, still years behind the approval track record SK Hynix, Samsung, and Micron have built up across multiple GPU generations.

That’s bifurcation, not migration. Nobody is walking away from ordinary DRAM. It’s simply becoming the tier where competition is real and margins compress, while HBM stays the tier where three companies hold an approval-and-experience advantage that’s genuinely hard to shortcut.

One honest complication worth sitting with, because a shorter version of this piece would skip it:

Has HBM actually escaped memory’s old boom-and-bust pattern? Nobody knows for certain yet, and it’s worth being straight about that rather than pretending otherwise. Memory has lived through more than a dozen of these cycles over five decades. Demand surges, prices spike, manufacturers race to add capacity, that capacity lands two or three years later all at once, and the market floods. Margins that once ran above 50% fall into the low twenties or worse.

The optimistic case says this time is genuinely different - multi-year contracts already signed, real capacity discipline learned the hard way from past crashes, and approval cycles long enough to stop the kind of overnight oversupply that broke every prior cycle. The skeptical case says exactly this story gets told near the top of every cycle, and that HBM’s premium just buys the incumbents time before it eventually compresses toward what memory has always become (a standardized part any approved vendor can sell). Both readings come from serious people. What’s actually true is somewhere in between: the mechanics described above make this cycle different in kind from the ones before it. That’s not the same thing as a guarantee the industry has escaped its own history for good.

Which brings us to this week:

SK Hynix reported a record quarter, revenue and operating profit both up sharply from a year ago, and missed Wall Street’s forecast on both lines anyway. Look at where the miss actually came from. HBM4 mass production shipments did begin in the second quarter, on schedule, and the company said so directly. What hadn't happened yet was the ramp. Management told analysts on the same call that shipments of some high-value products were pushed into the second half, and that the resulting change in product mix is what pulled the average selling price down. Starting and ramping are two different things, and only one of them had happened by the end of June.

One more thing: Samsung reported on July 30 too - a record quarter. The company says it was the first to ship HBM4 commercially this year, and it has already sent HBM4E samples to major customers. In 2024, Samsung was the company that couldn’t get HBM3E through Nvidia’s testing. Now it’s shipping the next generation and sampling the one after that. SK Hynix still holds the larger share of HBM, and that advantage matters. But qualification resets every cycle. Losing one generation doesn’t mean losing the next.

The mirage isn't that the quarter was fine. It's that the market is about to spend a few days treating a timing problem as a demand problem. Shipments moving from one half of the year to the next tells you something about a ramp schedule. It tells you nothing about whether the layer that's genuinely scarce - the one with the eighteen-month approval cycles and the compounding math above - is any less scarce than it was in April.

Memory is one layer. It’s not the only one that works this way. The AI Hardware Primer maps the rest of the stack, the foundry, the packaging, the GPU itself, and the same pattern runs through every layer of it:

Financial data sourced from company earnings reports and public market data as of July 30, 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

The full archive is at williamdavid.substack.com.

Share

Every technological revolution runs into a physical constraint eventually. Capital can solve one layer, but that just exposes the next one underneath it. The semiconductor series documented that pattern at the fabrication layer - this piece follows it to where it moved next.

Homer City, Pennsylvania used to be home to the largest coal-fired power plant in the United States. It’s now being rebuilt as the largest gas-fired power station in the country (seven GE Vernova turbines producing up to 4.4 gigawatts) feeding a 3,200-acre data center campus on the same transmission lines the coal plant used. The transmission capacity already existed. The scarce asset was generation, not wires.

The turbines took three years to get.

That gap - between how much power AI data centers need and how fast anyone can actually build it - is what this piece is about. Not because the AI buildout is slowing. It isn’t. The physical layer underneath it just hasn’t caught up yet.

The Same Mechanism, A Different Layer

The deeper you move into the stack, the fewer substitutes exist and the longer the timelines become. Policy can subsidize factories. It cannot accelerate qualification.

That framework applies to power with the same precision it applied to chips.

2,100 gigawatts currently sit in US interconnection queues - more than the total installed generating capacity of the entire US grid. Somewhere between 30 and 50 percent of planned 2026 AI data center capacity has already slipped to 2028 - not because the chips aren’t ready, but because the power connection is. Interconnection rights - the legal permission to connect a facility to the grid and draw firm, reliable load - are the scarce asset that capital alone cannot manufacture on a political timeline.

Washington already tried to fix this. In April 2026, grid infrastructure was designated a national defense priority under the Defense Production Act - the kind of move meant to fast-track exactly this problem. It’s the right call. But transformers still take more than two years to deliver, because policy can order urgency and still can’t manufacture the supply chain that urgency requires.

The law arrives before the thing it requires.

The Bypass and Its Own Constraints

If you can’t get a grid connection for years, the obvious move is to build your own power instead. That’s what behind-the-meter gas generation is - a power plant built right on-site instead of waiting on the utility grid - and it can be up and running in 18 to 24 months instead of years. Microsoft did exactly this in June: seven GE Vernova turbines for a Texas data center, built alongside Chevron. Homer City isn’t an isolated case. It’s the template everyone else is now following.

The order book already shows how fast this is moving. GE Vernova’s first quarter 2026 orders made the demand explicit: $18.3 billion, up 71% from a year earlier, with data center electrification orders in a single quarter exceeding all of 2025. Gas turbine backlog and slot reservations grew from 83 to 100 gigawatts in one quarter, prices are up 300% over three years, and available slots are already sold through 2028, with 2029 now under negotiation.

But the bypass has its own physical constraints. This is where the framework reasserts itself one layer deeper.

Gas turbines require yttrium. Specifically, yttrium oxide - which stabilizes the thermal barrier coatings inside turbine blades (the ceramic layer that allows components to operate at temperatures exceeding their own melting points). Without it, blade life collapses. Yttrium is a rare earth. The grid buildout and the critical minerals thesis are the same argument running at a different layer of the same stack.

The physical floor has its own physical floor.

The Materials Underneath

Copper runs through every layer of this buildout, more of it than most AI infrastructure models seem to have priced in. Bank of America estimates 60 to 75 tonnes of metal per megawatt of AI data center capacity - at 200 gigawatts by 2030, that’s 12 to 15 million tonnes of metal for data centers alone, over four years.

But copper’s constraint is not a story AI created. AI arrived while transportation, electrification, defense, and the energy transition were already competing for the same conductor. The US currently supplies roughly 86% of its own copper needs. By 2035, BloombergNEF projects that figure falling to 62% as demand compounds across every industrial buildout simultaneously. AI is not creating a new demand curve. It is colliding with every existing one at once.

Engineers are already trying to design around the copper problem, and it’s worth understanding why their fix doesn’t actually make the constraint disappear. AI data centers are shifting to 800-volt DC power, because higher voltage means less copper is needed per circuit. But that same shift deepens dependence on gallium nitride, a compound semiconductor used in the power conversion systems that make high-voltage distribution work in the first place. So the engineering solves one constraint and quietly creates another - less copper, more gallium nitride. That’s the pattern worth watching throughout this piece: one constraint compresses, and another expands right underneath it.

The transformer constraint is where this lands most visibly today, with delivery times now running above two years. Grain-oriented electrical steel - the specialized magnetic material inside every transformer the buildout requires, which allows electricity to be stepped up or down in voltage without excessive heat loss - has effectively one domestic integrated producer. Policy mandates domestically built grid infrastructure. The supply chain for the steel inside that infrastructure is not ready at the required scale.

Every Layer Becomes Industrial Capacity

Step back from the individual constraints and look at what they add up to. Chips ran into fabrication capacity. Fabrication ran into memory. Memory is now running into power, and power is running into the materials and industrial base needed to build it out. Each fix reveals the next unsolved layer beneath it, and every layer eventually becomes the same kind of problem - not software, not capital, but industrial capacity, with timelines measured in years and constraints that intent alone cannot compress.

The investors who recognized the chip constraint early captured the first layer of that value. The investors who recognized the memory constraint captured the second. The power constraint is the third layer now visible. Below it, already in view for anyone willing to look, are the materials and the industrial capacity to produce them.

The deepest constraints carry the longest timelines, the fewest substitutes, and the most durable pricing power.

The Situations

The framework identifies the gaps. The companies below sit inside them.

The question is not who benefits from the AI buildout, because nearly everyone in the stack benefits from the boom. The question is who benefits from the gap between the constraint and the resolution (the period between when the problem is visible and when the supply chain can actually address it). That gap is where pricing power concentrates.

The situations below are organized by proximity to the constraint itself.

These are not recommendations. The thesis and bear case for each is named. The reader decides.

POWELL INDUSTRIES (POWL, $232.21 as of July 24, 2026)

Start here, because most coverage does not.

Powell makes the equipment that sits between where the transformer ends and where the data center’s internal power begins - switchgear, power control rooms, electrical houses. It’s the layer that disappears inside the building the moment it’s installed and stays invisible until something goes wrong. Most people covering AI infrastructure have never written about it - it doesn’t show up in the GPU story or the power story, only in what connects them.

Every hyperscaler that builds behind-the-meter generation to avoid the interconnection queue still needs Powell’s equipment to route that power through the facility once it’s on-site. The bypass that avoids the grid doesn’t avoid Powell - it creates Powell’s order book.

The numbers back it up. New orders came in at $490 million last quarter, nearly double the prior year, and backlog reached $1.8 billion, up 33%. For every dollar of work completed, Powell signed $1.70 in new contracts, and the balance sheet carries no debt and $545 million in cash.

After the quarter closed, Powell landed a data center order exceeding $400 million - the largest in the company’s 70-year history. The company has roughly 3,000 employees. That order alone exceeds what a full year of prior quarterly revenue would suggest is manageable at their historical scale. They are going to have to grow into it.

At $232.21, the stock has pulled back roughly 29% from its all-time high of $328 while the order book has grown materially. Backward-looking, it looks expensive. Forward-looking - at what the next two years of backlog conversion implies - the picture is materially different. That is where the situation lives.

Bear case: Powell’s historical base was oil and gas, LNG terminals, and utilities. Data center is newer territory. The $400 million order is the largest single revenue concentration in company history. A delivery miss or cost overrun on a project of that scale - new territory for an organization of this size - lands hard on a stock priced for continued execution.

GE VERNOVA (GEV, $1,014.75 as of July 24, 2026)

GEV is the company that actually owns this bottleneck. Every hyperscaler that can’t get a grid connection needs a gas turbine instead, and at the scale this buildout requires, those turbines come from a short list of suppliers. GEV sits at the top of that list.

The most telling number in the second quarter print wasn’t the headline growth - it was who’s paying for what, and when. Orders came in at $24.2 billion, up 88% year over year, on revenue of $11.1 billion, up 22%. Gas turbine slot reservations grew from 100 to 116 gigawatts in a single quarter, with the company now targeting 125 gigawatts by year-end and backlog reaching $176.3 billion. But the number that actually matters is free cash flow: $5.1 billion, up from just $194 million a year earlier. That jump happened because hyperscalers are paying upfront to reserve turbine slots before the equipment even exists. That’s not a projection about future demand. That’s customers putting cash down today because they’re worried about not having a slot tomorrow.

Here’s what to watch in the next print: gas turbine orders in gigawatts. If bookings keep pace with the current run rate, that upfront cash keeps confirming real demand. If they slow, it means front-loading has outrun GEV’s ability to actually add production - a natural ceiling, since most capacity through 2030 is already spoken for.

The stock is up roughly 49% year-to-date and trades at more than double the sector average on forward earnings. That premium is really just the market agreeing GEV owns the physical constraint. The real question is whether the valuation already assumes the backlog converts at these higher prices - or whether there’s still room if it does.

Bear case: the wind segment is running at an operating loss guided around $400 million for the full year. It drags on overall profitability and competes with the stronger segments for management attention. A deceleration in backlog conversion timing would test a premium valuation that leaves limited room for disappointment.

EATON CORPORATION (ETN, $404.07 as of July 24, 2026)

If GEV generates the power, Eaton is the company that manages it once it exists - switchgear, circuit breakers, power distribution units, the equipment that takes high-voltage power and routes it safely to wherever it’s actually consumed. Every new behind-the-meter power source needs Eaton’s equipment before a single server in the building can draw from it.

Revenue was strong in the first quarter - a record $7.45 billion, up 17%. But that’s not the number worth paying attention to. The number that actually stops you is what’s underneath it: electrical orders from data center customers in the Americas were up 240% on a rolling twelve-month basis. Revenue tells you what already happened. Orders tell you what’s coming.

There’s also a catalyst the stock hasn’t fully priced in yet. Eaton is spinning off its vehicle powertrain business into a separate company, targeted for early 2027. What’s left behind is a cleaner, more focused electrification and power management business - data centers, grid infrastructure, aerospace power systems. Other industrial companies that have done similar separations have generally seen the remaining high-growth business re-rate once investors stop averaging it against a slower-growth segment. Whether that happens here, and when, is really what the current valuation is trying to price in.

Bear case: first quarter operating margins declined slightly, driven by integration costs across roughly $11 billion in acquisitions completed in a single quarter. Management targets significant margin improvement in the second half. Executing three major integrations simultaneously while the separation is underway is a lot to ask of any organization.

VERTIV HOLDINGS (VRT, $290.36 as of July 24, 2026)

Everyone else in this piece is focused on generating power or moving it. Vertiv is focused on what happens once that power is already inside the building. It makes the cooling systems that pull heat out of servers running at maximum load, the power protection systems that prevent outages, and the power distribution infrastructure inside the facility itself - a different layer from GEV and Eaton, but the same constraint logic underneath it.

The number worth paying attention to isn’t the revenue - it’s who’s betting on Vertiv directly. First quarter revenue came in at $2.65 billion, up 30%, with the Americas up 53%, backlog above $15 billion, and full-year guidance raised. But the more significant signal happened outside the earnings report: NVIDIA took a $2 billion equity stake in Vertiv and signed a multi-year supply agreement. A customer that takes an ownership position is telling you it plans to stay a customer, not quietly build a competing solution in-house.

At $290.36, the stock has pulled back roughly 24% from its 52-week high of $379.93 while the fundamentals have continued to strengthen. It trades at roughly 73 times trailing earnings, which means the market is already pricing what comes next rather than what’s already happened - and the backlog and order trajectory either justify that premium or they don’t.

Bear case: if major hyperscalers move toward building proprietary cooling and power distribution systems at scale, Vertiv’s addressable market compresses. The NVIDIA relationship is the strongest evidence against this scenario. It is one relationship, not a structural guarantee.

QUANTA SERVICES (PWR, $625.84 as of July 24, 2026)

Not the turbines. Not the switchgear. Quanta builds the actual grid - the transmission lines, substations, and physical interconnection infrastructure that policy keeps mandating and the supply chain keeps struggling to deliver.

This isn’t a diversified contractor that happens to have grid exposure. It’s a grid company, full stop - 82% of its revenue comes directly from electric infrastructure. And the backlog shows just how far behind everyone actually is: $48.5 billion, roughly eighteen months of revenue at the current run rate, and still growing, because for every dollar of work Quanta completed last quarter, it signed $1.60 in new contracts.

That’s the moat, and it took decades to build - accumulated engineering expertise, specialized equipment, trained crews, established contractor relationships. You can’t compress that by hiring more people or spending more money in the short term. Quanta has built it over time, and the backlog reflects that scarcity.

At $625.84, down roughly 21% from a 52-week high of $788.75, the stock sits well below analyst price targets ranging from $784 to $940. Whether that gap closes depends on how fast the backlog converts - which depends on permitting timelines and grid operator coordination outside Quanta’s direct control.

Bear case: the projects are signed. Physical execution still depends on regulatory approvals and grid operator coordination that Quanta cannot control unilaterally. Permitting delays do not move the backlog number. They move the conversion timeline.

FREEPORT-McMoRan (FCX, $62.60 as of July 24, 2026)

The long-term story on FCX hasn’t changed. What’s complicated is the next few quarters, and it’s worth being honest about both at once instead of picking a side.

FCX is the largest publicly traded copper producer in the Western world, and management has framed the company’s position explicitly around the electrification buildout - electricity equals copper - which the demand data genuinely backs up. But first quarter 2026 revealed a real operational problem at Grasberg, the company’s largest mine, in Indonesia. A mud rush back in September 2025 left more of the mine’s ore extraction points in wet, unstable condition than anyone expected. Fixing that means modifying equipment at each affected extraction point individually. It takes time. Capital doesn’t speed that up.

That’s why the company revised its copper production forecast down roughly 9% over the next five years.

But here’s why the structural argument still holds regardless: BloombergNEF projects US copper self-sufficiency falling from roughly 86% today to 62% by 2035, because AI is colliding with transportation electrification, grid buildout, and defense manufacturing - all competing for the same metal at the same time. That demand doesn’t depend on Grasberg running at full capacity. It was already going to happen.

So Grasberg is a timing problem, not a resource problem. The copper is still there. The engineering fix is already underway, a $700 million insurance recovery is agreed, and Indonesia’s operating rights are secured beyond 2041. The next few quarterly results will show whether the engineering timeline actually holds.

Bear case: the Grasberg modification timeline proves longer than estimated. Indonesia political risk around converting operating rights into final legal form. A meaningful slowdown in global industrial activity compressing copper prices before the mine ramp completes.

What the Bottleneck Teaches

Homer City is the argument made visible at a single address - a coal plant becoming gas turbines becoming a data center campus, on the same transmission lines, in the same Pennsylvania county. The whole transition of the American energy stack, happening on one property.

Every bottleneck solved exposes the next one beneath it. Fabrication became memory. Memory became power. Power is becoming materials. AI isn’t solving its bottlenecks - it’s digging downward through them.

This is how industrial revolutions spread through an economy - not by removing friction but by relocating it. The railroad required steel, and steel required ore. The internet required fiber, and fiber required glass. The pattern is older than the technologies it describes. The materials change. The structural logic doesn’t.

The semiconductor series demonstrated the framework at the fabrication layer. This piece extends it to the power layer. The materials layer is already visible beneath it for anyone willing to look.

The investor’s job is not to predict which technology wins. It is to identify where the constraint moved before everyone else notices. And to understand that the deepest constraints - the ones furthest down the stack, with the longest qualification timelines and the fewest substitutes - carry the most durable pricing power.

The bottleneck keeps moving, and the framework for finding it has to move with it. Tracking that - not the headline, the layer underneath it - is what The Chokepoint exists to do.

Share

Financial data sourced from company earnings releases, SEC filings, and public market data as of July 24, 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

The full archive is at williamdavid.substack.com.

