NEWSLETTER · FRONTIER LABS AS NATIONAL LABORATORIES · October 2026 · read online (no email list yet)

Frontier AI labs are already treated like national labs. Just not the way that counts.

The US now has a voluntary AI accord and a task force. It still lacks the contract that makes national labs accountable.

The short version

Frontier AI developers get lab-style treatment from government, but through one-off levers, not the contract that makes national labs open, graded and replaceable. Since late September the White House has added a voluntary accord and a task force, and courts have started doing part of the job a contract would do. The likeliest next steps are still more of the same levers, conditions on chips, and labs as users of government compute.

Why now

Three things landed in the weeks since late September. On September 29 the President and six AI companies signed the White House Accord on Super Intelligence.^[1] On October 13 Anthropic asked the full DC Circuit to rehear the case over its Pentagon designation; the government’s response is due November 13.^[2] On October 15 the task force built to carry out the accord adopted its charter, which starts a 120-day clock.^[3] On October 8 eleven companies also pledged $2.4 billion to the Genesis Mission, the federal AI-for-science program.^{[4] [5]}

If you follow AI companies, what government wants from them is now a business question.

What a national lab actually is

A national laboratory is a government-owned facility run by a contractor the government can grade and replace.^{[6] [7]} The Department of Energy has 17 of them, 16 run under management-and-operating contracts.^[6] The contract is the point. In 2018 the operator of Los Alamos “would lose the work after successive poor ratings in their performance evaluation, specifically related to safety,“ and the contract went to Triad National Security, LLC.^{[7] [8]}

The second feature is openness. DOE user facilities are “open to all interested potential users without regard to nationality or institutional affiliation,“ allocate time by merit review, charge nothing for work published in the open literature, and require full cost recovery for proprietary work.^[9] Since 1985 the rule for fundamental research has been that it stays unrestricted unless it is classified.^[10]

What frontier labs get instead

Since 2025 the pattern for AI developers runs the other way:

Testing, voluntarily. Executive Order 14409 (June 2, 2026) asks for a voluntary framework giving government access to covered models “for a period of up to 30 days” before release, and says it authorizes no mandatory licensing.^[11]
Export controls on named models. On June 12 the government applied export controls to Anthropic’s Claude Fable 5 and Mythos 5, barring any foreign national, “including foreign national Anthropic employees.“ They were lifted June 30.^{[12] [13]}
A procurement designation. The Pentagon named Anthropic a supply-chain risk in March after a dispute over use restrictions; the DC Circuit upheld it 2-1 on September 25.^{[14] [15]}
Talk of ownership. OpenAI and the White House discussed an equity donation to seed something like a “Public Wealth Fund.“ Semafor called it “an extremely tough sell for most of the industry beyond OpenAI.“^{[16] [17]}
A voluntary accord and a task force. The September 29 accord asks each developer for “four layers of controls and audits,“ including “an independent external auditor or evaluator” and “an independent committee of the board of directors,“ and says “it may make sense to codify these steps into laws or regulations.“^[1] The Super Intelligence Force, chaired by the Director of National Intelligence, was announced October 4.^[18] As reported, no legal basis, rulemaking power or enforcement power has been published; one official named export controls as a possible enforcement tool.^{[18] [3]}
Courts. Eight states sued OpenAI under consumer-protection law on October 14, and the EU’s AI Office has sent information requests to OpenAI and Anthropic after agent incidents.^{[19] [20]}

My reading: access is limited by nationality, set case by case, and tied neither to classification nor to a priced proprietary channel. That is the opposite of the user-facility rules. The accord and the task force coordinate voluntary commitments. Courts apply existing law one case at a time. Neither is a replaceable-operator contract.

Meanwhile the DOE labs have become hosts for the same companies. Argonne’s Solstice, with NVIDIA and Oracle, is 100,000 Blackwell GPUs meant to “develop and train new frontier models and reasoning models for open science.“^[21] Of the first 278 Genesis projects, 19 are led by companies.^[22] The companies pay in with tools and compute credits, which looks closer to a research partnership (a CRADA) than to an operator contract.^{[4] [23]}

Where this probably goes

I put an estimate on nine paths for what could happen by December 31, 2030. They are not additive, and they are judgment calls from the public record. The October 22 column is the first real measurement. The three earlier columns are simulated: on October 24 I reconstructed what the record then available would have supported at each quarter-end. Nobody made those estimates at the time, and I knew how things turned out, so treat them as a way to see the shape of the story, not a track record.

Path	Mar 31 (simulated)	Jun 30 (simulated)	Sep 30 (simulated)	Oct 22 (measured)
Ad hoc levers continue	0.80	0.85	0.83	0.85
Conditions on chips and contracts	0.78	0.80	0.80	0.80
Labs as users of national facilities	0.55	0.53	0.60	0.60
National-champion bargain	0.22	0.30	0.30	0.30
A public lab spins out a company	0.21	0.22	0.20	0.25
AI national lab under a management contract	0.18	0.19	0.20	0.20
Regulated circuit (licensing)	0.15	0.14	0.15	0.15
Contestable champion program	0.12	0.12	0.12	0.12
International lab	0.05	0.05	0.06	0.05

Estimates are the chance each path happens by December 31, 2030; they are independent, so they do not add to 1. Hollow dots on the tracker are the simulated points.

