Public AI Infrastructure

Nearly every notable AI model is now built inside a company, on computers few universities can afford. The national resource meant to widen access still runs as a pilot, on contributed capacity, with no statute behind it. Congress should put that resource in law, open it to newcomers first, buy from many suppliers, and build public capacity where it costs less over time.

The Problem

In 2025, industry produced 93 notable AI models; academia produced two.¹ The cost of the largest training runs has grown about 2.4-fold a year since 2016 and, on trend, will pass $1 billion by 2027.² The White House's 2025 AI Action Plan says long-term cloud contracts are "far beyond the budgetary reach of most academics and many startups."³

The gap reaches past campus. Independent scientists test whether models are safe and accurate, universities train the engineers every company hires, and public-interest work rarely pays for itself. The talent is growing: new AI PhDs in the United States and Canada rose 22% from 2022 to 2024.⁴ The computing power sits elsewhere.

Washington already owns world-class machines: Department of Energy labs run three of the world's four fastest supercomputers.⁵ Yet the National AI Research Resource, NSF's effort to open AI computing to researchers, began in January 2024 as a pilot drawing on 10 other agencies and 25 private and nonprofit partners.⁶ It has supported over 800 projects,⁷ but current law authorized only a task force to plan it.⁸ Three problems keep access narrow:

1. Access runs on goodwill. The pilot's private partners include Anthropic, Google, Meta, Microsoft, NVIDIA, and OpenAI,⁹ so researchers who test frontier systems can end up relying on capacity donated by the companies that build them.
2. Capacity fills fast. New York's first Empire AI system reached maximum capacity, with more than 200 researchers, four months after launch.¹⁰
3. Funding stays small. The task force estimated a full national resource at about $2.6 billion over six years;¹¹ NSF's new operations center is a $35 million, five-year award.¹²

Why legislation: No statute establishes the National AI Research Resource; the pilot runs on agency budgets and partners' contributions, and it can shrink as quietly as it grew. Both parties want it to last. The 2025 AI Action Plan calls for "a lean and sustainable NAIRR operations capability,"³ and in June 2026 the House Science Committee voted 29–0 to advance the CREATE AI Act, which would establish the resource in law.¹³ Americans should not need a company's permission to study the technology reshaping their lives.

The Solution

A four-step staircase: each step stands alone, and each step up adds public money and public ownership. It serves U.S. universities, nonprofits, small businesses, and public institutions; "utility" means dependable, fairly allocated public access, and private clouds are neither regulated nor acquired. Antitrust for AI markets and a federal research lab on AI's harms are addressed separately.

Step 1 — Make the resource permanent. Establish the National AI Research Resource in statute at NSF, as the CREATE AI Act would, offering computing time, licensed datasets and models, secure research environments, and technical help. Scale security and misuse controls to the work, and protect confidential data.

Step 2 — Open the door to newcomers. Allocate by published scientific and public-benefit criteria under conflict-of-interest rules, and reserve a meaningful share for smaller institutions and first-time applicants. Publish allocations and waiting times so Congress can see who gets served.

Step 3 — Buy from many, answer to none. Procure from multiple qualified providers, require that data and workflows can move between them, and bar any donor or supplier from vetoing lawful research findings. Publish publicly funded results where feasible, with exceptions for privacy, security, and small firms' commercial work.

Step 4 — Build public capacity where it pays. Fund five years in stages, tied to a published demand-and-cost plan that compares owned hardware with leased capacity, including energy, water, upgrade, and exit costs. Where independence and lifetime cost justify it, build publicly controlled capacity at national labs, which already pair public sites with private builders, as in Argonne's planned 100,000-GPU Solstice system.¹⁴ Independent science needs independent machines.

Where to start: Step 1 is the floor; a House committee has already approved it without dissent. Step 2 is the heart: a public resource should reach the researchers the market does not.

Administration and enforcement: NSF, coordinating with the Department of Energy, issues an implementation and procurement plan within 12 months and opens competitively allocated access within 18 months, subject to appropriations. Inspectors general and GAO audit spending and allocation; misuse brings suspension and recovery, with notice and appeal.

Risks and Mitigations

• An expensive asset that ages fast: Chips date quickly, and public projects run over budget. Buy in stages, compare owned and leased capacity, and count upgrade and exit costs before building; public ownership must earn its place, and some purchases will still age badly.
• Capture, by companies or politics: Scarce capacity invites favoritism. Diversify suppliers, disclose conflicts, protect publication, and use reviewable merit criteria; allocation choices will still be contested.
• Dangerous uses and sensitive data: Public access should not mean unrestricted access. Match controls to the risk of the work, keep sensitive data in secure environments, and hold incident response accountable; no control stops every misuse.

Similar Bills

Fit measures similarity to this proposal's mechanisms: High = direct precedent; Partial = useful component with material differences; Related = adjacent approach.

Federal

State

What this adds: The CREATE AI Act would put the national resource in law. This proposal adds what keeps it independent and fair: reserved access for newcomers, published allocation data, multi-supplier procurement with no donor vetoes, and a tested path to publicly owned capacity where it saves money.