July 26, 2026

The customs data shows zero. What it doesn’t show is the applications that were never submitted - and now, the reason why.

The Four Cliffs

Japan received nothing. The mechanism beneath the number just got sharper.

Chinese customs confirmed it for June: zero dysprosium, zero terbium, zero yttrium, zero gallium to Japan. Not rationed. Not delayed. Zero. Four simultaneous lines dropping to the floor and staying there across all four materials.

Paul Triolo, who tracks Chinese technology policy as closely as anyone, added the layer the customs data can’t show: Chinese suppliers won’t even try to apply for an export license if a Japanese company is involved. The chilling effect precedes the data by months. The market watches export volumes. The signal is the applications that never get filed.

This week, the reason behind that chilling effect became clearer.

On July 1, China quietly formalized something most Western analysts missed. MOFCOM Announcement No. 26 - a public reporting mechanism for strategic mineral export control violations. Dedicated hotline. Online portal. Anonymous tips accepted. Rewards available for verified reports. Voluntary self-disclosure is a mitigating factor. In June, a Chinese precision-optics company chairman was detained for falsely declaring germanium-bearing lenses as ordinary optical glass to sidestep licensing.

The grey-channel routing that once allowed materials to move around licensing - third-country transshipment, component disassembly, re-labeling - just got a lot riskier for everyone involved. The enforcement mechanism is now in place underneath it.

In May, two Japanese nationals employed by a major Japanese company were detained in Dalian for alleged export control violations involving rare-earth-related items. The Japanese government confirmed it in late June. At the time it looked like an isolated incident. The MOFCOM whistleblower system announced July 1 suggests it wasn’t.

“Won’t even try to apply” now has a corollary: “won’t even try to route around it.”

For investors: the companies already qualified inside the Western supply chain - MP Materials, Lynas, USA Rare Earth - aren’t competing against a tight market. They’re competing against a market that Japan has effectively been cut out of - and the workarounds are closing.

The License Is the Weapon, Not the Ban

November 10 is not an embargo date. That distinction matters more than the headline.

The market is pricing November 10 as a binary: controls resume or they don’t. The reality is more precise and more dangerous. China’s export control mechanism is a licensing regime. Selective. Reversible. Capable of targeting one country while leaving another untouched. The same shipment that clears for Malaysia routes through Lynas separation. A different shipment to a Japanese defense contractor never gets an application filed.

The yttrium price divergence makes the selectivity visible in a single number. Domestic Chinese yttrium oxide: approximately $7.88 per kilogram. European warehouse price: approximately $1,175 per kilogram. That gap isn’t a market inefficiency. It’s the price of the licensing regime expressed in dollars per kilogram.

The investment case for Western processors isn’t about what happens on November 10. The licensing architecture is already operating at full effect. November 10 changes the paperwork. The commercial reality is already here.

The West Is Building the Alternative. Scale Is the Question.

Lynas produced its first terbium oxide outside China in June 2026. The gap it’s filling is still enormous.

Lynas confirmed first terbium oxide production at its Malaysia separation facility in June 2026 - the first scaled terbium separation outside China, following dysprosium in May 2025. The circuit is rated up to 1,500 tonnes per annum of heavy rare earths.

To be clear: this matters. It’s also worth naming the scale mismatch directly. Japan received zero terbium from China through June. Lynas’s 1,500 tpa heavy rare earth circuit is a first proof point - not a replacement. The qualification clock applies: first production is not qualified supply for defense or semiconductor applications. That process takes 12-18 months minimum after first production.

The West is building the alternative. The question is whether the qualification timeline lines up with the enforcement timeline. Right now it doesn’t. That gap is the only clock that matters for the investment case.

The common thread

Whether it’s the zero shipments, the enforcement mechanism now operating beneath the licensing regime, or the scale mismatch between Lynas’s first terbium oxide and Japan’s zero imports, the constraint isn’t political. It’s already embedded in the physical supply chain. Policy can respond. The qualification clock doesn’t negotiate.

Three things I’ll be watching next week

AXTI Q2 earnings July 30 - first earnings print since the indium price inversion confirmed. Any commentary on MoC license delays or order backlog movement is the signal.

MP Materials Q2 earnings July 30 - NdPr pricing commentary against the $110/kg DoD floor price contract. If spot is diverging from the floor, that’s the qualification moat pricing argument in real time.

November 10 countdown: 107 days. Watch for any MOFCOM guidance on extractant control resumption - and whether the whistleblower system generates any publicly reported enforcement actions before then.

The clock

November 10 - REE extractant controls resume.
November 27 - gallium truce expiry.
January 1 - DFARS 252.225-7052.

107 days to extractant controls. 126 days to gallium truce expiry. 161 days to DFARS.

Live critical mineral prices at scarceearth.io.

The full archive is at williamdavid.substack.com. The Week in Minerals publishes every Saturday.

Prices directional from SMM domestic China benchmarks and ScarceEarth.io admin values as of July 24, 2026. Paul Triolo intelligence confirmed July 24, 2026 via X. MOFCOM Announcement No. 26 sourced from Chinese Ministry of Commerce, effective July 1, 2026. Japanese nationals detention confirmed by Japanese government via Reuters, late June/early July 2026. Lynas terbium oxide production confirmed via mainrichinternational.com July 2026 update. China-Japan export data sourced from General Administration of Customs China via Reuters, confirmed through June 2026. Yttrium price divergence sourced from CMP benchmark and ScarceEarth.io admin values.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

Nvidia is down. The frontier labs are heading toward public markets at eye-watering valuations. A Chinese AI lab just released a model that rivals the best American ones at a fraction of the cost. If you’re trying to make sense of what’s happening - and what it means for where to put capital - most of what you’ve read hasn’t explained the part that actually matters. This post does.

Start with what everyone is looking at.

Nvidia stock has sold off. The explanation circulating: AI models are getting more efficient, which means less demand for chips, which means less revenue for Nvidia. Meanwhile Anthropic and OpenAI are reportedly preparing for public markets at valuations in the hundreds of billions. And Kimi K3 - a Chinese AI model released last week - just demonstrated that a non-American lab can match the best frontier models, openly, at a fraction of the cost.

Those three facts feel like they tell a coherent story. Cheaper AI, less hardware demand, American AI companies racing to go public before the window closes. Sell the chips, maybe buy the labs before they list.

That story has a problem. It’s built on a misunderstanding of where the constraint in this buildout actually lives. And the misunderstanding starts with something almost no coverage of the AI trade bothers to explain: what the hardware actually does, and why certain parts of it cannot be made cheaper by software improvements.

The AI buildout runs on a stack - a set of layers, each one resting on the one beneath it. The application you use sits at the top. The materials required to build the whole thing sit at the bottom. Each layer has its own constraints, its own suppliers, its own timelines. What’s happening at the top - models getting cheaper, labs going public - is real. It’s just not the whole picture. The diagram below maps the full stack. Everything in this publication fits somewhere on this diagram. If you understand the diagram, you understand how the rest of the series connects.

Start at the bottom and work up.

What a chip actually is

A semiconductor chip is a piece of silicon covered in billions of tiny switches. Silicon is useful because it can be made to conduct or resist electricity depending on conditions - that controllable property is what makes computation possible. Those switches - transistors - turn on and off billions of times per second. That switching is computation. Everything an AI model does, every word it generates, every question it answers, is the product of those switches firing in sequence.

The switches have to be impossibly small. The most advanced chips today are built at dimensions measured in nanometers - billionths of a meter. A human hair is roughly 80,000 nanometers wide. The transistors in a leading-edge chip are a few nanometers across. Building them requires machines of almost incomprehensible precision, chemicals of extraordinary purity, and process knowledge accumulated over decades. There is no shortcut. You cannot will a chip fabrication facility into existence with capital alone. The knowledge has to be built, qualified, and proven - and that takes years regardless of how much money goes in.

Why AI runs on a specific kind of chip

Not all chips are the same. The processor in your laptop - a CPU - is optimized for sequential tasks. It does one complex thing after another, very quickly. A GPU was originally designed for graphics rendering, which requires doing millions of simple calculations simultaneously rather than one complex calculation at a time.

AI training and inference happen to require exactly that: massive parallel computation. The GPU became the engine of the AI buildout almost by accident. Nvidia happened to be building the best ones, and its software platform created a switching cost that competitors have spent years trying to overcome.

When a company announces it is spending billions on AI infrastructure, most of that money is going toward clusters of GPUs. Nvidia’s chips - each costing $25,000 to $40,000 - are the workhorses of the buildout. A single large AI training run might use tens of thousands of them simultaneously.

Why one company in Taiwan makes most of them

Here is the fact that most AI coverage buries or ignores entirely: Nvidia does not manufacture its own chips. Neither does AMD, Apple, Google, or most of the companies designing cutting-edge silicon. They design the blueprint and send it to a foundry - a specialized manufacturer that does nothing but fabricate chips.

Taiwan Semiconductor Manufacturing Company is that foundry. TSMC manufactures approximately 90% of the world’s most advanced chips. The GPU powering most AI infrastructure today was designed by Nvidia and built by TSMC. The next generation will be too. There is no meaningful alternative at the leading edge.

TSMC sits on an island 180 kilometers from the Chinese coast. Its most advanced manufacturing capacity - the facilities that can build the chips AI requires - is concentrated in Taiwan. Billions are being committed to build new facilities in Arizona. Arizona is also approximately 2% of TSMC’s current revenue, and its most advanced facilities won’t reach production until 2028 at the earliest.

The constraint isn’t just where TSMC is located. It’s what TSMC does that nobody else can replicate at scale - and the years it takes to qualify any alternative.

Why a chip needs more than fabrication

A chip doesn’t go straight from the fabrication facility to a data center. It has to be packaged - physically connected to memory, power delivery, and other chips in a way that lets them communicate fast enough to be useful.

This is the part of the AI buildout that receives almost no coverage, and it is where some of the most important constraints live.

CoWoS is TSMC’s advanced packaging process. It physically integrates high-bandwidth memory with a GPU on the same silicon platform - a thin layer that allows the two to communicate at speeds a conventional circuit board cannot match. Without this integration, the most powerful AI chips cannot function at their rated performance. The memory and the processor need to be that close to each other, physically, for the data transfer speeds that AI inference requires.

TSMC is the only commercial supplier of this packaging at scale. It is fully allocated through 2027. When TSMC’s CEO said on the most recent earnings call that packaging capacity is “so tight it is limiting my customers’ growth,” he was describing a physical constraint that no amount of capital can quickly resolve. You can announce a new facility. You cannot qualify one in a quarter.

The memory the chip depends on

The memory that sits alongside the GPU in that package is not the same memory as in your laptop. Regular memory is fast enough for most computing tasks. AI inference at scale requires something faster - memory that can transfer data to the processor almost instantaneously, in massive volumes, without lag.

High Bandwidth Memory - HBM - stacks multiple memory chips vertically, connects them through microscopic channels, and places them directly beside the GPU. It is physically inseparable from the chip it serves. Three companies make it at commercial scale: SK Hynix, Micron, and Samsung. SK Hynix holds approximately 58% of the market.

Here is the detail that matters most: switching HBM suppliers requires redesigning the entire chip package from scratch. The memory and the GPU are co-designed. A chip built around one manufacturer’s HBM cannot simply swap in another’s without an 18-month redesign cycle minimum. That is not a commercial negotiating position. That is a physical constraint.

This is the layer the Semiconductor & AI Hardware Series spends the most time on. If you want the full argument - why the shortage persists through 2027, why the switching cost is structural, and what the investment implications are - that series is where it lives.

What a model actually needs to run

This is where everything above connects to the news that sent Nvidia lower and left many investors confused.

When an AI model generates a response, it draws on everything it learned during training. That accumulated knowledge is stored as numbers - billions or trillions of them - called weights. They encode, statistically, how language works: what words follow other words, how concepts relate, what a coherent answer looks like. Every response the model generates is the product of those weights doing work.

All of those weights have to be loaded into HBM and kept there the entire time the model is running. The model cannot predict which weights it will need next, so all of them have to be within reach at all times.

Think of it like a chef. A great chef needs two things: the skill to cook, and the ingredients within reach. Compute is the cooking. The weights in memory are everything on the counter.

You can make the chef faster. You cannot make a meal with ingredients that aren’t there.

What Kimi K3 actually showed

The important point isn’t the architecture itself. It’s that software efficiency and hardware requirements are not the same thing.

Kimi K3 uses an architecture called Mixture of Experts. Instead of activating all 2.8 trillion weights for every response, the model routes each piece of a question to a specialized subset - roughly 104 billion weights at a time. Less active computation per response. A more efficient model.

The market drew a conclusion: models are getting more efficient, chip demand will fall.

Here is what the market missed. All 2.8 trillion weights still have to be on the counter. Kimi K3 requires several terabytes of HBM to exist in a runnable state - regardless of how efficient each individual response is. When Kimi K3 paused new subscriptions 48 hours after launch because demand overwhelmed capacity, it was not a compute problem. It was a memory problem. The model needed more HBM than it had available.

More efficient model. Same memory requirement. Different hardware. Different constraint.

The efficiency argument compresses the compute bill. It does not compress the memory bill. Those are different chips, different supply chains, and they are not moving in the same direction.

Why the frontier labs aren’t where the constraint lives

Which brings us to Anthropic, OpenAI, and what to make of these companies when they eventually reach public markets.

The frontier labs have built the most capable AI systems in the world and are genuinely changing how businesses work. They are also sitting on top of a supply chain they don’t control.

OpenAI trains on TSMC chips. Anthropic runs inference on servers full of HBM. Neither manufactures the hardware. Neither controls the materials. If CoWoS packaging capacity tightens - and it has - they pay more and wait longer. If HBM is allocated to Nvidia first - and it is - they buy what’s left. Their compute bills are enormous. The hardware that determines how much they can grow and what it costs is controlled by companies they have no ownership in.

The frontier labs are the most visible part of the AI buildout. They are not the most constrained part. The most constrained parts - the fabrication layer, the packaging layer, the memory layer - tend to have more durable pricing power than the application layer above them.

This is not an argument against the frontier labs. It is an argument for understanding where in the stack the constraint actually lives. The market spends most of its time pricing what’s visible. The more durable question is what makes the visible part possible.

The layer underneath everything

One more thing.

Every layer described above depends on materials that are difficult to produce, geographically concentrated, and in several cases controlled by China through export licensing. Tungsten is in the contacts of every memory chip. Gallium - under Chinese export control - goes into the power electronics that run data centers. Indium phosphide is at the heart of the optical connections that move data between chips at the speeds AI infrastructure requires.

The semiconductor story and the critical minerals story are the same argument at different layers of the same stack. The chip constraint is real. The constraint underneath the chip is also real. And the further down the stack you go, the fewer substitutes exist and the longer the timelines become before anything changes. That argument runs through the Critical Minerals Series at the full archive.

What to do with this

This post is not a recommendation. It is a map.

The AI buildout is the largest capital deployment in modern industrial history. Most coverage focuses on which model is winning, which lab is ahead, which application is growing fastest. Those are real questions. They are not the only questions.

The questions that tend to matter more over time: where does the constraint live? Who controls the layer that everything above it depends on? What would have to be true for that constraint to resolve - and how long would that take?

Next time you read that AI is getting cheaper, ask which layer is getting cheaper. The model layer might be. The memory layer isn’t. The packaging layer isn’t. The materials layer isn’t.

Those are different parts of the same buildout. They are not all going in the same direction.

The Semiconductor & AI Hardware Series and the Critical Minerals Series explore each of these layers in depth.

Ready to go deeper? The Semiconductor & AI Hardware Series starts here:

The full archive is at williamdavid.substack.com.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

Share

This is the closing post of the Semiconductors & AI Hardware Series. The series documented the physical constraint layer underneath the AI buildout. This post follows the money.

This post is the first edition of a quarterly format. The situations named here reflect where the physical constraint intersects a valuation dislocation today. When the constraint moves - and it will - the list moves with it.

Two things happened in the same 48-hour window last week.

ASML reported €9.3 billion in Q2 revenue - above the ceiling of its own guidance. TSMC reported $40.2 billion, a 77% year-over-year profit increase, raised full-year guidance to $43-45 billion, and committed another $100 billion to Arizona. C.C. Wei said on the call that packaging capacity is “so tight it is limiting my customers’ growth.”

And chip stocks went down.

That divergence - record earnings, declining prices - is the starting point for this post. Not because it signals a turn. Because it tells you which layer the market has been pricing and which layer it hasn’t.

The framework the series built doesn’t produce dozens of ideas. It eliminates most of them. What remains after the framework runs are the situations where the bottleneck and the valuation have begun moving in different directions. The pullback created more valuation dislocations than existed three months ago.

The series made four arguments. None of them have changed.

First: the AI buildout depends on a fabrication and packaging layer concentrated in a single geography. TSMC is 90% of leading-edge logic, 180 kilometers from the Chinese coast, with CoWoS packaging as the sole commercial supplier at scale. The $265 billion Arizona commitment is the right answer to that problem. It runs on a 5-7 year construction timeline. The constraint is still where the series said it was.

Second: HBM is not a commodity. Switching suppliers requires redesigning the chip package from scratch. Three suppliers exist. SK Hynix's guidance is that demand exceeds its supply capacity beyond 2030. Micron has declined to name a date at all, saying it has no line of sight on when supply catches up. The materials that build HBM - tungsten contacts in every DRAM die - route through the same licensing architecture the series documented at the indium layer.

Third: the policy layer is aimed at the wrong level. The MATCH Act targets equipment. China is supplying 35% of fab equipment domestically. The qualification clock runs 2-5 years after any restriction takes effect. Policy can disrupt. It cannot resolve the constraint on the timeline the law assumes.

Fourth: inference efficiency is compressing at the application layer. Open-weight models are absorbing enterprise adoption. The $0.50 barrel of intelligence is real. But every cheap derivative still traces to a training run that needed leading-edge silicon - and the training layer still routes through the physical constraint the series documented.

One thing has changed since Part 1. Intel’s 18A is in production on Panther Lake. EMIB-T packaging is matching CoWoS yields. The monoculture has a credible challenger for the first time in years. The timeline remains multi-year. The direction is no longer purely hypothetical. Both things can be true.

Ten situations follow. Organized by where they sit in the stack.

One note before proceeding: these ideas are intentionally concentrated around one thesis. If the physical-constraint framework is wrong, several of these will likely underperform together. This is not a diversified list. It is ten expressions of the same underlying argument.

