The Compute Reserve

The problem

Taiwan produces something over 90 percent of the world's advanced-process semiconductors. At the leading edge the concentration is tighter still: a single company, TSMC, derives roughly three-quarters of its wafer revenue from nodes at 7 nanometers and below, with about a quarter from 3-nanometer alone.

Every argument in this block — the grid load in № 024, the displaced worker in № 026, the public compute in № 029 — sits downstream of a supply chain with one physical chokepoint, 100 miles off the coast of a country that claims the island.

The CHIPS and Science Act of 2022 put $52.7 billion into changing that, and it has moved real capacity onshore. It is also a supply-side program with a decade-long lead time, and it addresses fabrication while the tighter bottleneck has often been advanced packaging — the step that assembles logic dies and high-bandwidth memory into a finished accelerator, capacity for which has been the binding constraint on AI hardware output more often than lithography has.

So America has begun building fabs and has no answer at all to the question the fabs do not address: if access were cut tomorrow, what could the United States actually call on, and how fast?

There is no mechanism. No reserved capacity, no priority claim, no pre-negotiated surge. The Defense Production Act allows compulsion after the fact, which is a blunt instrument that arrives late and poisons the commercial relationships it seizes. That is the entire toolkit.

The proposal

A Compute Reserve built on contracted options, not stockpiled hardware: the United States pays a premium for a pre-negotiated priority claim on domestic fabrication, packaging, and datacenter capacity, callable in a declared emergency.

The instinctive design — buy chips, store them — is wrong and should be rejected before anyone builds it. The Strategic Petroleum Reserve works because oil in a salt dome is the same oil in ten years. An accelerator in a warehouse loses most of its relative value in three, because the frontier moves. A national stockpile of GPUs would be a monument to depreciation.

What does not depreciate is a claim on future output.

How it would work

1. Capacity options, bought like insurance. The government pays an annual premium to domestic fabs, packaging facilities, and datacenter operators for the right to call a defined share of their output or capacity at a pre-agreed price during a declared emergency. Firms keep selling commercially in normal times; the public buys the option, not the product. This is how airlines hedge fuel and how the Defense Department already contracts for sealift through the Civil Reserve Air Fleet, which is the right precedent and an American one.

1. Packaging first. Priority goes to advanced packaging and high-bandwidth memory rather than to fabrication alone, because that is where the queue has actually formed. A proposal that funds only fabs is fighting the last bottleneck.

1. A rolling, small physical buffer. A deliberately modest inventory of current-generation accelerators and critical components, continuously rotated into public compute (№ 029) as it ages rather than left to rot. The buffer covers weeks, not years. Its purpose is to bridge to the options, not to substitute for them.

1. Power as part of the reserve. A fab or a datacenter without firm power is a building. Reserve contracts must include interconnection and generation commitments, coordinated with the large-load rules in № 024 so that the reserve is not itself a hidden charge on household bills.

1. Publish the posture. An annual unclassified statement of how much capacity is under option, at what cost, and what it would supply in an emergency. A reserve nobody can audit is a press release.

1. Allies, not autarky. Options extended to fabs in Japan, South Korea, and the Netherlands' equipment supply chain. Full domestic self-sufficiency at the leading edge is not achievable this decade at any price, and pretending otherwise produces expensive theatre.

The numbers

Options are cheap relative to the thing they cover, which is the whole argument for this structure.

Suppose the target is a callable claim on 10 percent of domestic advanced packaging and accelerator output. Against 2026 private AI capital expenditure running $660 to $690 billion across the five largest providers, the relevant domestic hardware slice is on the order of low hundreds of billions. An option premium in the range of 2 to 4 percent of the covered value implies something like $2 to $5 billion a year.

That is under a tenth of the CHIPS Act's one-time appropriation, paid annually, for a capability the CHIPS Act does not provide. It is also roughly a third of what № 029 asks for public compute, and the two should be built together: the physical buffer rotates into the Model Commons, and the Commons gives the reserve somewhere useful to put hardware in peacetime.

The honest uncertainty: nobody has priced options on fab capacity at scale, because the market does not exist. The 2 to 4 percent figure is imported from other capacity-reservation markets and could be badly wrong in either direction. The first tranche should be small and competitively bid precisely to discover the price.

The honest objections

"This is industrial policy with a hedging vocabulary." It is industrial policy. We are not going to pretend otherwise, and the case rests on a specific market failure: no private actor internalises the national cost of a Taiwan contingency, so no private actor buys enough insurance against it. That is the textbook condition for public purchase. What we would resist is the version that becomes permanent subsidy to incumbents, which is why the premiums should be competitively bid and the posture published.

"An option is worthless if the emergency is a war. The government will just seize capacity anyway." The sharpest objection. In a true crisis the Defense Production Act supersedes contracts, so what exactly is being bought? Three things a seizure does not provide: the capacity exists to be seized only if someone invested in it, and the option premium is what funds that investment; the terms are pre-negotiated, so the call happens in days rather than after months of litigation; and firms that have been paid for years have planned for the call. Seizure is a legal power. This buys readiness.

"Ten percent of output is nowhere near enough to matter." Probably true for a total cutoff, and we should not claim this substitutes for onshoring. It is a bridge measured in months, meant to prevent the worst weeks while other measures spin up. A proposal that promised self-sufficiency would be lying about physics and lead times.

"Depreciation kills the buffer too, not just a stockpile." Correct, which is why the buffer is small and why it rotates into public compute rather than sitting idle. Some value is lost. The alternative — no physical bridge at all — has a worse failure mode.

"Allied options are exactly the dependency you are trying to fix." Partly. But a portfolio of dependencies across four allied jurisdictions is meaningfully different from a single point of failure, and the equipment chokepoint in the Netherlands is not substitutable at any price this decade. Honest risk reduction beats dishonest independence.

Sources

• Taiwan's share of advanced-process semiconductor production, in excess of 90%; TSMC wafer revenue by node, with 7nm and below at roughly 74% and 3nm at about 25%
• CHIPS and Science Act of 2022, $52.7 billion in semiconductor manufacturing and R&D funding (nist.gov)
• Advanced packaging and high-bandwidth memory as the binding constraint on accelerator output
• Civil Reserve Air Fleet and Voluntary Intermodal Sealift Agreement as capacity-option precedents (defense.gov)
• Defense Production Act, 50 U.S.C. § 4501 et seq.
• Hyperscaler 2026 capital expenditure guidance of $660–690B
• Proposals № 024 (The Grid Dividend), № 029 (The Model Commons)