The Model Commons

The problem

In 1862, in the middle of the Civil War, Congress passed the Morrill Act. It transferred federal land to the states on the condition that the proceeds endow colleges teaching agriculture and the mechanic arts. It was not charity and it was not a research program. It was a bet that if you put the means of learning within reach of ordinary people in every state, the country would compound.

Cornell, MIT, Berkeley, Texas A&M, Purdue, and dozens more came out of that bet. It is arguably the highest-return act of American economic policy ever passed, and it worked by making a scarce input — advanced knowledge — into public infrastructure rather than a private good.

The scarce input now is compute, and it is not public infrastructure.

The five largest American cloud and AI infrastructure providers have committed to capital expenditure somewhere between $660 and $690 billion in 2026 — roughly double 2025 levels, with about three-quarters of it going to AI-specific infrastructure. Amazon alone is projected near $200 billion, Alphabet $175 to $185 billion, Meta $115 to $135 billion, Microsoft above $120 billion.

Against that, America's public compute effort is the National AI Research Resource. The NAIRR pilot has been a genuine success on its own terms: more than 600 research projects and 6,000 students across all fifty states, with roughly 3.77 exaFLOPS aggregated from federal partners including the DOE national laboratories, and the National Science Foundation is now standing up an operations center to move it from pilot to program.

It is also, next to $660 billion, a rounding error.

The consequence is not that American research stops. It is that the frontier moves inside a handful of firms, that the graduate student and the state university and the small company work at borrowed scale on borrowed terms, and that the ability to ask a question of the frontier becomes a commercial relationship rather than a civic capability. A country in which only five organisations can train a frontier model is a country in which only five organisations can check one.

The proposal

Make NAIRR permanent and fund it at national scale. Fund open-weight American models as public infrastructure. Attach the land-grant condition: what the public pays to create, the public can use.

How it would work

1. NAIRR as a permanent institution, not a pilot. Statutory authorisation, a decade-scale appropriation, and an operations center with its own budget line — the NSF is already building this and it should be finished rather than left to annual discretion. Allocation by scientific merit, open to universities, non-profits, startups, and state and tribal agencies, as the pilot already is.

1. Public compute at the national labs. The DOE already runs the machines and the security model. Expand the federal fleet with a standing share reserved for public-interest use: model evaluation and safety testing, health and materials research, and any domain where the commercial return does not justify frontier-scale training but the public return does.

1. Fund open-weight models with public money. Not a state-run frontier lab competing with industry. Targeted funding for open models with published weights, published training data provenance, and permissive licences — the way the government funds reference genomes, reference implementations, and standard reference materials. The purpose is a floor, not a champion: an American open model good enough that no small company, school district, hospital, or state agency is forced into a single vendor.

1. The land-grant condition. Any model trained principally on publicly funded compute publishes its weights and evaluations, subject to a narrow security exception reviewed on a published schedule. This is the Morrill bargain: public means, public access.

1. Compute for evaluation, explicitly. A standing allocation for independent testing of commercial systems — the capability that makes № 027's right to a human decision enforceable, and that currently exists mostly inside the firms being evaluated.

The numbers

Set the target at $15 billion a year for a decade for public compute and open-model funding combined.

That is a large number in the language of science budgets — roughly a third of the NIH — and a small one against what it sits beside. It is about 2 percent of a single year's private AI capital expenditure. It would not put the government anywhere near the frontier and is not meant to.

What it buys, on the pilot's demonstrated ratios, is public compute capacity one to two orders of magnitude beyond NAIRR's current 3.77 exaFLOPS, which would move a state university from running fine-tunes to running real experiments.

The Morrill comparison is worth doing honestly. The land grants transferred roughly 10 million acres, whose value at the time was a substantial share of federal assets, to institutions that took decades to produce a return. The proposal here is comparatively cheap and asks for far less patience. If it returns a hundredth of what the land-grant colleges returned, it is the best money in this catalog.

The honest objections

"Government compute will be obsolete before it is installed." The most practically damning objection. Public procurement cycles run years; hardware generations run months, and a federal cluster specified today may be two generations behind at ribbon-cutting. This is a real and partly unavoidable cost. The mitigations are to buy access as well as iron — allocations on commercial clouds alongside owned capacity — and to fund on a rolling replacement basis rather than as capital showpieces. We would rather fund a smaller, continuously refreshed capability than a large monument.

"Open weights are a proliferation risk." The serious safety objection, and we do not think it resolves cleanly in either direction. Open models are the basis of most independent safety research and most of the competitive floor; they are also irrevocable once released. Our position is the narrow one: publication is the default for publicly funded models, with a defined security exception, decided on published criteria, reviewed on a schedule, and reported to Congress. Anyone who claims this tradeoff is obvious in either direction is not being straight with you.

"This subsidises the AI industry's customers and props up demand." Partly, and the design should minimise it by favouring open outputs and public-interest use over commercial capacity. But note the alternative: without a public floor, the entire research and evaluation ecosystem is a customer of the firms it is meant to study. That is a worse subsidy, paid in independence.

"Why not just regulate the labs instead?" Regulation and capability are complements. A regulator with no compute cannot verify a claim, cannot reproduce a result, and cannot test a system it is regulating. The evaluation allocation exists exactly because rules without instruments are aspirations.

"$15 billion a year, when this catalog also wants a caregiver's wage and a permanent fund?" A fair challenge and the honest answer is sequencing rather than denial. Of everything in this block, this is the one whose return is most speculative and most distant, and if only some of these pass, it should not be first. The Morrill Act took forty years to look like genius. We are asking for the bet, not claiming the payoff is near.

Sources

• Morrill Land-Grant Acts of 1862 and 1890, 7 U.S.C. § 301 et seq.
• National Science Foundation, National AI Research Resource pilot — 600+ projects, 6,000 students, all 50 states, approximately 3.77 exaFLOPS from federal partners; NAIRR Operations Center solicitation (nsf.gov, nairrpilot.org)
• U.S. Department of Energy, NAIRR pilot allocations at the national laboratories (energy.gov)
• Hyperscaler 2026 capital expenditure guidance of $660–690B across the five largest US providers, roughly 75% AI-related
• Proposals № 020 (The Sovereign Equity Act), № 027 (The Right to a Human Decision), № 030 (The Compute Reserve)