These are not recommendations. The thesis and bear case for each is named. The reader decides.

FABRICATION

TSMC (TSM, $398.37 as of July 17, 2026)

The series argued TSMC in four elements: 90% concentration of leading-edge logic, Arizona at approximately 2% of revenue, CoWoS as sole-supplier chokepoint fully allocated through 2027, and geographic risk 180 kilometers from the Chinese coast. Every element was confirmed on the July 16 earnings call. Nvidia alone secured approximately 60% of TSMC’s CoWoS output for 2026, leaving remaining AI and ASIC customers competing for what’s left.

Since the call: reports in early August describe TSMC expanding outsourced CoW volume to ASE and SPIL, beyond the WoS work already routine at those partners. Confirmed on July 16 was TSMC's allocation share. Not yet confirmed at the time was how much of that allocation TSMC would keep executing itself.

The stock being flat on record earnings tells you the demand narrative was already priced. The $60-64 billion capex raise - an $8 billion increase in a single quarter - is not a forecast. Companies do not commit capital at that scale without multi-year customer agreements already locked. C.C. Wei said demand will remain strong “all the way to probably 2029, 2030.” That is the longest demand visibility window the company has ever provided publicly.

On the Arizona question: the AP1 and AP2 advanced packaging facilities are targeted for mass production in 2028 at the earliest. Until they are online, every AI accelerator that uses TSMC’s most advanced packaging still routes through Taiwan. Arizona diversifies production. It does not yet diversify capability.

Bear case: Taiwan cross-strait escalation before AP1 and AP2 reach commercial scale. Intel’s EMIB-T packaging, if it continues improving, provides the first credible alternative to CoWoS at the layer that currently binds before fabrication.

Next confirmation: Monthly revenue releases from TSMC investor relations - available within the first two weeks of each month. The single most important real-time data point in the series.

ASML (ASML, $1,747.58 as of July 17, 2026)

The two-sided thesis was confirmed and sharpened by Q2. The bull side: full-year guidance raised to €43-45 billion, 2027 EUV orders nearly fully received, 30% capacity expansion planned for 2027. The risk side: China at 14% of system sales in Q2 versus 19% in Q1. The specific exposure the MATCH Act targets is already compressing without the law passing.

A substantial portion of every leading-edge capacity expansion eventually routes through ASML’s order book before it appears in ASML’s results. That’s a forward order embedded in someone else’s guidance. TSMC’s $60-64 billion capex raise is the clearest expression of that dynamic: equipment orders already locked, revenue recognition to follow.

Meanwhile, Chinese chipmakers are accepting ASML’s price increases while TSMC resists. Neither can walk away. That’s the monopoly confirmed in a single data point.

Bear case: MATCH Act passes and triggers DUV export restrictions before allied demand absorbs the revenue gap. The H2 China system sales figure - implied at approximately €5.9 billion to hit 20% of the full-year guide - does not materialize if export controls tighten mid-year. That’s the number management didn’t explicitly state on the call but that the arithmetic requires.

Key confirmation event: Q3 results and China revenue as a percentage of system sales. If that figure continues compressing toward 10%, the MATCH Act risk is already resolving without legislation.

Onto Innovation (ONTO, $279.85 as of July 17, 2026)

Part 3 documented the CoWoS packaging constraint in detail. ONTO makes the inspection and metrology tools that qualify and monitor that constraint. The Dragonfly G5 system targets 2.5D and 3D chiplet packaging inspection - the fastest-growing segment of semiconductor equipment as advanced packaging replaces node scaling. Management projected 2026 revenue growth above 30%, driven by advanced packaging. Every packaging line in Arizona’s AP1 and AP2 facilities will require process control tools before it enters production.

Why ONTO over KLA? KLA is the dominant name across all inspection categories. ONTO is more concentrated in advanced packaging specifically - which is where the growth is - at a substantially lower valuation. ONTO trades at a meaningful discount to KLA on EV/Revenue despite competing directly in the fastest-growing segment. Camtek (CAMT) covers similar ground, primarily in bumping and advanced packaging inspection, but with more exposure to memory packaging than the logic-and-CoWoS focus that makes ONTO’s Dragonfly positioning relevant to the specific constraint the series documented.

Bear case: KLA expands its advanced packaging inspection suite and absorbs the CoWoS inspection opportunity before ONTO establishes dominant design wins. The competitive question - whether packaging inspection becomes winner-take-most or remains competitive - is a business risk more than a macro risk.

Catalyst to watch: Q2 earnings August 6 and advanced packaging backlog as a percentage of total order intake.

MEMORY

SK Hynix (SKHY, $154.03 as of July 17, 2026)

The series argued the HBM constraint in two sentences: no substitute exists, and switching suppliers requires redesigning the chip package from scratch. SK Hynix holds 58% of the market and delivered HBM4 to Nvidia nearly a year ahead of Samsung. The Nasdaq listing on July 10 removed friction that kept US institutional capital out of this position for 14 years.

The tension is structural, not just timing. SK Hynix sells HBM under long-term supply agreements at pre-negotiated contract prices. When conventional DRAM spot prices rose approximately 30% quarter-over-quarter in Q2, fixed HBM pricing dragged SK Hynix’s blended ASP below what a fully spot-priced product mix would have delivered. This isn’t a Q2 problem. It is the permanent arithmetic of a company whose dominant product trades at pre-negotiated terms. The bull case is that HBM4 contracts will be negotiated at higher price points as older agreements roll off. That’s a 2027 story.

Q2 earnings land July 29 on a stock with no US trading history and no established institutional base. On July 13, the stock fell 15.4% in Seoul on a single broker note. On July 14, it reversed entirely. That volatility is the price of the position.

Bear case: LTA pricing structure continues to cap upside capture in an upcycle while competitors on spot pricing benefit more directly. 600%+ appreciation on the Korean-listed shares over twelve months compresses the margin for error. DFARS clock begins January 1, 2027 - SKHY is Korean, which cuts against the geopolitical diversification thesis for defense-adjacent customers.

Catalyst to watch: Q2 earnings July 29. HBM4 volume ramp guidance and any commentary on 2027 contract pricing - that’s where the LTA constraint begins to ease.

Micron Technology (MU, $848.95 as of July 17, 2026)

The results make this the strongest fundamental situation on the list. Q3 FY2026: revenue $41.46 billion, up 346% year-over-year. Gross margin 84.9%. EPS $25.11. Q4 guided to $50 billion revenue and 86% gross margins - a further acceleration. HBM3E and HBM4 are fully booked through calendar 2027, with demand extending into 2028. The CEO said Micron can only fulfil 50-66% of customer demand in the medium term. The company generated as much cumulative cash flow in the last two quarters as in its entire prior history.

The geopolitical argument has strengthened since Part 1. DFARS-qualified fabs in Idaho and New York. $22 billion in strategic customer agreements with approximately $18 billion in cash deposits already received. Ford supply agreement signed July 6. New York fab pouring first concrete July 9. $3 billion additional strategic investment announced July 9. Korean fabs face energy cost shocks and helium supply constraints that Micron’s domestic fabs do not - though domestic manufacturing reduces geopolitical risk without eliminating global supply chain exposure entirely.

The stock is down materially from recent highs on broader semiconductor sentiment. The fundamentals have not moved in the direction the stock has.

Bear case: CXMT bonded DDR5 yields continue improving and qualify for datacenter workloads at scale - memory architecture substitution at the server layer, not inference efficiency, is the scenario that breaks the shortage thesis. At 84%+ gross margins, the stock is pricing a cycle that has limited room for disappointment.

Catalyst to watch: Q4 FY2026 earnings in September. Any update on 2028 demand visibility would extend the structural case beyond what management has already guided.

MATERIALS & INFRASTRUCTURE

Entegris (ENTG, $138.74 as of July 17, 2026)

The series spent four parts arguing that the constraint isn’t at the chip - it’s at the chemistry and materials one layer beneath it. Entegris is the company that most directly embodies that argument.

Entegris supplies ultra-pure process chemicals, advanced filtration systems, and specialty materials to semiconductor fabs. Every domestic fab - TSMC Arizona, Micron Idaho and New York, Intel Ohio - requires Entegris materials. The qualification process for process chemistry runs 12-18 months and cannot be compressed. A new fab that opens in Arizona cannot run its processes until the chemistry is qualified. Entegris is the qualification gate that sits inside every CHIPS Act announcement - and the first supplier to finalize its CHIPS Act award agreement.

Management stated in Q1 2026 that “materials intensity and process complexity continue to increase” - the direct mechanism by which content per wafer grows with every node transition. The company articulates a 24-month demand cycle: fab construction products ramp first, then WFE and filtration, then unit-driven business. Memory is currently in wave 1. Advanced logic is transitioning through waves 2 and 3.

Bear case: AI capex digestion causes customers to defer chemical procurement, compressing near-term revenue before the CHIPS Act fab ramp absorbs the volume. The company is navigating a CFO transition (new CFO Sukhi Nagesh replacing the departing CFO) which adds execution uncertainty at the leadership level.

Catalyst to watch: TSMC Arizona Phase 2 equipment installation confirmation - Entegris materials qualification must precede any volume production ramp. Each new domestic fab is effectively a new multi-year Entegris customer.

AXT Inc (AXTI, $45.86 as of July 17, 2026)

This is the highest-risk, most asymmetric situation on the list - and the one most directly traceable to the physical argument made in Part 2.

AXT’s Tongmei subsidiary is the primary Western-aligned supplier of indium phosphide substrates - the foundational material for optical transceivers, silicon photonics, and 5G components. Every InP wafer that ships for these applications routes through their substrate production. The indium licensing gate Part 2 documented - China’s Ministry of Commerce licensing every export shipment - sits directly in their supply chain as both a risk and a structural barrier to entry for anyone trying to replicate their position.

On July 2, 2026, AXT signed a $22 million supply agreement with a major optical transceiver customer. That is the first confirmed commercial-scale validation of the indium licensing chain argument from Part 2. The Tongmei IPO pivot from Shanghai to Hong Kong removes the primary valuation overhang on the Chinese subsidiary.

This situation stacks multiple conditions: InP licensing stability, optical demand acceleration, Tongmei IPO success, continued Western customer preference, and manageable political risk. All can hold. That is a different statement than saying all will. This is the most asymmetric name on the list, not necessarily the highest conviction.

Bear case: China tightens MoC licensing on InP exports specifically - the exact leverage point Part 2 documented becomes the investment risk. A negative regulatory decision on the Tongmei Hong Kong IPO removes the subsidiary valuation catalyst.

Catalyst to watch: Tongmei Hong Kong IPO filing timeline. Additional supply agreements confirming commercial-scale InP substrate demand from Western optical customers.

Coherent Corp (COHR, $277.60 as of July 17, 2026)

Part 2 documented the indium licensing gate - China’s Ministry of Commerce licensing every InP export shipment. The CEO of Coherent flew to Beijing personally to raise delays. That constraint sits in Coherent’s cost structure as a risk. The stock is down 38% from its all-time high.

What the pullback prices: execution risk, InP supply dependency, and the broader optical name repricing. What it doesn’t change: Q3 FY2026 showed record bookings, backlog visibility extending to calendar 2028, and 6-inch InP wafer yields now exceeding the 3-inch lines they replaced. The execution concern that was legitimate twelve months ago has largely resolved. NVIDIA confirmed it by taking a $2 billion equity stake and signing a multiyear CPO supply agreement - the most significant commercial validation in Coherent’s history.

The more interesting argument: silicon cannot generate light. As silicon photonics scales, InP demand grows with it rather than being displaced by it. Coherent’s integration of the InP substrate layer through DustPhotonics positions it to capture the demand growth from the technology that appears to be competing with its existing products. The company plans to double InP output capacity by next quarter and more than double again by end of 2027.

Bear case: CPO deployment timeline slips into 2028, delaying the revenue ramp the NVIDIA agreement implies. The execution risk has reduced. The demand timing risk has not.

Catalyst to watch: Q4 FY2026 earnings August 13. CPO revenue as a percentage of total datacom sales and any update on the NVIDIA supply agreement volume timeline.

ASM International (ASMIY, $995.94 as of July 17, 2026)

The series named ASML as the unreplaceable equipment layer. ASM International sits one layer deeper in the same argument - at the atomic layer deposition process that makes the gate-all-around transistor transition physically possible.

At 2nm and below, gate-all-around architecture is the structural requirement. ALD is the process that enables it - applying ultra-thin material coatings one atomic layer at a time with the precision that no other deposition technique can match at this geometry. ASM International holds the dominant ALD market position. There are no legacy process substitutes at the 2nm node. The qualification timeline for ALD chemistry and tooling runs years, not quarters.

Q1 2026 revenue was €862.5 million - at the high end of guidance, with a record adjusted operating margin of 33.1%. Management described gate-all-around as “shaping up to be a large node” and confirmed Q2 is guided to €980 million with H2 expected stronger than H1. The SAM uplift at 1.4nm is expected to be even larger than what ASMIY captured at 2nm. Pilot line investment for 1.4nm begins later this year.

Less followed than ASML. Less covered than KLAC. Positioned at a layer the market has not yet priced with the same precision it has applied to lithography.

Bear case: Gate-all-around adoption timeline slips as leading-edge logic customers delay node transitions on cost grounds. MATCH Act export restrictions reduce China revenue. The OTC ADR structure (ASMIY) means lower liquidity than exchange-listed names.

Key confirmation event: TSMC N2 and Intel 18A volume production ramp confirmation - both require ALD at the gate-all-around layer. Any pilot line order disclosure for 1.4nm would validate the SAM expansion thesis ahead of schedule.

Linde plc (LIN, $513.22 as of July 17, 2026)

Nobody in semiconductor coverage is writing about Linde as a semiconductor play. That classification gap is the opportunity.

Every semiconductor fab runs helium continuously for wafer cooling, leak detection, and process chambers. Qatar’s Ras Laffan complex - approximately 33% of global helium supply - sustained damage that will take 3-5 years to repair per QatarEnergy. China banned helium exports July 10. Fabs entered 2026 with emergency buffer stocks that are depleting.

Linde holds long-term supply contracts with semiconductor fabs globally. Q1 2026: EPS up 10%, electronics segment up 10% driven by AI chip investments. Management described the helium market as facing “acute global shortages” and confirmed it is prioritizing long-term agreements and anticipating continued price increases throughout the year. The detail that matters most: full-year 2026 guidance was set with no implied helium upside. Every helium pricing improvement is incremental to the numbers already guiding the stock.

Why Linde over Air Products or Air Liquide? All three benefit from helium supply disruption. Linde is the largest industrial gas company by market share and revenue, with the deepest semiconductor fab contract exposure globally. Air Products has a stronger hydrogen thesis for the longer-duration energy transition play. Air Liquide carries more European fab exposure. For the specific helium-as-semiconductor-constraint argument the series implies, Linde is the most direct expression.

Bear case: Ras Laffan repair timeline accelerates beyond current guidance, restoring helium supply before pricing power compounds. US-Iran diplomatic resolution restores Hormuz stability and removes the geopolitical premium from industrial gas pricing.

Catalyst to watch: Q2 earnings July 31 and helium segment pricing commentary. Guidance already excludes helium improvement - any upside statement from management is additive to consensus.

How to Read This List

Not all ten situations carry equal conviction or risk. A loose grouping helps:

Highest conviction - where the physical constraint is most observable, substitution most remote, and the thesis requires the fewest additional assumptions:

TSMC, ASML, Micron, Entegris, Linde.

High conviction - where the structural thesis is sound but timing, competitive dynamics, or execution add meaningful uncertainty:

SK Hynix, ONTO, ASM International, Coherent.

Highest asymmetry - where the framework fits but multiple conditions must hold simultaneously:

AXT.

This is not a ranking. It is a map of where different risks live within the same underlying argument.

What Would Make This Wrong

Every thesis in this list depends on one assumption: that qualification remains slow and substitution remains difficult. If qualification timelines compress dramatically, if advanced packaging becomes commoditized faster than expected, or if a new manufacturing architecture removes today’s bottlenecks entirely, this framework will need to change.

Chokepoints don’t disappear overnight. But they do move. The series already acknowledged Intel’s EMIB-T as the first credible challenge to the CoWoS monoculture. That’s the template for how bottlenecks migrate: not sudden obsolescence, but gradual movement of the binding constraint to a different layer. The job is to track where the constraint moves, not to assume it stays put.

That’s why this list gets revisited every quarter. The framework is permanent. The names are not.

Readers may notice what isn’t on this list.

NVIDIA isn’t here. AMD isn’t here. Broadcom isn’t here.

That’s intentional.

Great products are not always great bottlenecks. The framework the series built starts where substitution becomes difficult and qualification takes years. That’s where pricing power tends to last longest - not because the product is better, but because the alternative doesn’t exist yet, or costs more time than the market has patience for.

The market spends most of its time pricing products. This publication spends its time pricing constraints. Over long periods, the second often matters more than the first.

The series begins with Semiconductors & AI Hardware, Part 1: The Compute Layer.

The full archive is at williamdavid.substack.com.

Share

Financial data sourced from company earnings releases, SEC filings, and public market data as of July 17, 2026. Stock prices sourced from live market data, closing prices July 17, 2026: TSM $398.37, ASML $1,747.58, ONTO $279.85, SKHY $154.03, MU $848.95, ENTG $138.74, AXTI $45.86, COHR $277.60, ASMIY $995.94, LIN $513.22. All prices require live verification before any investment decision. References to specific price levels reflect conditions at time of writing and should not be relied upon as current.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

July 18, 2026

Chinese customs data shows Japan has received zero dysprosium or terbium since November. Eight months. Still zero. The same week, Japan signed a trilateral agreement with the US and Australia to fund gallium production on Australian soil. One country. Two moves. Both pointed in the same direction.

Gallium
The first credible Western gallium facility just got funded - 133 days before the truce expires.

Alcoa’s final investment decision landed July 14. A new gallium production plant at Alcoa’s Wagerup alumina refinery in Western Australia. $307 million. 100 metric tons per year - roughly 10% of current global gallium supply. US-Australia-Japan trilateral framework. Production targeted for late 2026.

The truce suspending China’s most restrictive gallium export controls expires November 27. That is 133 days. The Alcoa plant will not be producing gallium at scale in 133 days. What the FID does is start the clock on an alternative that didn’t exist before. The chokepoint hasn’t moved. The trajectory has.

Alcoa (AA) is now a gallium story in a way it wasn't three months ago. The FID creates a direct line between the trilateral framework and a publicly traded company with existing alumina infrastructure. That's a different investment profile than a junior miner starting from scratch.