What the history shows:

The ranking never changed. Ad hoc levers and chip conditions led throughout; an AI national lab under a management contract stayed near 0.20.
The national-champion bargain moved most, from 0.22 to 0.30, on the June reports of equity talks.^{[16] [17]}
Labs as users of national facilities dipped in June, when export controls barred foreign nationals from two models, then rose to 0.60 with the first Genesis awards.^{[12] [22]}

The top two paths need no new institutions. Conditions on chips are already running: Commerce approved G42 and HUMAIN purchases in November 2025 “conditioned on both companies meeting rigorous security and reporting requirements.“^[24] The third borrows national-lab machinery, because DOE already allocates time on light sources and supercomputers by merit review and offers full-cost proprietary access.^{[9] [25]}

The actual national lab, a government-owned secure AI facility under a replaceable operator, sits at 0.20. I found no solicitation or legislation for one, and DOE’s own partnering guide says “the laboratory cannot compete with industry.“^[23]

Not yet re-scored: the newest news. My first read is that the task force and the rehearing fight nudge ad hoc levers up slightly, because both work firm by firm, and leave the regulated circuit roughly flat, because the accord floats codification while the Justice Department litigates against state AI rules.^{[1] [3] [2] [26]} Those are candidate moves, not applied. The tracker will be re-scored, with the reasons shown, when a dated event moves a path.

Allies have built pieces of the toolkit. The UK’s access rests on memoranda that are “voluntary, not legally binding.“^[27] The EU’s General-Purpose AI Code of Practice lets signatories “show they comply with the AI Act by adhering to the code.“^[28] Korea dropped two of five sovereign-model teams in January 2026.^[29]

The strongest objection

Open science and frontier secrecy may not mix. User facilities are open regardless of nationality; a model with dangerous capabilities would need classified handling or new statutory authority, which takes it out of that model.^{[9] [10]} The national labs' own record is also mixed: a 2015 commission found “a strikingly consistent pattern of criticism” across more than 50 reports without comprehensive change.^[30]

Two newer objections cut the other way. Without a contract, accountability arrives through lawsuits, which are slow, backward-looking and unpredictable; a jury ended Elon Musk’s challenge to OpenAI’s structure on timing alone.^[31] And coordination between labs without a statute carries legal risk: a September 18 antitrust complaint treats a public call for “industry-wide coordination” on pace as an agreement among competitors (known only from a search snippet).^{[32] [33]} The accord has signatories “meet regularly to establish standards,“ so that risk is live.^[1]

One press account (reported, single source) also says the White House asked OpenAI and Anthropic to hold new models back from UK testers until US officials had checked them.^[34] If accurate, voluntary testing can compete as well as cooperate.

My reading: the sharpest challenge is that national-security pressure can also be pressure on a company’s own limits. The Anthropic designation turned on its use restrictions.^[35]

What to watch

November 13: the government’s response to Anthropic’s rehearing petition.^[2]
About November 28: proposed legislative language on “SI,“ the executive branch’s new term for AI, due 60 days after September 29.^[36]
The task force’s risk report and breach-reporting protocols inside its 120-day window; a press source puts the end at February 12, 2027.^{[18] [3]}
An INCITE-style call that allocates frontier-scale AI compute, and whether “without regard to nationality” survives for AI compute.^{[9] [25]}
Any signed government stake, and whether it carries governance rights.^{[16] [17]}

Notes and limits

Court opinions and the June 12 Commerce letter were read through secondary reports.^{[12] [35] [15]} One Reuters item on the Pentagon designation was seen only as a search snippet.^[14]
The task force charter was not read directly. Its details come from press and analysis, and its legal basis has not been published.^{[18] [3] [37]}
Model names and incident details are as stated by companies and press.
The UK-US access dispute rests on one press account.^[34]
Probabilities are draft estimates by the research team, not output from a model or market. The March, June and September columns are simulated retrospective estimates made with hindsight; only October 22 is contemporaneous. Estimates for the national-champion bargain and for labs as users of national facilities are the least certain, because a single announcement could settle either.
This issue keeps about a tenth of the research. Left out: the history of Berkeley and Argonne, the dated timeline, five of the nine path write-ups, the table of who proposes what, and most of the legal front (copyright suits in the US, Germany, India, the UK and Canada, and the Colorado preemption fight).

Keep following

The tracker, with all nine paths, what would move each, and every re-score: xanderpico.com/resources/frontier-labs-paths/
The numbers as quoted in this issue (October 22): xanderpico.com/resources/frontier-labs-paths/snapshots/2026-10-22.html
The thread for this project: xanderpico.com/threads/frontier-labs-national-labs.html

Sources

Written with AI assistance. Estimates are draft estimates by the research team, not forecasts from a model or market. Quotes were checked against the pages cited unless marked as seen only in a search snippet. Research and commentary, not investment advice.