Who Controls the Stack

The Cerebras IPO filed in early May 2026 with a price range of $115–125. By the time the book closed, demand was more than 20x oversubscribed and the range had been lifted to $125–135. That is not routine oversubscription. It is a market declaration that AI compute infrastructure is scarce — and that institutional capital has priced that scarcity explicitly.

That single data point connects to every other supply chain signal from the same week. Not because they involve the same company, but because they involve the same stack.

The Mineral Layer

On January 1, 2027, two expansions to DFARS — the Defense Federal Acquisition Regulation Supplement — take effect simultaneously. The first prohibits Chinese magnets in US defense contracts, not just at the melting and production stage as current rules require, but from the point of mining through final production. The second applies identical restrictions to tantalum. Full supply chain exclusion, confirmed.

The current DFARS (effective through December 31, 2026) only captures Chinese magnets at the processing stage. A magnet smelted in a non-Chinese facility but using Chinese-mined ore is technically compliant today. The 2027 expansion closes that gap. The explicit target is the upstream source, not the downstream process.

This is not a tariff. It is an architectural decision. The US defense industrial base is severing Chinese critical mineral inputs at the origin, across the full vertical. The magnets at issue — neodymium-based permanent magnets — appear in missile guidance systems, radar arrays, submarine propulsion, and drone actuators. Tantalum goes into capacitors throughout military electronics. The policy logic is precise: if the mineral can be traced to Chinese soil, it is out.

The Fab Layer

The same week, details clarified on the Sony-TSMC joint venture in Kumamoto, Japan. Sony takes a majority and controlling stake. TSMC holds a minority position. The facility sits in Koshi City, inside a new Sony fab under construction for the purpose.

The target market shift matters as much as the ownership structure. This JV is not aimed at smartphones or consumer electronics — the domains where Sony's CMOS image sensors have historically dominated. The targets are automotive cameras, lidar systems, and robotics visual recognition. These are the sensor systems that will operate in the same physical environment as the drone platforms being funded under Japan's SHIELD defense program.

Japan's government has already committed substantial subsidies to TSMC's first Kumamoto fab, JASM. This JV is on the same policy track — one more data point in a deliberate national strategy to internalize semiconductor supply chain capacity within an allied nation. The subsidy architecture and the joint ownership structure are both intentional: sovereign fab capacity, built with explicit government support, targeting industrial sectors that sit directly adjacent to defense applications.

The Compute Layer

Cerebras designs wafer-scale AI chips — individual dies that span an entire silicon wafer rather than the smaller chiplets that most AI chip companies produce. The 20x-plus oversubscription of its IPO as of May 2026 reflects a simple market signal: investors believe AI compute infrastructure is capacity-constrained, and they are paying a premium to own exposure to that constraint.

The immersion cooling market forecast published the same week — covering the period through 2037 — names the next constraint clearly. As chip TDP (thermal design power) climbs, driven by larger dies, denser interconnects, and higher clock rates, air cooling reaches a physical ceiling. Immersion cooling, which submerges compute hardware directly in dielectric fluid, is the engineering response. The market includes Green Revolution Cooling, Midas, Iceotope (backed by Meta), Chemours, and Alfa Laval. The constraint is not demand — it is the physics of heat dissipation at scale.

The Golden Pass LNG facility on the Texas Gulf Coast recorded its first feed gas demand above 0.4 BCF/day as of May 9, 2026, a record for the facility. Early commissioning data projects 0.46 BCF/day — approximately 17.6% of total capacity. This is the energy infrastructure connecting North American natural gas production to the power grids that will supply AI data centers in allied nations. The mineral exclusion policy and the fab reshoring strategy both require power. LNG is part of how that power gets there.

Three Layers, One Stack

The pattern is structural. Take it vertically.

The mineral layer is being reorganized by policy. DFARS 2027 does not merely restrict procurement — it signals to the entire US defense industrial base that Chinese mineral inputs are being phased out across dozens of materials categories, forcing qualification of alternative sources through a process that takes years, not months. The policy creates a structural demand for non-Chinese mineral supply chains that did not exist at this scale before.

The fab layer is being rebuilt in allied nations under government subsidy. The Sony-TSMC Kumamoto JV sits alongside TSMC Arizona, the planned TSMC Germany facility, and Samsung's Texas expansion. Each represents sovereign capacity constructed outside China, specifically positioned for the industrial sectors — automotive, robotics, defense — where the next decade of compute demand will concentrate. These are not market-driven location choices. They are policy-driven infrastructure decisions, with governments as co-investors.

The compute layer is signaling genuine scarcity. A 20x-oversubscribed IPO at a lifted price range is not sentiment. It is capital pricing a structural supply shortage in AI computing infrastructure — and doing so in public, with names and numbers attached.

These three layers are not parallel trends. They are vertically connected. Minerals are the inputs to the fab process. Fabs produce the chips. Chips run the AI workloads. The chokepoints at each layer are being actively contested — through procurement policy, through capital deployment, through geopolitical industrial strategy operating at the nation-state level.

The Open Question

The throughline is correct: the open question is who controls the chokepoints at the intersection.

The answers are not settled. US policy is clear on mineral exclusion, but the alternative source qualification process is slow — multi-year timelines for qualifying new suppliers across materials categories that were previously sourced without scrutiny. Allied fab capacity is expanding, but still years from matching China's production scale in the relevant material classes. AI compute demand is already outpacing supply, and the cooling constraint means that even as new fabs come online, the physical infrastructure to run the chips remains a binding bottleneck.

What is settled is the direction. The stack from minerals to compute is being reorganized around a specific set of allied nations and supply chain architectures. The chokepoints are moving. Capital markets — as the Cerebras IPO shows — are already pricing the reorganization.

On-chain perpetuals on these markets trade 24/7. The structural shifts in mineral supply chains, semiconductor capacity, and AI infrastructure are exactly the kind of slow-moving, high-conviction positions where timing is the variable. Blackboard.