Rare Earths
Eight months of zero dysprosium and terbium exports to Japan. The duration is the signal.

That’s not a trade disruption. That’s a sustained stockpiling condition. The two heavy rare earths at the center of China’s export control crisis haven’t moved to Japan since November. Draw your own conclusions.

Meanwhile the price tape reflects what the silence implies. Terbium is up more than 20% month-over-month. Dysprosium is the outlier - down roughly 5% - but it’s the only name moving against the trend in a rare earth complex where NdPr and germanium are both moving as though Beijing has already cut supply. China’s H2 mining quota announcement still hasn’t arrived. It usually does in June or July. The market is pricing the silence.

The companies sitting downstream of that eight-month zero-export condition include every manufacturer that uses permanent magnets for motors, actuators, and precision drives. Coherent and AXT sit at the InP substrate layer that shares the same licensing architecture the series documented. The silence in Japanese customs data is already in their supply chain.

IEA Outlook
The argument has shifted. The IEA just confirmed it.

The IEA’s Global Critical Minerals Outlook 2026 dropped Thursday. The framing matters more than the numbers: concern has shifted from whether supply can meet demand to vulnerability from concentrated supply chains - specifically processing and refining. That is the Mine Is Running argument in primary-source language from the world’s most cited energy authority. Strategic minor mineral prices more than doubled. Tungsten surged sixfold. The IEA estimates China’s expanded rare earth export controls put approximately $6.5 trillion of downstream production at risk across automotive, defense, energy, and high technology. The IEA now has a supply risk framework specifically for aerospace and defense materials.

The market has been treating this as a mining story. The IEA is calling it a chemistry story.

The IEA put a number on what the companies already qualified inside the Western supply chain are sitting on - $6.5 trillion of downstream production at risk from China's export controls. Entegris, Linde, MP Materials. The qualification moat just got a dollar figure attached to it by the world's most cited energy authority.

The common thread.

Whether it’s gallium, dysprosium, terbium, or tungsten, Japan’s eight-month zero-export condition and the Alcoa FID are saying the same thing: the bottleneck isn’t the rock. It’s everything that happens after it leaves the mine - and the West is only now starting to build that layer.

Three things I’ll be watching next week.

• Does the indium inversion return? It held for two consecutive weeks in July before closing. Any re-emergence signals renewed domestic Chinese hoarding.

• Linde Q2 earnings July 31 - first major industrial gas print since China’s helium ban. Any upside commentary is additive to guidance already set without helium improvement.

• November 27 countdown: 133 days. Watch for any EU or US policy response to the gallium truce expiry before it happens rather than after.

The clock.

November 10 - REE extractant controls resume.
November 27 - gallium truce expiry.
January 1 - DFARS 252.225-7052.

133 days to the gallium truce. 168 days to DFARS.

Subscribe now

Live critical mineral prices at scarceearth.io.

The full archive is at williamdavid.substack.com. The Week in Minerals publishes every Saturday.

Prices directional from SMM domestic China benchmarks and ScarceEarth.io admin values as of July 1, 2026. Alcoa FID sourced from Alcoa Corporation press release, July 14, 2026. Japan zero-export condition sourced from Chinese customs data via Reuters and Japan Times, confirmed through May 2026. IEA figures sourced from IEA Global Critical Minerals Outlook 2026, July 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

Share

The series began with a question: what happens at each layer of the AI hardware stack when the physical constraint tightens, and who profits from the gap? Part 1 documented the materials constraint at the compute layer. Part 2 showed the same mechanism propagating through memory and connectivity. Part 3 demonstrated why policy is aimed at the wrong layer. Post 4 asks the question every reader has been waiting for: given all of this, who actually benefits?

The $700 billion figure - projected combined AI infrastructure spend by the five largest US hyperscalers in 2026 - is the demand-side number the series has been measured against. Before naming who benefits from the investment layer, it is worth understanding what portion of that number is structural and what portion is self-referential. The distinction matters.

The Circular Demand Problem

The AI infrastructure boom has a financing structure worth naming from primary sources rather than commentary.

CoreWeave’s S-1 disclosed $17.3 billion in debt against GPU clusters pledged as collateral. Cerebras’s S-1 disclosed a loan agreement prohibiting the company from selling chips to OpenAI’s competitors, paired with approximately 33 million warrants that OpenAI holds in Cerebras. Nvidia’s 13F discloses equity stakes in CoreWeave, xAI, Anthropic, and Mistral - the companies representing the largest sources of Nvidia’s data center demand. The demand aggregator holds equity in the demand generators.

None of this makes the demand fake. Enterprises are renting GPUs, models are being trained, and revenue is being generated. The question is not whether the demand exists. It is how much of today’s demand has been pulled forward by financing structures that assume tomorrow’s adoption curve.

The underlying infrastructure buildout is real. The financing layered on top of it makes the boom appear larger than it otherwise would. Investors who price only the headline spend without distinguishing real demand from self-referential demand are pricing the wrong number. The companies that sit at the physical constraint - the materials layer, the fabrication layer, the memory layer - earn revenue regardless of which AI startup is writing the check. The companies that depend on specific AI startups continuing to raise capital face a different risk profile entirely.

A third risk sits underneath both. One prominent version of the architectural bear case was articulated by Michael Burry in a July 2026 Substack post: that AI development went language-first when it should have gone reasoning-first, and that scaling parameters improve simulation without closing the gap to genuine reasoning. This is a different objection from the circular financing problem. Circular financing questions how demand is funded. The architectural objection questions whether the destination itself justifies the investment. If AI scaling ultimately encounters a reasoning ceiling, training compute demand could weaken in ways the financing structures above cannot paper over. Burry’s disclosed short positions in Nvidia, Micron, and AMAT are one investment expression of that thesis. The thesis may prove wrong. Burry has been wrong before on timing. But the argument belongs in the bear case.

The geopolitical rebuttal to all three objections is not that they are wrong. It is that they change the threshold at which investment slows rather than the direction of the buildout. The United States cannot assume China will slow because commercial returns disappoint. China cannot assume the United States will do the same. Once AI becomes a strategic capability rather than a purely commercial product, investment decisions are driven by relative positioning as much as expected returns. Governments can tolerate lower commercial returns than private markets because strategic competition changes the objective function from maximizing profit to preserving relative technological position.

Markets can price a slowdown. They struggle to price an arms race.

Who Sits at the Physical Constraint

The series has documented the constraint at every layer. The investment question is not who benefits from the boom - almost everyone in the stack benefits from the boom. The question is who benefits from the gap between the constraint and the resolution. Those are different companies.

The deeper one moves into the stack, the fewer substitutes exist and the longer qualification timelines become. Investors often begin with AI models because they are the most visible layer. Supply chains begin with chemistry because it is the least replaceable. The further upstream the constraint, the greater the pricing power when demand accelerates.

Fabrication

TSMC (TSM, $419.48 as of July 15, 2026): Every AI roadmap ultimately converges on one question: how many advanced packages TSMC can physically ship. TSMC is the fabrication layer. 90% of leading-edge logic production concentrated on one island 180 kilometers from the Chinese coast. Arizona accounts for approximately 2% of revenue. CoWoS packaging - the advanced packaging process that integrates HBM onto the same silicon interposer as the GPU die - is fully allocated through 2027 with TSMC as the sole commercial supplier at scale.

Update, August 1, 2026: TSMC’s CoWoS position is no longer unchallenged. Reports say TSMC is developing an EMIB-like packaging technology with Kinsus Interconnect specifically to stop customers from defecting to Intel’s competing approach - Google is named as a live prospect. Sole supplier still describes today’s volume. It no longer describes the absence of an alternative.

Update, August 5, 2026: A second thread is moving alongside the first. TSMC is expanding outsourced CoW volume to OSAT partners ASE and SPIL, beyond the WoS outsourcing already standard. This isn’t a competing technology, it’s TSMC’s own process running on someone else’s line, under TSMC’s license. The chokepoint hasn’t moved. Who physically executes inside it has started to.

Q2 2026 confirmed the argument: gross margin 67.7%, above the ceiling of its own guidance, on $40.2B in revenue. Capex raised to $60-64B. C.C. Wei: “The AI megatrend continues to drive the need for more and more computation.” The company raising capital expenditure by $8B in a single quarter is not a forecast. It is a physical commitment.

The constraint is not TSMC’s market position. The constraint is the physical limitation of what CoWoS can produce at any given capacity level, and TSMC is the only company that can produce it. If AI demand doubles, CoWoS throughput does not double without capital investment that takes years to deploy. TSMC’s pricing power comes from being the constraint, not from being the best option among several. Every GPU Nvidia ships, every HBM stack SK Hynix produces, every advanced package that connects them - all of it passes through TSMC’s fabrication capacity before reaching a data center. There is no practical way around it. That is not a competitive advantage. That is physics.

Bear case: geopolitical disruption to Taiwan operations or faster-than-expected packaging competition from Samsung or Intel Foundry compresses TSMC’s pricing power before CoWoS capacity expands to meet demand. Intel qualifying High-NA EUV in production on Panther Lake - with EMIB-T packaging reportedly matching CoWoS yields - is the first concrete data point for that scenario. The timeline remains multi-year. The direction is no longer purely hypothetical.

Memory

SK Hynix (SKHY, $165.26 as of July 15, 2026): 58% of global HBM market share. Samsung and Micron are the only alternatives, and all three have now qualified for HBM4 production for Nvidia’s Vera Rubin platform - but SK Hynix delivered HBM4 samples nearly a year before Samsung. The shortage persists beyond 2030 per the company's own guidance. Data center DRAM already accounts for 50% of global consumption, and is projected above 60% by the end of the decade. The July 10 Nasdaq listing removes the Korean exchange friction that has kept US institutional capital out of the position for 14 years. The listing does not change the operating margins, the order book, or the HBM market share. It changes who can access it. Bear case: 600%+ appreciation in the Korean-listed shares over twelve months compresses the margin for error. A Q2 earnings miss on July 29 - three weeks after listing - lands on a newly listed stock with no US trading history and no established institutional base. The listing and the earnings report are, effectively, the same bet.

Micron (MU, $904.28 as of July 15, 2026): The domestic US HBM position. Q3 FY2026 results beat estimates substantially, with $22 billion in strategic customer agreements including $18 billion in cash deposits - customers prepaying for access to future production. In a policy environment where DFARS 252.225-7052, the CHIPS Act, and the broader decoupling agenda increasingly value domestic manufacturing, Micron’s Idaho and New York fabs carry differentiation that SK Hynix and Samsung structurally cannot match. Bear case: technology gap versus SK Hynix in HBM yield and volume; the differentiation is real but the premium it commands has limits.

Rambus (RMBS, $102.89 as of July 15, 2026): The royalty that does not require any specific manufacturer to win. Active patent license agreements at SK Hynix (through mid-2034), Samsung, and Micron - the three HBM manufacturers - compound with every new HBM generation regardless of share shifts between them. The investment thesis requires only that HBM grows, which it will structurally through 2030. Bear case: Rambus disclosed in its Q1 2026 10-Q that it has been responding to a US federal grand jury subpoena connected to a Department of Justice antitrust investigation since Q1 2026. The stock has declined approximately 40% from its 52-week high as a result. The IP thesis is unchanged - active license agreements through mid-2034 are not affected by the investigation - but the legal overhang is the primary near-term risk. Secondary bear case: a future HBM standard change would require JEDEC to ratify a successor specification that memory manufacturers could implement without drawing on Rambus’s interface and timing IP, which requires industry consensus among all three current licensees to collectively move away from a framework they are already under multi-year agreements to use. Possible. Not the base case.

Equipment

ASML (ASML, $1,815.27 as of July 15, 2026): Every leading-edge logic process below roughly 7nm ultimately depends on EUV lithography. There is no second supplier. ASML sits at the equipment layer with the two-sided thesis Post 3 documented: DUV restriction is real downside risk - China was approximately 20% of total 2026 net sales per management guidance, with system sales at 14% in Q2, the specific layer the MATCH Act targets. EUV dominance is a physical fact that no other company can challenge. The constraint tightens as leading-edge nodes advance, and every leading-edge node requires EUV. Bear case: Chinese DUV substitutes advance faster than the 2028-2030 timeline most Western analysis assumes, compressing the window before ASML can replace China revenue with Western volume.

For investors seeking more torque than the durability names above: Entegris (ENTG, $138.69 as of July 15, 2026) sits inside the constraint rather than around it. Entegris makes the specialty chemicals, filtration systems, and advanced packaging materials that flow through every leading-edge fab - the consumables that fabs cannot operate without and cannot easily substitute. The qualification barrier is high: a new Entegris product requires years of in-fab validation before it ships at volume. That validation cycle, once cleared, is the moat. Every new semiconductor node creates new Entegris qualification cycles. Bear case: fab capex cycle downturn compresses consumables volumes faster than leading-edge node transitions can replace them; customer concentration at TSMC, Samsung, and Intel creates revenue volatility tied to their individual production decisions.

The Efficiency Counterargument

The series would be incomplete without addressing the strongest opposing thesis directly.

Algorithmic efficiency gains - mixture-of-experts routing, multi-head latent attention, FP8 mixed precision - are compressing per-query compute and memory requirements at the inference layer. This is real. OpenAI halved its own inference costs. Jalapeño, the custom inference chip OpenAI co-developed with Broadcom, targets 50% lower cost per inference token than current-generation Nvidia GPUs. Open-weight model share on OpenRouter rose from 34% in January 2026 to 65% by June 2026.

What the efficiency argument does not change: the fabrication layer. Building the next generation of frontier training silicon still requires the same tungsten contacts, the same gallium compounds, the same TSMC CoWoS capacity, the same HBM4 integration. Training remains compute-intensive and remains Nvidia’s stronghold. Inference is where efficiency compression bites.

The companies sitting at the fabrication, packaging, and memory layers are structurally protected from inference-layer efficiency compression because their constraint is at training and leading-edge fabrication, not at inference workloads. The efficiency argument also runs in the opposite direction from what most coverage assumes. Open-source compression isn’t hollowing out the physical layer - it’s filling it. Companies that were building proprietary foundation models to compete with Anthropic and OpenAI are dropping out of that race and becoming infrastructure customers instead. They still need GPUs, HBM, CoWoS packaging, and the optical interconnects that move data between racks. More players consolidating around the same physical infrastructure layer means the physical ceiling tightens, not loosens. China’s domestic AI ecosystem makes this concrete: CXMT’s bonded DRAM pilot separates memory cell array and peripheral circuitry onto separate wafers, then bonds them - producing high-density DRAM using only DUV lithography, bypassing EUV entirely. It doesn’t need to beat HBM. It needs to be good enough to run China’s own AI workloads at scale, freeing China’s AI buildout from Western supply chains while more Western players compete for the same constrained physical infrastructure. The chip is cheaper per token with open source. The system isn’t - because you still need more of everything to match the same output at scale. Inference efficiency compresses the application layer. The infrastructure layer stays scarce. The companies that depend on inference-layer GPU volume growth face a more complicated picture. Understanding which bucket a position sits in determines whether inference efficiency is a threat or a distraction.

The Next Constraint

The next constraint is already visible at the power layer.

The 2,100 gigawatts currently sitting in US interconnection queues exceed the total installed capacity of the US grid. Approximately 30-50% of planned 2026 AI data center capacity has already slipped to 2028. Interconnection rights - the legal permission to connect, transmit, and draw firm load at a specific node - are the scarce asset that capital alone cannot manufacture on a political timeline. The mechanism will be familiar: policy is aimed at one layer while the binding constraint has already moved to another.

The series that follows this one maps that constraint. The companies at the power delivery and grid infrastructure layer - transformer manufacturers running six-year backlogs, thermal management specialists, power electronics suppliers - are the next iteration of this argument.

The Series Conclusion

Every technological revolution eventually encounters a constraint that capital cannot solve. Investors spend enormous effort forecasting demand. They spend far less time identifying the physical bottleneck that determines supply. Throughout this series, the bottleneck appeared to move - from compute to memory to connectivity to policy - but the underlying mechanism never changed.

The deeper one moves into the stack, the fewer substitutes exist and the longer qualification timelines become. Materials sit further upstream than equipment. Equipment sits further upstream than fabrication. Fabrication sits further upstream than the AI models that run on it. The further upstream the constraint, the greater the pricing power when demand accelerates - and the longer it persists when policy tries to resolve it.

Markets price narratives. Supply chains price physics.

Investors consistently overestimate what governments can accomplish and underestimate what physics, chemistry, and manufacturing qualification determine. That is not a semiconductor conclusion. It is a framework for understanding every sector where policy mandates outrun physical supply chains: critical minerals, defense manufacturing, nuclear, grid electrification, water infrastructure, pharmaceutical manufacturing, shipbuilding.

Policy can subsidize factories. It cannot accelerate qualification.

Share

The series begins with Semiconductors & AI Hardware, Part 1: The Compute Layer:

The full archive is at williamdavid.substack.com

The Chokepoint

Structural research across seventeen sectors — energy, defense, semiconductors, space, critical minerals, and more. Every piece maps the supply chains, chokepoints, and cross-sector dependencies that headline coverage ignores.

By William David

SKHY ADR confirmed listing price $149 per ADR sourced from Nasdaq, July 10, 2026. Stock prices verified July 15, 2026 closing prices: TSM $419.48, ASML $1,815.27, NVDA $212.50, MU $904.28, RMBS $102.89, SKHY $165.26, ENTG $138.69. CoreWeave S-1 circular financing sourced from SEC filing. Cerebras S-1 loan agreement and OpenAI warrant disclosure sourced from SEC filing. Nvidia 13F stakes sourced from SEC quarterly filing. $700B hyperscaler capex figure sourced from CNAS analysis. $22 billion Micron strategic customer agreements and cash deposits sourced from Micron Q3 FY2026 earnings release and earnings call, June 24, 2026. TSMC Q2 2026 results sourced from TSMC earnings release and conference call, July 16, 2026; capex raise to $60-64B confirmed on call. TSMC CoWoS allocation status sourced from company guidance and industry reporting, 2026. HBM shortage duration and data center DRAM consumption projections sourced from SK Hynix CEO guidance and company filings, 2026. OpenRouter open-weight share data sourced from Citi Research via Reuters, June 29, 2026. OpenAI inference cost reduction sourced from The Information, June 2026, corroborated by Digg and investingLive. Interconnection queue data sourced from CNAS and Lawrence Berkeley National Laboratory analysis. Michael Burry architectural critique and parameter trap argument sourced from Michael Burry Substack, July 10, 2026 (substack.com/@michaeljburry). Burry disclosed short positions in Nvidia, Micron, Applied Materials, and SOXX ETF sourced from SEC 13F quarterly filing, Scion Asset Management. Rambus DOJ antitrust grand jury subpoena sourced from Rambus Q1 2026 Form 10-Q, filed April 29, 2026. ASML Q2 2026 China system sales 14% of net system sales sourced from ASML Q2 2026 earnings release, July 15, 2026. Intel 18A High-NA EUV production qualification and EMIB-T packaging yield sourced from KeyBanc analyst note, July 2026, and Reuters, July 15, 2026. Facts verified July 16, 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

Every multinational semiconductor company operating in China now faces an impossible choice: comply with US export controls or comply with Chinese law. Increasingly, doing both is impossible. Washington hoped export controls would slow China’s semiconductor progress. Beijing hoped industrial policy would reduce its dependence on foreign technology. Instead, both sides have accelerated the race toward materials, chemistry, and supply chain control.

This is the policy layer. It is where the most confident claims in the AI technology competition are being made - and where the gap between the claim and the physical reality is widest.

The Compliance Catch-22

State Council Order No. 834, promulgated March 31, 2026, is China’s first unified supply chain security framework. It integrates export controls, countermeasures, data security obligations, and investment screening under a single national security mandate. Every entity operating within China’s jurisdiction now carries whole-of-firm compliance obligations.

The catch-22 it creates is structural, not procedural. Article 13 of China’s Anti-Foreign Sanctions Law imposes penalties on any entity that complies with foreign sanctions designations targeting Chinese companies. The US Entity List is a sanctions-adjacent designation tool. TSMC, Samsung, and SK Hynix are all managing annual export licenses for their China fabs after their validated end-user exemptions expired at the end of 2025.

Comply with US export restrictions and face potential penalties under Chinese law. Comply with Chinese law and face potential penalties under US law. The companies most exposed are the ones with the most sophisticated China operations - which are, by definition, the ones that matter most to the global semiconductor supply chain. That is not a negotiating position. That is a structural constraint on how AI infrastructure gets built.

The MATCH Act - and What It Misses

The Multilateral Alignment of Technology Controls on Hardware Act passed the House Foreign Affairs Committee on April 22, 2026 - the largest export control markup in congressional history by number of bills advanced in a single session. The Senate companion has bipartisan backing from Senators Risch, Ricketts, Kim, and Schumer. The bill has not yet passed the full House or Senate floor.

The core mechanism designates CXMT, YMTC, SMIC, Hua Hong, and Huawei as covered facilities subject to Entity List-equivalent restrictions on equipment sales, servicing, and technical support. A 150-day diplomatic window gives allies structured time to align their own controls. If they do not, the Foreign Direct Product Rule extends US jurisdiction to equipment made abroad using US software or components.

The law was written for the constraint as it existed in 2022. The center of gravity has shifted below the equipment layer into chemistry, materials, and process inputs. Export controls slowed access to advanced tools. They did not stop China from moving investment into the layers beneath those tools.

Domestic Chinese suppliers accounted for roughly 35% of equipment purchased by Chinese fabs in 2025, ahead of Beijing’s own targets - up from approximately 25% in 2024. The substitution is not happening at the equipment layer the MATCH Act targets - it is happening at the deposition chemistry, process gas, and specialty materials layer, one level below where the law is aimed. China’s imports of semiconductor manufacturing machinery reached a record $51.1 billion in 2025 even as six previous rounds of US export controls attempted to constrain that flow.

The 150-day diplomatic window is where the allied alignment problem becomes concrete. ASML’s China revenue represented 29.1% of its total sales in 2025. Tokyo Electron’s exceeded 40%. Neither government can absorb those losses on a congressional timeline. The cryogenic etch ban that appeared in earlier drafts of the MATCH Act was rolled back before the committee vote. What survived is the part of the bill where allied resistance is highest.

The MATCH Act targets the tools China buys. China is increasingly supplying those tools domestically. That is not a reason to abandon the law - it is a reason to understand what it will and will not accomplish.

The Gallium Truce - and What It Actually Is

In November 2025, following the Busan summit between Presidents Trump and Xi, China suspended its ban on gallium, germanium, and antimony exports to the United States until November 27, 2026. The announcement was widely described as a trade de-escalation gesture.

In 2025, China’s exports of unwrought gallium fell 94% from 2024 levels. They have not yet recovered in 2026. The metals remain on China’s dual-use export control list, meaning exporters must still apply for licenses for every overseas sale. The military end-use ban - prohibiting exports to US military users or for military uses - was not addressed by the suspension announcement. It remains in effect.

The market priced the truce correctly. Spot gallium prices reported by Fastmarkets reached a record high of $1,850 per kilogram in April 2026 - an increase of over 200% since the start of 2025, six months after the truce was announced. A genuine resumption of Chinese gallium exports would put downward pressure on Western spot prices. Prices rose through the truce. The market understood what the announcement actually said.

November 27 is not a resumption date. It is a decision date. On that date, China can re-impose the full prohibition with no additional announcement, no diplomatic process, and no transition period - because the underlying control architecture never left. The truce suspended the ban. It did not resume the supply.

Together, these three policies reveal the same pattern: governments are competing over equipment while supply chains are reorganizing around materials.

The Investment Layer Underneath the Policy Layer

The policy collision creates a specific investment situation. There are four categories of policy winners at the equipment layer, organized by where they sit in the fabrication process.

Lithography

ASML (ASML, $1797.32 as of July 10): The most exposed and the most protected simultaneously. The MATCH Act’s DUV provision targets 29.1% of ASML’s revenue - the China business Western governments want ASML to reduce. The same Western governments have made ASML’s EUV monopoly a strategic asset they are committed to protecting. No other company on earth can manufacture EUV machines. That is not a competitive position. That is a physical fact. The two-sided thesis: DUV restriction is real downside risk, EUV dominance is structurally protected upside. Bear case: Chinese DUV substitutes advance faster than the 2028-2030 timeline most Western analysis assumes, compressing the window before ASML can replace China revenue with Western volume.

Deposition and Etch

These three names sit at the layer where China’s domestic substitution is furthest advanced - and where the MATCH Act’s pressure will be felt most directly if it clears the full House and Senate.

Tokyo Electron (8035.T, ¥72,940 as of July 10): The allied equipment name most exposed by the MATCH Act’s 150-day diplomatic window. Over 40% of revenue from China. TEL’s position in deposition and etching makes it directly relevant to the chemistry substitution argument: the equipment China is building domestic alternatives for is the equipment TEL currently dominates. Bear case: Japanese diplomatic alignment with MATCH Act timelines forces a revenue cliff before TEL can replace that volume elsewhere.

Applied Materials (AMAT, $602.50 as of July 10): The broadest equipment exposure across deposition, etch, and CMP. Bear case: MATCH Act passage compresses China revenue simultaneously with domestic substitution reducing future addressable market.

Lam Research (LRCX, $350.33 as of July 10): Etch dominance is the specific capability China is attempting to substitute domestically fastest. Bear case: same as AMAT, with additional exposure if Chinese etch substitution succeeds at mature nodes before Lam can replace that customer base.

Metrology

KLA Corporation (KLAC, $231.52 as of July 10): Process control is the hardest equipment category to substitute - you cannot make chips without measuring whether the process worked, and KLA’s metrology tools represent years of accumulated calibration data that cannot be replicated on a short timeline. For a Chinese fab trying to build process control in-house, the data itself is the moat, not just the hardware. Bear case: MATCH Act passage compresses China revenue, but KLA’s replacement window is the longest of any equipment maker because the calibration data advantage degrades slowly.

Advanced Packaging

ASM International (ASMIY, $1033.80 as of July 10): Atomic layer deposition at the gate-all-around transistor transition that defines leading-edge nodes from 3nm forward. ALD cannot be substituted with legacy processes for the chips that matter most to the AI buildout. Bear case: gate-all-around transition slower than expected, reducing the urgency of ALD adoption at scale.

BE Semiconductor Industries (BESI.AS, $255.30 as of July 10): Hybrid bonding capability that makes HBM4 integration possible. As the series has documented across Parts 1 and 2, advanced packaging is the chokepoint that breaks before the fab. BESI’s hybrid bonding equipment is fully allocated through 2027. Bear case: alternative bonding approaches reduce BESI’s addressable market in future generations.

The winners are not necessarily the companies with the best technology. They are the companies that reduce dependence on China faster than China reduces dependence on them.

Governments can subsidize factories. They can restrict equipment. They can negotiate trade agreements. None of those policies change the chemistry inside the fabrication process or shorten qualification timelines measured in years.

Part 4 leaves policy behind and follows the money. If governments cannot remove the physical constraints, investors still have to decide who benefits from building around them.

Policy can subsidize factories. It cannot accelerate qualification.

Share

The full archive is at williamdavid.substack.com. The compute layer argument begins in Semiconductor Series Post 1.

MATCH Act House Foreign Affairs Committee vote sourced from House Foreign Affairs Committee press release, April 22, 2026 (foreignaffairs.house.gov). MATCH Act bill text (H.R. 8170) sourced from Congress.gov, primary record. Senate companion bill (S.4281) sourced from Congress.gov; Senate sponsors Risch, Ricketts, Kim, and Schumer confirmed in Senate companion bill text. MATCH Act DUV provision and cryogenic etch rollback sourced from Tom’s Hardware, April 22, 2026, citing April 16 revised draft circulated by bill sponsors. R Street Institute MATCH Act analysis (ASML and TEL China revenue figures, equipment import data) sourced from R Street Institute, April 30, 2026. China gallium export data (94% decline in 2025, price record at $1,850/kg in April 2026) sourced from CSIS analysis, May 22, 2026. Gallium truce mechanics and military end-use ban status sourced from Fastmarkets, November 9, 2025; Ministry of Commerce announcement November 9, 2025. State Council Order No. 834 sourced from Andersen Institute analysis, April 29, 2026. All stock prices verified at market close July 10, 2026 via Yahoo Finance. ASML $1,797.32. Applied Materials $602.50. Lam Research $350.33. KLA Corporation $231.52. Tokyo Electron ¥72,940. ASM International $1,033.80 (ASMIY). BE Semiconductor $255.30 (BESI.AS). Financial data sourced from company earnings reports and public market data as of July 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

The AI shortage everyone is measuring is a GPU shortage. The shortage that is actually binding is memory. The constraint after that is bandwidth. Both trace upstream to a material China licenses one shipment at a time.

On July 1, 2026 - the same day the WF6 halt took effect at the compute layer - the indium price signal that ended the previous post became harder to ignore. Chinese domestic indium crossed above Western industrial warehouse price for what appears to be the first time in the pricing data available to verify: $775 per kilogram inside China against $717 per kilogram in Western warehouses. Under a functioning export-control regime, this is the opposite of what you would normally expect. July 1 broke that pattern.

Indium is the upstream input for indium phosphide - the substrate on which optical transceivers are built. Optical transceivers are the components that move data between GPU clusters inside AI data centers. Without them, a $40,000 GPU delivers no additional AI performance regardless of how much compute it contains. The previous post ended with WF6. It turns out WF6 never left the story.

The Memory Constraint

AI clusters require memory that is physically integrated onto the same silicon interposer as the GPU die. This is high-bandwidth memory - HBM - and switching suppliers requires redesigning the chip package from scratch. It is not a procurement decision. It is a fabrication decision.

SK Hynix holds 58% of global HBM market share. Samsung and Micron are the only alternatives. SK Hynix's own guidance is that the shortage persists well into the next decade. DRAM prices rose more than 600% in 2025. Data center DRAM now represents 50% of global consumption, projected above 60% by 2030. Up to 70% of all memory chips produced globally in 2026 are consumed by AI data centers.

The Vera Rubin platform - Nvidia’s next-generation GPU entering production this half - requires HBM4, the latest memory generation. The platform designed to relieve one constraint deepens another before any relief arrives. HBM3E and HBM4 production at SK Hynix and Micron is fully booked through 2027, with demand extending into 2028.

WF6 never left the story. The same tungsten hexafluoride that Kanto Denka Kogyo and Central Glass halted production of on July 1 is required for tungsten contacts inside every DRAM die. The compute shortage and the memory shortage trace to the same upstream material. One halt. Two layers of the AI hardware stack affected simultaneously.

Here is what that means practically: the AI buildout needs more memory than exists. The companies that make the memory need materials that just got harder to source. The GPU designed to solve the compute shortage requires a new generation of memory that makes the memory shortage worse before it gets better. The semiconductor sector has shed trillions in market value since late June as investors rotate out of AI hardware on valuation concerns. The physical constraint doesn’t rotate with them.

SK Hynix (000660.KS, ₩2,201,000 as of July 8) begins trading on Nasdaq as SKHY on July 10 - two days from now. Priced at $149 per ADR, the $28-29 billion raise is the largest ADR listing in history. Korean-listed shares have appreciated nearly 800% over the past twelve months. SK Hynix has already begun mass production of 192GB SOCAMM2 memory modules built specifically for the Vera Rubin platform - HBM4 demand not as a future thesis but as current commercial production. HSBC assigned a roughly 20% valuation premium to the ADR versus the Korean listing, arguing that US investor access historically narrows Korea’s valuation discount. Q2 2026 earnings land July 29 - three weeks after the listing. Investors who buy SKHY on listing day are simultaneously betting on Q2 results. Bear case: ~800% appreciation in twelve months compresses the margin for error. A Q2 miss lands on a newly listed stock with no US trading history and no established institutional base.

Micron (MU, $948.80 as of July 8) is the only US-headquartered company producing HBM at commercial scale. Q3 FY2026 results beat estimates substantially - $25.11 EPS against a $20.49 estimate, $41.46 billion revenue against $35.69 billion expected. HBM3E and HBM4 fully booked through 2027, with $22 billion in strategic customer agreements including $18 billion in cash deposits. In a policy environment where DFARS 252.225-7052, the CHIPS Act, and the broader decoupling agenda increasingly value domestic manufacturing, Micron’s Idaho and New York fabs carry differentiation that SK Hynix and Samsung structurally cannot match. Bear case: Samsung and SK Hynix outpace on yield and volume; Micron’s technology gap narrows but doesn’t close.

Rambus (RMBS, $109.64 as of July 8) holds the dominant IP position in HBM controller design, with active patent license agreements at SK Hynix, Samsung, and Micron - the three HBM manufacturers - and over 100 HBM design wins across generations. The royalty compounds with every new HBM generation regardless of which manufacturer wins volume. The investment thesis does not require any specific memory producer to win. It requires only that HBM grows - which it will structurally through 2030. Bear case: a future HBM standard change routing around core Rambus patents. For that risk to materialize, JEDEC would need to ratify an HBM5 or successor standard that specifies a fundamentally different controller architecture - one that memory manufacturers could implement without drawing on Rambus’s interface and timing IP. That requires industry consensus among SK Hynix, Samsung, and Micron to collectively move away from a specification framework all three currently license and have multi-year agreements to continue using. Possible. Not the base case.

The Connectivity Constraint

The memory layer determines how much compute can actually be utilized. The connectivity layer determines whether that compute can communicate.

At 800 gigabits per second - the current leading-edge deployment standard - fiber optic transceivers are the only commercially viable option for moving data between servers at the speeds AI workloads require. The 1.6 terabit per second generation is beginning deployment at major hyperscalers. Fiber optic transceiver lead times now exceed forty weeks at major suppliers, underscoring that bandwidth has become a manufacturing constraint rather than merely a networking one.

The substrate that makes these transceivers work is indium phosphide - InP, a compound semiconductor grown from refined indium metal. China controls 55-60% of global refined indium supply through an active export licensing regime. Every shipment requires Ministry of Commerce approval. A licensing regime also functions as an intelligence-gathering regime: China knows exactly who needs indium, at what volume, for what application, before approving each shipment.

The price inversion that opened this post suggests the licensing regime is beginning to bind domestically as well as internationally. The material that connects the compute layer to the bandwidth layer is under the same licensing architecture as gallium and germanium.

AXT Inc. (AXTI, $60.12 as of July 8), through its Beijing-based subsidiary Tongmei, is the primary Western-aligned supplier of InP substrates. A backlog exceeding $100 million is waiting on Ministry of Commerce export approval. On June 25, 2026, AXT-Tongmei signed a three-year Master Development and Supply Agreement with Coherent Corp - a primary-source-confirmed direct supply relationship between the InP substrate manufacturer and the optical transceiver company that most depends on it. Bear case: the MoC licensing gate tightens before the backlog clears; AXT’s manufacturing remains predominantly in China, limiting its ability to route around the constraint it depends on. Thesis breaker: VCSEL-based co-packaged optics becoming commercially viable at scale, reducing InP substrate demand at the data center layer.

Coherent Corp. (COHR, $317.05 as of July 8) sits directly at the AI data center bandwidth buildout - and directly inside the InP licensing constraint. The three-year AXT-Tongmei supply agreement formalizes that dependency into a contracted relationship. The CEO flew to Beijing with the Trump trade delegation specifically to raise export license delays on indium phosphide - primary-source evidence of a supply chain dependency operating at the executive level. Raymond James raised its price target to $435, Strong Buy. The 1.6T transceiver qualification cycle at major hyperscalers is the multi-year demand driver. Bear case: hyperscaler in-house photonics development displacing merchant silicon suppliers. Thesis breaker: hyperscaler adoption of VCSEL co-packaged optics architecture displacing InP-based transceiver demand at the rack scale.

Astera Labs (ALAB, $393.16 as of July 8) makes PCIe and CXL retimers - the components that manage signal integrity at the interfaces connecting memory, compute, and storage inside AI server racks. Retimers are not optional. Electrical signal degradation across rack-scale distances makes them a physical requirement. Q1 revenue was $308.4 million, up 93% year-over-year. Q2 guided up to $365 million. The Scorpio X-Series switch platform extends the addressable market from retimers into full rack-scale networking. Next earnings August 11. The agentic AI shift - where the ratio of CPUs to GPUs moves toward 1:1 as AI agents run increasingly complex multi-step tasks - is the structural catalyst that extends ALAB demand beyond the current training cycle. Primary bear case: hyperscaler in-sourcing of retimer functionality. Secondary: Amazon warrant structure compressing gross margins if dilution accelerates.

Corning (GLW) and Amphenol (APH) sit underneath the optical layer at the physical infrastructure level - fiber cable and interconnects respectively. Corning signed the largest fiber deal in its history with Meta. Amphenol is reporting record bookings and joined the optical standard that will define how AI data center racks are wired. Both illustrate the scale of the physical buildout pushing against the same upstream materials constraints.

The Signal Underneath

The compute layer established the thesis: physical constraints are binding at the materials and fabrication layer while markets price narratives about those constraints rather than the constraints themselves.

The memory and connectivity layer proves the thesis is not limited to one material or one chokepoint. WF6 at the compute layer. Indium at the connectivity layer. HBM at the memory layer. The same upstream dependencies - China’s licensing authority over the materials that flow into every layer - appear at each level of the stack, each with its own qualification clock and its own gap between announced capacity and operational reality.

Policy can subsidize factories. It cannot accelerate qualification.

The next part asks what governments can do when the binding constraint isn’t capital - it’s chemistry.

Share

The full archive is at williamdavid.substack.com. The compute layer argument begins in Semiconductor Series Post 1:

SK Hynix (000660.KS) price as of July 8, 2026 (Korean Stock Exchange close). SK Hynix ADR priced at $149 per ADR, July 9, 2026. HBM shortage duration: SK Hynix CEO guidance, company filings 2026. DRAM price appreciation and consumption data: industry reports 2025-2026. SOCAMM2 HBM4 production for Vera Rubin: SK Hynix press release March 15, 2026. Micron (MU) Q3 FY2026 earnings: company earnings release June 24, 2026. Micron strategic customer agreements and cash deposits: company earnings call June 24, 2026. Rambus HBM IP licensing: active patent license agreements with SK Hynix (through mid-2034), Micron (December 2024), and Samsung; Rambus investor presentation Q2 2026; Rambus HBM4E Controller IP press release March 4, 2026. Astera Labs (ALAB) Q1 2026 results: company earnings release May 5, 2026. ALAB Scorpio X-Series: company press release April 2, 2026. ALAB personal position disclosure: publisher. COHR CHIPS Act letter of intent: Coherent press release June 16, 2026. AXT-Tongmei/Coherent supply agreement: AXT press release June 25, 2026; Coherent confirmation June 25, 2026. AXTI backlog: company earnings call and filings Q1 2026. Transceiver lead times: industry procurement data 2026. Indium prices: SMM domestic China July 1, 2026 ($775.34/kg); strategicmetalsinvest.com Western warehouse July 2, 2026 ($717.50/kg). China indium supply concentration: USGS Mineral Commodity Summaries 2026. All stock prices verified July 8, 2026. Financial data sourced from company earnings reports and public market data as of July 8, 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

The AI industry is spending $700 billion this year on infrastructure it cannot build without China. Not because it chose the dependency. Because the physical reality of what an AI chip requires runs through supply chains China controls at every layer simultaneously - and on July 1, one of those layers went into an unconfirmed halt.

On that date, Kanto Denka Kogyo and Central Glass - the two Japanese companies that together supplied roughly 25% of global tungsten hexafluoride production - halted output. WF6 is the specialized gas deposited inside every advanced semiconductor to form the conductive contacts that make the chip function. Without it, there is no 3D NAND flash. No high-bandwidth memory. No AI accelerator at the nodes that matter. Samsung, SK Hynix, and TSMC were notified in April. They had three months to find immediately deployable commercial alternatives capable of replacing the lost Japanese capacity.

There were none. The reason is worth understanding precisely.

WF6 is made from high-purity tungsten powder - a material that is roughly 80% Chinese. In early 2026, China tightened export controls on tungsten-related items to Japan under its dual-use licensing system. Japan’s imports of the tungsten precursor used to produce WF6 collapsed by as much as 70% on key trade routes. The Japanese plants didn’t run short of demand. They ran short of raw material. On July 1, the last inventory was consumed and production stopped - with no confirmed restart date.

That is the supply story. There is a second story running on the same timeline - and it is the one most coverage missed entirely.

The Information Operation

As the Japanese supply disruption became visible in April, Chinese state-adjacent financial media began circulating a claim: CSSC Peric (688146.SH) - China’s largest WF6 producer, with 2,000 tonnes of annual capacity and another 1,000 tonnes under construction - could fill the gap. Its product already meets the purity requirements for advanced memory and AI chip fabrication. CSSC Peric was the solution.

The claim was technically defensible on capacity. It was functionally false on timeline - and the distinction between capacity and qualification is the argument most coverage missed.

Capacity and qualification are not the same constraint.

Leading-edge fabs typically maintain multiple qualified suppliers where process requirements allow - but for specialty materials like WF6, diversification is constrained by both chemistry and qualification. Introducing a new WF6 source into a semiconductor fabrication process requires 12 to 18 months of testing - purity specifications measured in parts per billion, yield validation across production runs, supply reliability verified at volume. Qualification cannot manufacture molecules that do not exist. CSSC Peric was not a qualified supplier at Samsung or TSMC when the claim circulated. No qualified replacement could exist within the remaining window.

The Korean bridge supply that chipmakers pivoted toward carries its own version of this problem. SK Specialty signed a long-term agreement to supply Samsung with 150 tonnes of WF6 per month - a real response. But Foosung, the second Korean producer stepping into the gap, has initiated qualification with CSSC Peric to begin large-scale Chinese WF6 imports in August. The bridge that markets are pricing as a Western solution runs through the same Chinese source the narrative named as the replacement.

What This Means: China cut off the tungsten powder that Japanese producers needed. As those producers halted output, Chinese media promoted a Chinese company as the solution.

Whether intentional or not, the effect was the same: the narrative suggested the shortage was temporary, reducing the market’s perceived urgency of qualifying non-Chinese alternatives while positioning CSSC Peric as the obvious replacement before the qualification race had properly begun.

This is what supply chain leverage looks like. Physical disruption and market narrative operate together, whether by design or by incentive - and a market response that lags the physical reality is the result either way.

The Efficiency Argument, Stated Correctly

Before the fabrication constraint is named, it needs to clear one challenge that has been running in the coverage for weeks.

A Wall Street Journal headline in late June announced that a Chinese open-weight model had matched Anthropic’s frontier capability in cybersecurity - and by extension, that algorithmic efficiency was closing the gap that hardware controls were supposed to maintain. The headline traveled fast. The underlying reporting did not support it. The WSJ’s own sourcing said something narrower: under guided, narrow conditions, a publicly available Anthropic model and the Chinese model could match the frontier specifically on bug-finding tasks. That is not general capability parity, and Anthropic’s own published red-team data shows the actual difference is categorical, not narrow.

But there is a legitimate argument underneath the overstated headline, and it deserves a direct answer.

Inference-layer efficiency compression is real. Open-weight models now do work that cost $4 per million tokens six months ago at a fraction of that price, and open-source models grew from 34% to 65% of usage on the OpenRouter platform through the first half of 2026. The cost differential is one driver - the second is structural: enterprises running sensitive workflows on open-weight models retain full control over their data stack in a way that frontier API access cannot guarantee, regardless of contractual terms. On June 24, OpenAI and Broadcom unveiled Jalapeño - a chip built from scratch specifically for running AI models in response to user requests, as opposed to training them. Broadcom’s CEO stated that early testing shows roughly 50% lower cost per inference token compared to current AI GPUs. That figure is the company’s own, not yet independently validated. But the architectural argument is sound: a purpose-built inference chip strips out compute capacity that those workloads never use.

What This Means: Inference is getting cheaper. Efficient software is compressing hardware costs at a specific layer of the AI stack. This would matter the same way regardless of which country’s lab or company achieved it first. It is not a China story. It is an economics story.

What efficiency gains do not change: the fabrication layer. Building the next generation of frontier training silicon - the Vera Rubin platform entering production this half, the custom ASICs hyperscalers are ordering, the HBM4 memory stacked on the same interposer as the GPU - still requires WF6 for tungsten contacts. Still requires gallium compounds in power delivery systems. Still requires the TSMC capacity every name at this layer depends on. Algorithmic efficiency compresses demand at the inference layer. It does not move the physical constraint at the fabrication layer. Both things are true simultaneously, and a framework that acknowledges only one of them is incomplete.

The question this creates for investors is not whether Nvidia loses. The question is which part of Nvidia’s economics becomes commoditized first. The market often prices Nvidia as though inference and training are one business. Economically, they are beginning to diverge. On July 1 - the same day the WF6 halt took effect - Meta announced it was building a cloud business to sell excess AI computing capacity, an admission that inference-layer infrastructure has outrun its internal demand. Inference-layer pricing power compresses first - purpose-built chips and open-weight models are attacking it from two directions simultaneously. Training-layer demand is where the materials constraint is most acute and where Nvidia’s position is most durable. Those are different economic phenomena operating at different layers of the same stack.

The Foundational Constraint

There is one company underneath nearly every leading-edge AI chip in this stack.

Taiwan Semiconductor Manufacturing Company fabricates over 90% of the world’s leading-edge logic chips. Not 50%. Not 60%. 90%. Every AI GPU, every custom ASIC, every chip the $700 billion buildout requires runs through one company’s fabs - most of them on a single island 180 kilometers from the Chinese coast.

Arizona is producing. The $165 billion US commitment to TSMC’s American expansion is real, and Phase 2 equipment installation begins this quarter. Arizona currently represents approximately 2% of TSMC’s total revenue. Washington has spent hundreds of billions attempting to reduce dependence on Taiwan. The result so far is production capacity equal to roughly two cents of every dollar TSMC earns.

There is a packaging bottleneck that binds before fabrication itself - and most AI hardware coverage hasn’t found it yet. Advanced packaging is no longer a post-fabrication step. It has become part of the semiconductor itself. CoWoS - the process that physically integrates HBM memory with a GPU on the same interposer, enabling the data bandwidth that AI workloads require - has TSMC as its sole commercial supplier at scale. CoWoS capacity is fully allocated through 2027.

Update, August 5, 2026: TSMC's CoWoS execution is no longer entirely in-house. TSMC has expanded outsourced CoW volume, the chip-to-interposer step previously kept mostly in-house, to OSAT partners ASE and SPIL, building on the WoS outsourcing already in place. The technology and the license remain TSMC's. The hands doing the work no longer are.

A leading-edge AI GPU without advanced packaging is not a commercially deployable AI accelerator. A cross-strait disruption scenario doesn’t hit the fabrication line first. It hits the packaging line. The constraint that binds before the most visible constraint is the one that breaks first.

One detail worth noting: the chip designed to reduce Nvidia’s inference dominance is manufactured by TSMC. The solution to one layer of constraint runs through the foundational constraint at every other layer.

What the Constraint Means

The physical constraint is binding now. Markets routinely price capacity expansions long before they exist physically. The question investors should ask is whether current valuations assume qualification timelines that the underlying chemistry and manufacturing process cannot actually support.

On the WF6 constraint: the qualification window for non-Chinese alternatives began in April under emergency conditions. If a leading fab announces a qualified non-Chinese WF6 source before year-end, it means the clock started earlier than the shortage implied and the supply gap closes faster than the current timeline suggests. If no qualification announcement arrives by Q4, the shortage deepens into 2027 precisely as the price trajectory has been pricing. That is the event to watch - not another narrative about capacity.

Nvidia (NVDA, $194.83 as of July 2) generated $215.9 billion in data center revenue in FY2026, 88% of total revenue. The inference-layer pricing power compresses first, for the reasons above. The training-layer materials constraint - WF6, gallium, the TSMC capacity that Vera Rubin deepens rather than relieves - is where Nvidia’s position is most durable and where the physical constraint is most acute. If materials tighten before new qualified capacity comes online, scarcity pricing power concentrates in the qualified capacity that already exists. Those are different conditions from inference compression, and they require different frameworks to hold simultaneously.

Taiwan Semiconductor (TSM, $434.16 as of July 2) reports Q2 earnings on July 16. Analysts have raised targets to $575-590 on AI demand for leading-edge capacity. The materials constraint doesn’t change what TSMC earns from the demand that exists. It defines the ceiling on how quickly the demand it cannot yet serve can be served.

One signal the compute layer coverage has not yet named: Chinese domestic indium ($775/kg as of July 1) crossed above Western warehouse price ($717/kg as of July 2) for the first time on record. Indium phosphide - refined from indium - is the substrate for the optical transceivers connecting GPU clusters inside AI data centers. Acute domestic tightening inside China on a material that connects the compute layer to the memory layer is how the next constraint appears before the coverage does.

That is where this series goes next.

https://williamdavid.substack.com/p/semiconductors-and-ai-hardware-part-09e

Share

WF6 supply warning: Kanto Denka Kogyo and Central Glass client notifications, The Elec (April 3, 2026), TrendForce (June 12, 2026). WF6 5N-grade price: up 232.7% year-on-year per ChemNet via TrendForce. CSSC (Handan) Peric Special Gases Co., Ltd. (688146.SH): company investor disclosure and ChemNet (May 15, 2026). Foosung/CSSC Peric qualification: SunSirs and Futunn via KuCoin News, June 2026. SK Specialty Samsung contract: SunSirs via KuCoin News, June 2026. Jalapeño: OpenAI and Broadcom primary press releases, June 24, 2026. 50% cost figure: Broadcom CEO Hock Tan, Bloomberg/Reuters, June 24, 2026 - self-reported, no independent validation at launch. WSJ headline: McMillan, Huang, Ramkumar, June 27, 2026. Open-weight adoption data: Citi Research via Reuters, June 29, 2026. Nvidia FY2026 data center revenue and customer concentration: Nvidia earnings report. TSMC Arizona revenue share: company disclosures. Indium prices: SMM domestic China July 1, 2026 ($775.34/kg); strategicmetalsinvest.com Western warehouse July 2, 2026 ($717.50/kg). NVDA ($194.83) and TSM ($434.16) prices verified July 2, 2026. Financial data sourced from company earnings reports and public market data as of July 5, 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

The full archive is at https://williamdavid.substack.com. The processing layer argument begins in the Critical Minerals series: https://williamdavid.substack.com/p/the-chokepoint-guide-to-critical.

Update July 6, 2026: Reporting from Capital Mischief raises credible questions about whether the Japanese production halt is permanent or a supply warning rather than a confirmed cessation. The conversion chain argument, the 12-18 month qualification window, and GTP's scarcity pricing power position are unchanged regardless of the precise status of the Japanese lines - and are now confirmed by independent analysis.

Today, Kanto Denka Kogyo and Central Glass - the two Japanese companies that together supplied approximately 25% of global tungsten hexafluoride (WF6 - the specialized gas that deposits conductive tungsten contacts inside every advanced semiconductor, from DRAM to HBM to AI accelerators) - permanently ceased production. June 30 was the last shipment date. Today is when the guarantee expires.

Samsung, SK Hynix, and TSMC were notified in April. They have had three months to find alternatives at commercial scale. There are no immediately scalable alternatives capable of replacing the lost Japanese capacity. Qualifying a new WF6 supplier at any semiconductor fab takes 12 to 18 months. Building new WF6 synthesis and purification capacity from scratch takes longer. The price already reflects it: 5N-grade WF6 - the semiconductor-grade material Samsung and SK Hynix actually use - is up 232.7% year-on-year. Korean suppliers SK Specialty and Foosung notified customers in spring that contract prices for 2026 would rise 70 to 90%.

On March 16, Almonty Industries commissioned Phase 1 of the Sangdong tungsten mine in South Korea - historically one of the world’s largest tungsten deposits, returning to production after more than 30 years. Phase 1 is designed to produce roughly 2,300 tonnes of tungsten concentrate per year. Phase 2, targeting 2027, doubles that. At full capacity Sangdong is expected to supply approximately 40% of global tungsten demand outside China.

That is the most credible non-Chinese tungsten supply response in decades. It is real. It is producing. The market celebrated it.

The shortage appeared one layer downstream.

The Conversion Chain Most Coverage Misses

Sangdong produces tungsten concentrate - WO3 (tungsten trioxide, the raw output of mining and initial ore processing). What a semiconductor fab needs to deposit tungsten contacts inside a chip is WF6. Between those two things sits a conversion chain: concentrate to APT (ammonium paratungstate - the primary tungsten intermediate compound), APT to high-purity tungsten oxide, oxide to tungsten metal powder, powder to WF6 via fluorination under controlled conditions. Each step requires specialized facilities, specialized chemistry, and its own process qualification.

Kanto Denka and Central Glass were not just running the final step. They were running the end of a chain whose upstream half - high-purity tungsten powder - was supplied almost entirely from China. When China restricted tungsten exports to a whitelist of 15 designated firms in 2026, Japan’s imports of Chinese APT collapsed by as much as 70% on key trade routes. The plant couldn’t make the gas because the powder stopped arriving. The powder stopped arriving because China stopped approving the shipments.

Today the plant closes.

European APT rose from approximately $900 to $940 per metric tonne unit (MTU) in January 2026 to $2,900 to $3,180 per MTU by early June - up more than 230% year-to-date - against a long-term historical average below $300 per MTU.

What This Means: China set the price. Japan built the plant. China cut the powder. The plant stopped working. Today the plant closed. The mine that just opened produces concentrate. The gap between concentrate and the gas a semiconductor fab can actually use is the problem the mine hasn’t solved yet.

The One Western Processor Already Inside the Bottleneck

The obvious question is who already occupies that missing conversion layer.

Global Tungsten & Powders (GTP), a division of the CERATIZIT Group’s Plansee Group, operates the primary Western tungsten powder processing facility in Towanda, Pennsylvania. GTP processes approximately 10% of the world’s tungsten production - the largest such operation outside China. As of February 2026, Plansee’s CEO stated plainly: “We are working at capacity. We could produce 50 to 60 percent more and sell it on the market. We are sold out for the next six to nine months. Some of our customers are desperate.”

GTP is also the exclusive Western customer for Sangdong’s concentrate. The ore ships from South Korea to San Francisco and then overland to Towanda. Plansee is a partial owner of Almonty Industries. The Sangdong mine and the Towanda powder facility are not two separate supply chain developments. They are two halves of the same integrated response.

In May 2026, Plansee announced a joint venture with Manhattan Five Partners to expand GTP’s tungsten oxide production capacity at Towanda to approximately 12,000 tonnes per year, with initial deliveries intended to serve a strategic US government and defense reserve. That expansion is beginning now.

GTP’s capacity constraint is precisely the qualification moat that makes it investable: companies already qualified at the conversion layer hold something that cannot be quickly reproduced. GTP is qualified. GTP is expanding. GTP is already sold out into 2027.

What This Means: The bottleneck in Western tungsten supply is not the mine. It has never been the mine. It is the powder processing step between ore concentrate and semiconductor-grade input - and the one Western facility that does this at meaningful scale confirmed in February that it cannot keep up with demand.

A Nuance on the Korean Replacement Supply

South Korean producers - SK Specialty and Foosung - are stepping in on the WF6 side. SK Specialty has signed a long-term contract to supply Samsung with 150 tonnes of WF6 per month. Foosung has begun qualification with Chinese suppliers for expanded tungsten powder imports.

That response is real. It also runs on a condition worth naming precisely. Korean WF6 producers themselves depend on imported tungsten powder. Most of that powder currently still comes from China. South Korea continues to receive Chinese powder access that Japan no longer does. That is an observable difference in market access, not a declared Chinese policy. Whether Beijing maintains that asymmetry as Korean WF6 production scales to fill the Japanese gap is the open question. If it does, the Korean response provides meaningful bridge supply. If it doesn’t, the crisis deepens before the Western conversion chain is ready.

What This Means for Capital

The market celebrated the mine because mines are visible. The shortage emerged one layer downstream because bottlenecks rarely sit where headlines do. That pattern is this publication’s thesis applied to tungsten.

The investment argument here has two distinct layers worth holding separately.

The first is option value: strategic assets become more valuable before their cash flows reflect it. When Japan’s WF6 capacity permanently exits and Chinese powder controls prevent quick Japanese reconstruction, every existing non-Chinese processing asset increases in strategic value regardless of near-term earnings. Yesterday these assets represented optionality. Today they represent capacity that already has nowhere else to go.

The second is scarcity pricing power, and it is different from option value. GTP is already sold out. SK Specialty and Foosung have already locked in 70 to 90 percent contract price increases with Samsung and SK Hynix. The existing qualified capacity is not merely becoming more strategically valuable - it is commanding premium pricing right now, because the alternative supply just permanently exited. Pricing power in an oligopolistic market where qualification takes 12 to 18 months and the primary supplier just closed is a structural margin expansion for the qualified survivors. That is a different economic phenomenon than option value, and sophisticated readers will notice if the two are conflated.

The qualification moat applies directly. Demand for qualified non-Chinese tungsten processing is accelerating faster than qualification itself can. New WF6 synthesis facilities face the same 12 to 18-month customer qualification cycle regardless of when they open. The companies already inside that qualification - GTP at the powder layer, SK Specialty and Foosung at the WF6 layer - hold a position that cannot be quickly reproduced.

The leading indicator runs through South Korea. Sangdong is producing on allied soil. Qualification within South Korea’s allied industrial ecosystem can meaningfully reduce the barriers to broader allied procurement, although each program still requires its own qualification pathway.

The thesis breaker: China can weaponize oversupply as well as scarcity. If Beijing floods Western tungsten markets with low-cost powder the moment Western processing alternatives approach commercial scale - the move it made against Molycorp in rare earths - the scarcity premium compresses before it converts into sustained earnings. Separately, if China decides to restrict Korean powder access the way it restricted Japanese access, the Korean WF6 bridge supply breaks and the crisis deepens simultaneously with the GTP expansion still incomplete.

The Short Version: The mine opened. The conversion chain gap didn’t close. Two things increased in value today: the option value of every non-Chinese tungsten processing asset, and the pricing power of the handful of companies already qualified at the conversion layer. Those are different phenomena. The first is about strategic worth before earnings. The second is about margin expansion right now. Both are real. Both live downstream of the mine.

The Pattern

Two days in a row. Different mineral. Same conclusion.

Yesterday’s Note documented two non-Chinese architectures cracking the rare earth chemistry dependency - SPREC at the mine gate and MAIL in Belfast. The argument was that the processing layer is the bottleneck, not the mine.

Today’s Note proves the same thesis in tungsten. The mine response arrived. The processing chain is still the constraint. The supply response is ahead of the infrastructure it requires. Again.

That is not a coincidence. It is the structure of every critical minerals supply chain built on decades of Chinese vertical integration. The mine is always the visible part. The bottleneck is always one or two layers downstream, where the chemistry runs through infrastructure that takes years and billions to replicate - and where China’s control was quietest while it was being built.

Share

The piece that started this argument:

WF6 permanent cessation confirmed via The Elec (April 2026, primary Korean semiconductor trade press), TrendForce (June 12, 2026), and CTOL Digital Solutions. WF6 price data: 5N-grade up 232.7% year-on-year per ChemNet via TrendForce, June 12, 2026. Korean contract price increases of 70-90% confirmed via The Elec primary trade press. European APT price: $2,900-$3,180/mtu as of June 2, 2026 per China Tungsten Industry Association via CTIA; January 2026 base of $900-$940/mtu per Fastmarkets. Sangdong Phase 1 commissioning sourced from Almonty Industries (NASDAQ: ALM) primary company announcement, March 16, 2026. GTP capacity, sold-out status, and Sangdong exclusive supply deal sourced from Philadelphia Inquirer interview with Plansee CEO Karlheinz Wex, February 11, 2026. Plansee/GTP tungsten oxide reserve JV with Manhattan Five Partners sourced from CERATIZIT Group primary press release, May 19, 2026. SK Specialty Samsung contract sourced from SunSirs via KuCoin News, June 2026. All prices verified July 1, 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

The full archive is at williamdavid.substack.com. The processing layer argument begins in the Critical Minerals series.

The standard argument against Western rare earth independence has never been about mining. Mines can be funded, permitted, and built. The argument has always been one layer deeper: the processing - the chemistry that converts ore into separated oxides - remains Chinese. And a magnet built on Western ore but processed through Chinese chemistry is not a Western magnet. It is a Chinese-processed magnet that started somewhere else.

That argument was correct when the Critical Minerals series made it. Two separate developments - one from a Chilean-Brazilian miner, one from a Belfast recycling facility - are now testing whether it stays correct.

What the Processing Problem Actually Is

Rare earth separation uses solvent extraction (SX) - a chemical process that pulls individual elements out of solution through a long cascade of mixer-settler tanks using organic extractants. The problem is not the cascade. The problem is the feed going into it.

Dirty feed - ore with high impurity loads of aluminium, iron, and phosphate co-precipitated into the carbonate - turns a technically viable separation plant into an expensive failure. Aluminium is the primary culprit. It poisons SX circuits, slows campaign cycles, and collapses recovery of dysprosium and terbium - the heavy rare earths that remain the genuinely constrained half of any high-performance magnet.

China built its processing dominance not by discovering superior chemistry. It built it by controlling everything upstream of separation: the P204 and P507 organic solvents that conventional SX circuits depend on, the equipment, the feed preparation. The know-how accumulated over three decades is embedded in Chinese industrial infrastructure.

The Short Version: The mine is not the bottleneck. The feed is. China controls the feed preparation layer and the chemistry layer simultaneously. A Western separation plant running on Chinese extractants is still inside China’s leverage.

What Aclara Did

Aclara Resources (TSX:ARA) announced SPREC - Super Pure Rare Earth Carbonate - developed at its pilot plant in Santiago, Chile.

Conventional ionic clay mining produces MREC (Mixed Rare Earth Carbonate) at 40 to 60% rare earth content, impurities and all. Every other ionic clay miner ships this contaminated product to a separation plant and makes that plant’s job harder. Aclara’s CMH (Circular Mineral Harvesting) process - patented in Chile, Brazil, the US, and China - applies SX logic upstream, within the mining flowsheet itself, before the product leaves South America. The aluminium is removed at the mine site. A second upstream SX stage strips remaining non-rare-earth impurities. The output is approximately 99% rare earth carbonate - SPREC - before it reaches the Louisiana separation plant.

CEO Ramon Barua was direct in the company announcement: “Impurities are precisely the main source of complexity in the solvent extraction plant.” SPREC moves that complexity upstream, where it can be controlled, before it becomes an expensive downstream problem.

One caveat stated plainly: SPREC is at pilot scale in Santiago. Commercial application at Carina (Brazil) and Penco Module (Chile - EIA unanimously approved June 9, 2026) is still ahead. The pilot demonstrates the process works. Commercial scale has not yet been demonstrated.

The Short Version: Aclara moved the finish line upstream. The mine now ships a product the separation plant can actually work with. That changes the separation economics. It does not change the separation requirement.

What Ionic Technologies Did

Ionic Technologies, a spinout from Queen’s University Belfast operating as a wholly owned subsidiary of Ionic Rare Earths (ASX:IXR), attacked the same problem from the recycling side - and with different chemistry entirely.

Its MAIL (Multifunctional Amide Ionic Liquid) process replaces P204 and P507 at the separation stage. Ionic liquids have been studied for decades, but cost, scale, and the difficulty of industrial validation delayed adoption until Ionic Technologies built a demonstration plant in Belfast’s Titanic Quarter and proved the process on real-world magnet waste. MAILs are non-volatile, non-flammable, and operate with two orders of magnitude higher selectivity for rare earth separation than conventional organic solvents. The Belfast demonstration plant has produced neodymium, praseodymium, dysprosium, and terbium oxides at 99.5%+ purity from pre-consumer NdFeB magnet scrap.

The supply chain assembled around Belfast is worth naming precisely. Vacuumschmelze (VAC) - now the subject of a $1.9 billion acquisition by Energy Fuels - supplies pre-consumer magnet manufacturing waste. Ionic processes it in Belfast using MAIL chemistry. Less Common Metals (LCM) - the UK Department for Business and Trade’s identified sole Western rare earth metal and alloy manufacturer - converts the separated oxides into metals and alloys. Those alloys return to VAC for magnet production.

The world’s first electric vehicle motor built entirely from recycled rare earths was produced through a collaboration between Ionic Technologies, LCM, and Ford. A peer-reviewed Minviro study confirmed 61% lower carbon emissions versus primary supply.

The same caveat applies. The Belfast facility operates at 30 tonnes per annum demonstration scale. The commercial-scale facility has received an offer in principle for a £12 million UK government capital grant, subject to conditions. The 40x scale-up has not been built.

The Short Version: Belfast demonstrated the chemistry works without Chinese extractants. The commercial plant is funded in principle but not yet operating. The VAC-LCM-Ionic chain is real. The scale is still ahead.

What China’s Own Architecture Reveals

This is not happening in a vacuum.

On October 9, 2025, China’s MOFCOM announced MOFCOM Announcement No. 61 - the first time China formally exercised extraterritorial jurisdiction in its export control regime. The mechanism works on two triggers. First: products manufactured outside China that contain more than 0.1% Chinese-origin rare earth content by value require a Beijing-issued MOFCOM export license. Second: products manufactured outside China using Chinese-origin rare earth processing technologies require a Beijing-issued license. Both triggers apply regardless of whether any Chinese party is involved in the transaction.

Those extraterritorial provisions were suspended until November 10, 2026 as part of the Busan trade truce. They are currently paused. They resume in 133 days.

The strategic implication is direct. MAIL-based separation uses no Chinese extractants - its chemistry originates at Queen’s University Belfast, developed independently of Chinese processing know-how. SPREC’s closed-circuit ammonium sulphate system uses no Chinese-origin reagents. Neither architecture trips the technology trigger. Neither trips the content trigger for the separated product.

China’s own extraterritorial architecture has defined, by design, exactly which processing approaches remain outside its licensing reach. MAIL and SPREC are two of them. That is not coincidental to the investment argument. It is the investment argument.

The Short Version: The November 10 resumption date is the deadline most investors aren’t watching. When China’s extraterritorial licensing architecture reactivates, every Western processing operation that still runs on Chinese-origin extractants or Chinese-origin content re-enters Beijing’s compliance reach. The architectures that don’t are structurally different assets after that date.

What This Means for the Investment Picture

The bear case on Western rare earth processing rested on a structural assumption: the processing layer won’t be solved in time, which means any company positioned below that layer is building on a foundation it doesn’t control. That assumption is weakening - not because the January 2027 DFARS deadline has moved, but because the ecosystem around it is changing.

Here is the investment implication worth stating plainly. If cleaner, non-Chinese-origin feed becomes the standard for Western ionic clay projects, the value does not accrue only to Aclara. It accrues to every qualified downstream separation asset that receives cleaner feed. Iluka’s Eneabba refinery - now more than 50% complete, with a first binding offtake of 1,200 tonnes across four years from a global automotive OEM at a US$155 million minimum floor - becomes more economically viable when the feed arriving at its gate is cleaner. Lynas, MP Materials, NEO Performance in Estonia, Energy Fuels at White Mesa: any separation plant operating outside China that can receive SPREC-grade carbonate runs faster campaign cycles, recovers more heavy rare earth per tonne of feed, and operates at lower reagent cost per tonne of separated oxide.

The option value of existing Western processing assets has increased. Not because their cash flows have changed yet. Because the feed quality available to them is improving, the Chinese extractant dependency is being attacked on two fronts simultaneously, and China’s own extraterritorial licensing architecture - resuming November 10 - has confirmed precisely which processing approaches it cannot reach.

The qualification timeline has not compressed. Defense qualification standards hold under deadline pressure - they do not relax. The heavy rare earth gap is real and unchanged. The DFARS deadline lands January 1, 2027, and neither SPREC nor MAIL is at commercial scale before it.

But the bear case’s strongest argument was that Western processing couldn’t improve fast enough because the processing layer problem was structural and static. That assumption is no longer obviously correct.

The Short Version: Aclara moves the finish line upstream. Ionic removes the Chinese chemistry. Both increase the option value of every qualified Western separation asset downstream. China’s own extraterritorial architecture, resuming November 10, defines the boundary between assets that remain inside its reach and assets that don’t. The deadline hasn’t moved. The ecosystem around it has.

The law forced the market to solve a problem it had spent decades assuming China would solve instead. The market is solving it.

Share

The full archive is at williamdavid.substack.com. The processing layer argument begins in the Critical Minerals series:

Aclara SPREC announcement and Penco EIA approval sourced from Aclara Resources (TSX:ARA) company press releases, June 2026. Ionic Technologies MAIL process and Belfast facility details sourced from Ionic Technologies website, IonicRE ASX filings, and Magnetics Magazine (November 2024). VAC-LCM-Ionic supply chain partnership confirmed via Magnetics Magazine and IonicRE ASX filing (July 2025). Iluka Eneabba offtake sourced from Iluka Resources (ASX:ILU) company announcement, June 23, 2026. MOFCOM Announcement No. 61 extraterritorial provisions sourced from Jones Day, White & Case, and CMS Law primary legal analysis of MOFCOM text, October 2025. Suspension until November 10, 2026 confirmed via MOFCOM Announcements No. 70 and No. 72. All figures cited for context only. Facts verified June 30, 2026.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.

Most defense analysis stops at who builds the system. This framework asks a different question: who still runs when the minerals break?

The Defense Series built the argument. This is what it actually finds. This is not a list of defense companies. It is a test of which ones still function when the supply chain fails. Part 4 made the deepest point: the mandate was built for a world where one enforcement architecture reached everywhere. That world is less universal than it was. The adversary built its supply chain dominance inside a different system entirely - one that doesn’t need Western rules to function. That’s not an argument against watching these names. It’s the argument for being precise about which ones survive that condition.

On June 22 - the day the series concluded - China placed MP Materials and USA Rare Earth on its export control list. The two companies the US government invested most heavily in to replace Chinese rare earth dominance are now explicitly cut off from Chinese-origin dual-use inputs. Filter 3 just went live.

Three filters run through every name below. Did they clear the qualification gate? How contractually committed is the revenue? Can they power through a mineral supply disruption, or does the structural trap cut their production? The first two are how most defense analysis works. The third is what this series adds.

The qualification moat takes years to build. The mineral supply chain can be cut faster than any compliance program can adapt. That gap is the whole argument. This framework is being tested against a fixed deadline. When constraints hit, value shifts from building the system to keeping it running.

Eight situations. Four tension cases. One explicit rejection of the materials layer. The names not on this list - including the obvious ones - with the specific reason each doesn’t make it.

This is not a buy list. These are situations where the framework surfaces something worth understanding before it becomes obvious. Do your own research before acting on anything here.

All prices verified June 25, 2026 close. Financial data sourced from company earnings reports and public market data.

Tension Case: Kratos Defense & Security Solutions (KTOS) - $46.32

Category: Autonomous Platforms

The most exposed name in this piece - and the one that most directly validated the thesis. The CCA (Collaborative Combat Aircraft - the Air Force program developing autonomous drone wingmen for manned fighter jets) production award confirmed what the series argued: Kratos cleared the qualification gate for the platform type DFARS 252.225-7052 (the January 1, 2027 defense procurement deadline requiring supply chains free of Chinese rare earth involvement) specifically requires. That moat is real. No competitor replicates it quickly.

The tension is the most honest in this piece. GaN (gallium nitride - the compound semiconductor material inside radar and electronic warfare systems) components and NdFeB (neodymium-iron-boron - the permanent magnet technology inside every drone actuator and guidance system) magnets are both under Chinese export controls. Both were affected by the June 22 designation. The company that most directly confirmed the Defense Series argument is also the most exposed to its structural trap. The question isn’t whether Kratos matters. It’s whether it can source what it needs before the deadline that validates it.

Record $2B+ funded backlog. $14B+ opportunity pipeline. Q1 2026 revenue +22.6%. Down 65% from its 52-week high of $134.

What to watch:

Catalyst: DFARS compliance supply chain announcements. Q3 earnings July 30.

Confirms thesis: Domestic GaN and rare earth sourcing confirmed, production ramp on schedule, backlog converts.

Breaks thesis: Chinese export controls cut inputs before domestic alternatives arrive. Appropriations cycle trims the production award.

Framework score: 2/3

The verdict: The contract is real. The backlog is contracted. The exposure is embedded. All three are true at the same time.

Situation 1: Viasat (VSAT) - $60.00

Category: Secure Communications

Anti-jam SATCOM (satellite communications hardened against electronic warfare jamming) is the nervous system of the autonomous warfare stack. Defense Series Part 2 named it explicitly - every drone in every Replicator program requires it in contested environments. Viasat is the pure-play on that specific layer.

Three clean filter passes. Cleared the qualification gate for Space Force programs. Revenue is contractually committed through long-duration Space Force relationships - just won a US Space Force Protected Tactical SATCOM Global prime contract. The mineral insulation here is structural, not incidental: service revenue from existing satellite constellations doesn’t require new mineral procurement. New orders add capacity; they don’t restart the supply chain from raw material. That’s a different risk profile from every physical manufacturer on this list.

FCF (free cash flow - cash generated after capital expenditures) turning positive. Down 33% from its 52-week high of $89.79.

What to watch:

Catalyst: Space Force contract deployment milestones. FCF positive confirmation.

Confirms thesis: Space Force relationship deepens, FCF turns consistently positive, Replicator production ramp drives anti-jam demand.

Breaks thesis: Commercial aviation division underperforms and offsets defense growth. Satellite production faces GaN component delays.

Framework score: 3/3

The verdict: Three filters, three passes. The mineral insulation is structural. Down 33% from its high.

Situation 2: L3Harris Technologies (LHX) - $288.52

Category: Multi-Layer Defense Systems

L3Harris isn’t a single-layer company. Electronic warfare systems, WESCAM sensor turrets, AN/PRC-163 communications radios, Vampire counter-drone systems, missile solutions - it operates across six of the ten layers of the autonomous warfare stack simultaneously. That diversification is both the argument and the mineral insulation. No single GaN or rare earth disruption concentrates on one program and takes down the company.

The value case is the least dramatic on this list - LHX is down 24% from its 52-week high of $379. The argument isn’t the discount. It’s the gap between the current price and what a record $40.7B backlog at 1.4x book-to-bill ratio (new orders versus revenue - a forward growth signal) should command. Analyst consensus implies 24% upside. The ceasefire selloff treated LHX the same as every other defense name regardless of its structural position.

What to watch:

Catalyst: Q2 earnings. Revenue converting from the record backlog.

Confirms thesis: Revenue grows into backlog, margins expand, analyst targets close.

Breaks thesis: Backlog slippage on a major program. GaN supply disruption hits EW component production.

Framework score: 2.5/3

The verdict: Record backlog. 24% analyst upside. Less mineral exposed than the obvious alternatives. The entry isn’t as clean as the rest of this list. The thesis is.

Situation 3: Huntington Ingalls Industries (HII) - $279.09

Category: Naval Shipbuilding

The most structurally insulated name on this list. And the only pure-play large naval shipbuilder in the Western hemisphere.

That second fact matters more than it looks. There is no competitor, no alternative builder, no second source for Virginia-class and Columbia-class submarines. HII doesn’t win contracts against rivals. It is the contract.

Submarine construction is steel, titanium, copper, and nuclear propulsion - not rare earth permanent magnets, not gallium nitride. The structural trap that constrains every drone and missile manufacturer barely touches HII’s production line.

The revenue here isn’t policy-independent - nothing in defense truly is. But Columbia-class submarine contracts are among the most contractually committed revenue streams in the defense universe, extending through the 2040s. The FY2027 $65.8B shipbuilding budget exists regardless of ceasefire developments. Q4 2025 operating earnings +89.3% year-over-year.

Cheapest prime on EV/EBITDA (enterprise value to earnings before interest, taxes, depreciation, and amortization). Down 25% from its 52-week high of ~$370.

What to watch:

Catalyst: FY2027 budget passage. Columbia-class delivery schedule updates.

Confirms thesis: Budget passes, delivery holds, ceasefire selloff reverses as market reconnects stock to contractually committed backlog.

Breaks thesis: Newport News shipbuilding labor and cost overruns worsen. Columbia-class falls further behind schedule. Budget tradeoffs shift capital toward drones and AI at the expense of shipbuilding.

Framework score: 3/3

The verdict: The backlogs are contractually committed. The production line doesn’t run on rare earths. The stock moved 25% lower. The thesis didn’t.

Situation 4: BWX Technologies (BWXT) - $204.77

Category: Naval Nuclear

No alternative supplier for naval nuclear reactor components exists anywhere in the Western world. That isn’t a competitive advantage. It’s a monopoly that exists by design - the naval nuclear reactor program was built around a single supplier for security and continuity reasons, and that design has never changed.

The qualification gate was cleared and locked permanently. Revenue is among the most contractually committed in the defense universe; Columbia-class provides 20+ years of visibility regardless of administration. Near-zero mineral exposure - reactor components are specialized steel alloys, zirconium, and hafnium. The mineral constraint that complicates every other name on this list is structurally irrelevant to BWXT’s production.

Q1 2026: revenue +26%, EPS (earnings per share) beat by 22%, backlog +77% year-over-year. Analyst average target $226-238. Down 15% from its 52-week high of $241.82. SMR (small modular reactor - next-generation nuclear reactors sized for distributed power generation) licensing deal announced June 2026 adds civilian optionality analysts aren’t pricing in.

What to watch:

Catalyst: FY2027 shipbuilding budget passage. SMR commercial development milestones.

Confirms thesis: Backlog converts, Columbia-class on schedule, SMR attracts commercial capital.

Breaks thesis: Columbia-class schedule slippage ripples to component orders. Budget tradeoffs reduce shipbuilding allocation. P/E (price-to-earnings) of 54x leaves limited room for execution shortfalls.

Framework score: 3/3

The verdict: A monopoly by design. Near-zero mineral exposure. Backlog up 77%. The valuation is the only honest bear case.

Situation 5: Leidos Holdings (LDOS) - $100.00

Category: Defense Systems Integration

The only name on this list that clears all three filters at the operating layer. No physical production. No rare earth magnets. No gallium nitride. The system Leidos integrates and operates still runs when the minerals break. That sentence is the entire argument.

The market is pricing Leidos as if it’s a government consulting firm - the kind that gets cut when budget pressures arrive. That’s the wrong category. Mid-delivery systems integration contracts are infrastructure. Leidos isn’t advising the Defense Information Systems Agency on what to build. It built the thing, it’s running the thing, and stopping mid-delivery costs more than continuing. Advisory budgets get cut. Infrastructure contracts get completed.

$2.2B intelligence community contract. $760M NASA health services contract. $540M AI counter-terrorism contract won 2025. Cloud-based Joint Management Tool deployed for Defense Information Systems Agency and US Space Command June 2026. These are delivery commitments, not consulting arrangements.

P/E of 9.82. Down 51% from its 52-week high of $205.77. The DOGE (Department of Government Efficiency - the federal cost-cutting initiative) risk must be named: it has caused real contract reductions. The argument is that the market is pricing permanent structural damage into a revenue model more resilient than the fear suggests.

What to watch:

Catalyst: Q2 earnings August 4. Intelligence or space domain contract awards.

Confirms thesis: Revenue holds through DOGE environment, contracts prove sticky, P/E re-rates toward peers.

Breaks thesis: DOGE expands into mid-delivery systems integration. Intelligence community budget cut structurally, not cyclically. Program overruns on classified contracts create write-downs that aren’t visible until they hit earnings.

Framework score: 3/3

The verdict: The system still runs when the minerals break. Down 51% from its high. P/E of 9.82. The market may be confused about what kind of company this is.

Situation 6: Parsons Corporation (PSN) - $47.05

Category: Defense Intelligence Infrastructure

Parsons doesn’t look like a defense contractor. That’s why it’s here.

Cyber, missile defense, SIGINT (signals intelligence - intelligence derived from intercepted electronic signals), electronic warfare, space ground systems, classified sensors. The Altamira Technologies acquisition closed January 2026, adding missile warning, space, and cyber capabilities for $375M - deepening the stack precisely where Part 2 named as essential. Parsons builds and operates the defense intelligence infrastructure regardless of which platform wins any program competition.

Three clean filter passes. Deeply embedded in missile defense and space intelligence. Revenue is contractually committed through physical infrastructure contracts that don’t get cut mid-delivery. Near-zero mineral exposure - technology and engineering, no physical production.

Record $9.3B backlog. Book-to-bill 1.4x. Q1 EPS beat by 18%. Record adjusted EBITDA (earnings before interest, taxes, depreciation, and amortization) margin. P/S (price-to-sales) of 1.02. Analyst average target $78.25 implies 66%+ upside. Down 47% from its 52-week high of $89.50, near its 52-week low of $48.23.

What to watch:

Catalyst: Q2 earnings August 5. Classified contract awards in space or intelligence domains.

Confirms thesis: Backlog converts, margins hold, market re-rates away from DOGE fear.

Breaks thesis: Federal Solutions segment hit harder than expected. Altamira integration costs weigh on near-term margins. Classified program overruns create write-downs not visible until they surface.

Framework score: 3/3

The verdict: Relative to platform manufacturers, this is a business that degrades slower under constraint. The market is confusing it with something more vulnerable than it is.

Situation 7: Elbit Systems (ESLT) - $732.71

Category: International Autonomous Platforms

The highest risk - highest upside name on this list. And the international situation most defense conversations aren’t having.

Israel’s largest defense company. Combat-validated in the Iran conflict in real time - not a procurement thesis, a battlefield proof point. EO-IR (electro-optical/infrared - sensor technology detecting heat and optical signatures) systems, unmanned platforms, and electronic warfare hardware proven under live fire conditions no other Western defense company can credibly claim in 2026.

Cheapest valuation in the autonomous platform tier - P/S of 2.41x versus Kratos at 7.98x. Down 28% from its 52-week high of $1,016. Q1 2026 EPS beat analyst estimates by $1.03. $350M international contract announced during the quarter. European NATO rearmament creates direct incremental export demand.

Filter 3 is moderate. The sensor and actuator dependencies are real. Israel has been building alternative mineral supply relationships longer than US companies, but that doesn’t make the exposure zero. The bear case requires a direct statement: Elbit’s combat validation comes from the Iran conflict. That carries geopolitical complexity and reputational considerations for some investors that are real and material. That isn’t a thesis call. It’s a disclosure.

What to watch:

Catalyst: European NATO procurement decisions. International contract awards.

Confirms thesis: European demand accelerates, valuation gap to US peers closes, mineral supply chains prove more insulated than assumed.

Breaks thesis: Geopolitical complexity reduces institutional appetite before the valuation gap closes.

Framework score: 2.5/3

The verdict: The cheapest combat-validated autonomous platform name on this list. The bear case is geopolitical, not fundamental.

Tension Case: AeroVironment (AVAV) - $136.68 | Active Legal Risk

Category: Loitering Munitions

The most direct Iran war replenishment thesis on this list - and the one carrying the most complicated near-term risk alongside it.

Switchblade loitering munitions (small one-way attack drones designed to circle a target area before striking) consumed at emergency rates during the conflict. Record $1.10B funded backlog in the Autonomous Systems segment. 13 consecutive Buy ratings from analysts, average target $267-301. Down 67% from its 52-week high of $417.86.

Two things require direct disclosure before the thesis. First, a securities class action lawsuit was filed in June 2026 alleging misrepresentations about the $1.7B SCAR Space Force contract, which was subsequently terminated for convenience by the Space Force. The lawsuit is active. The outcome is uncertain. Second, Filter 3 fails - GaN and NdFeB dependencies are identical to Kratos. The Clearing Trap from Part 4 also applies: OMB (Office of Management and Budget) approved doesn’t mean paid. The appropriations cycle is the funding leg that can break the framework agreement even when demand is real.

The loitering munitions thesis is intact. The Space Force legal overhang is not.

What to watch:

Catalyst: Q4 and full fiscal year 2026 earnings June 29 - if publishing after this date, check results and update this section. Securities lawsuit developments. Replicator 2 production award timing.

Confirms thesis: Lawsuit resolved without material damage, Replicator 2 awards arrive, domestic mineral sourcing confirmed.

Breaks thesis: Lawsuit expands, Space Force contract loss proves structural, appropriations cycle delays Replicator 2.

Framework score: 2/3 on the filters. Additional binary legal risk.

The verdict: The funded backlog is contracted. The exposure is embedded. The constraint is active. The legal overhang is the variable none of the other situations carry.

Situation 8: Rocket Lab (RKLB) - $80.69

Category: Space Launch

This name is not here for valuation. It’s here because it’s the cleanest way to own the ISR (intelligence, surveillance, reconnaissance) layer of the sovereign stack.

Responsive small-satellite launch for rapid constellation reconstitution - the surveillance layer of the Part 2 stack made investable as a standalone company. The Iran conflict validated satellite criticality for targeting and battle damage assessment in real time during this series’ publication window. That’s an operational fact from an active conflict, not a thesis.

The Electron launch vehicle runs on conventional propellants - mineral-resilient at the core product level. The Space Systems segment carries more mineral exposure, but the thesis is the Space Force relationship, not component manufacturing.

The value case is the weakest on this list - trading well below analyst consensus of ~$97-103, down 47% from its ATH (all-time high) of $151. KeyBanc upgraded with a $135 target following the SpaceX IPO-related selloff. The SpaceX pressure is real. The entry is murkier than every other situation here.

What to watch:

Catalyst: Nasdaq-100 inclusion. Space Force contract expansions. Neutron development milestones.

Confirms thesis: SpaceX overhang clears, Space Force relationship deepens, Neutron timeline holds.

Breaks thesis: SpaceX captures Space Force launch share directly. Electron anomaly disrupts cadence.

Framework score: 2.5/3

The verdict: The ISR layer of the sovereign stack. The entry point is the weakest on this list. The strategic position is not.

The Tension Cases

Four names carry the tension case designation. Kratos and AeroVironment are covered above - both pass the qualification gate and both fail Filter 3 directly. The thesis and the structural trap are the same fact seen from two directions. The two below complete the picture.

Mercury Systems (MRCY) is the near-monopoly on flight-qualified edge AI compute modules for US military drone programs. The qualification moat is real. MRCY is also the early warning indicator for GaN supply chain disruption - the first place margin compression shows up when Chinese gallium exports tighten. China placed MP Materials on its export control list the same week MRCY reached its all-time high of $122.17 on June 15. The stock has since pulled back to $105.57 - down 14% from its high. The thesis and the risk arrived in the same week. Watching, not acting.

Leonardo DRS (DRS) is the pure-play military EO-IR specialist with fewer than ten sell-side analysts. The institutional discovery argument is genuine. The GaN dependency in its sensor technology is equally genuine. It’s 10% from its 52-week high. The thesis is right. The discount isn’t there yet.

The Materials Layer - An Explicit Rejection

MP Materials and USA Rare Earth pass Filter 1 - Mountain Pass is the only active rare earth mine in the Western hemisphere and both companies have cleared the DoD qualification gate. MP fails Filter 2 - the DoD price floor is a permanent subsidy, not a bridge. Strip the $42.3M in grid electricity sales and the rare earth processing operation runs at a loss. Both fail Filter 3 most directly - they are the chokepoint. China chose to target the companies most central to solving the rare earth dependency specifically.

Strategic importance does not equal investment return. The thesis is real. The policy dependency is disqualifying until domestic input sourcing is confirmed and the DoD price floor is no longer required to sustain the operation. Those conditions don’t exist today.

Names Not on This List - and Why

Lockheed Martin, Raytheon Technologies, Northrop Grumman, and General Dynamics validated the mandate through their backlogs. The mandate is in their price. The investor who waited for this framework to find LMT missed the move. These are the primes that confirmed the thesis. They aren’t the thesis.

Moog (MOG-B) makes every Patriot interceptor actuator and every F-35 flight control system. The Iran war confirmed the thesis. Significant dysprosium (a heavy rare earth element - the material inside high-performance permanent magnets that prevents demagnetization at operating temperatures) exposure in every actuator. Three filter passes, all legitimate. The stock reached its all-time high of $420.65 on June 18, 2026, before pulling back to $420.00 at the June 25 close.

[NEW INTELLIGENCE — June 2026] Moog announced a partnership with Niron Magnetics in December 2025 to develop rare-earth-free Iron Nitride actuator designs for guided munitions defense applications. The partnership is in evaluation stage - not commercial deployment. Current production actuators retain the dysprosium dependency. The direction of travel confirms that Moog recognizes the mineral exposure as a material risk worth solving. That recognition is itself evidence of Part 4’s structural trap: the adversary built the dependency over thirty years inside a system that doesn’t need Western rules to function. Iron Nitride is Moog’s acknowledgment that the trap is real. It isn’t the exit. This strengthens the discipline argument: the framework finds it interesting. The stock near all-time highs means it doesn’t find it actionable.

This is what discipline looks like: the framework finds it interesting. It doesn’t find it actionable. Those are different things.

Booz Allen Hamilton (BAH) is down 37% from its reference price. The discount is real. The DOGE risk is structural, not cyclical. Recovery requires a policy shift. That makes it policy-dependent in the wrong direction - Filter 2 cuts it.

Mercury Systems (MRCY) is named above in the tension cases. Reached its all-time high of $122.17 on June 15 before pulling back to $105.57. Down 14% from its high. Highest GaN exposure on the list. The thesis and the risk are the same fact seen from two directions.

HEICO (HEI) has fifty consecutive profitable years and a permanent tailwind from the DoD cost-reduction mandate. The connection to this series’ specific argument is less direct than the names above. Next series candidate.

A Note on the Framework’s Assumptions

This framework rests on one bet: that the mineral constraint becomes binding before substitution, stockpiling, or policy adaptation closes the gap. That assumption may be wrong. Japan adapted after the 2010 rare earth shock. Substitution pathways exist - LFP (lithium iron phosphate - a battery chemistry that eliminates cobalt and reduces rare earth dependency) batteries, reluctance motors, alternative sensor architectures. The market may already be pricing adaptation speed faster than disruption speed. If it is, Filter 3 loses its edge and the relative advantage compresses.

The system might also fail first somewhere else entirely. Budget reallocations, software integration delays, and classified program overruns have historically disrupted more defense programs than material shortages. If that’s the failure mode, the integration and infrastructure names on this list face a different kind of exposure than the one the framework surfaces.

The case for the framework is this: January 1, 2027 is six months away. Adaptation takes longer than six months. The Vulcan Elements magnet facility - one of the largest Western rare earth magnet buildouts underway, targeting 10,000 metric tonnes annually - hits meaningful production in 2027 and 2028, after the deadline. The gap between what the law requires and what exists is not theoretical. It’s documented, dated, and closing slower than the clock.

That’s the bet. It could be wrong. Know what you’re holding.

The system has been mapped. The constraints are now investable.

The highest-quality exposure sits with the companies that clear all three filters and are priced as if they don’t.

January 1, 2027 does not move. The supply chain it requires is behind schedule. The gap between what the law requires and what currently exists is not closing. The market is still pricing as if it does.

That is the trade.

Share

The full archive is at williamdavid.substack.com

All prices verified June 25, 2026 close. Financial data sourced from company earnings reports and public market data.

This post is for informational purposes only and is not investment advice. The Chokepoint is an independent investment research publication. Nothing in this publication should be construed as a recommendation to buy, sell, or hold any security. All company references and price data are provided for informational and contextual purposes only. Conduct independent due diligence and consult a qualified financial advisor before making any investment decisions.