TL;DR: Anthropic, OpenAI, and Google DeepMind have been meeting in working groups since July to create an industry-led AI safety standards body. Anthropic backs a federal-agency model (FAA), Google DeepMind a federally overseen industry body (FINRA), OpenAI an international certification forum (IAEA). They agree on the diagnosis -- frontier models need pre-release scrutiny -- and disagree on who holds the pen. On September 14, Trump called Jensen Huang live onstage at the All-In Summit; Huang said "we're not going to let it happen." The antitrust waiver Amodei's plan required is not coming from this White House. Three vendor contract questions for enterprise teams.
The week that started with Dario Amodei's "We Must Pace the Frontier" essay ended with a phone call.
While Nvidia CEO Jensen Huang was onstage at the All-In Summit in Los Angeles on September 14, he received a live call from President Trump. Huang put it on speaker. Trump's message was unambiguous: the AI safety push is politically motivated, possibly a Chinese scheme, and the United States should be racing, not braking. Huang did not push back. He said "we're not going to let it happen."
That exchange happened 48 hours after Amodei's essay called explicitly for the US government to mediate industry coordination and issue "a narrow antitrust waiver for certain kinds of safety conversations." Those two data points are in direct tension. The antitrust cover Anthropic's plan requires is not coming from the current White House.
Meanwhile, The Information reported on September 13 that Anthropic, OpenAI, and Google DeepMind have already been meeting in working groups since July to build the thing Amodei described -- an industry-led AI safety standards body. The meeting is happening. The structure is not settled. And the government involvement that would give it teeth is now explicitly opposed by the administration.
The three models on the table
The working group has converged on a shared diagnosis and diverged on the prescription. All three CEOs agree that frontier models should face outside scrutiny before they are released to the public. That convergence is meaningful -- it is the first time the three leading frontier labs have publicly aligned on any structural change to how AI is tested before deployment.
The divergence is on structure, and it matters for enterprise buyers because the structure determines what "my vendor is a member of the safety body" actually means.
The FAA model (Anthropic / Amodei): A federal agency modeled on the Federal Aviation Administration, with the authority to block a model's public release from Day 1. Under this model, a frontier AI model would be required to pass a pre-release safety evaluation conducted by an independent government-empowered body before any public or commercial deployment. Amodei has connected this to the embedded evaluator commitment he announced separately -- evaluators with badge-level access to the lab would be the instrument of verification.
This is the most government-involved of the three proposals. It is also the furthest from the current US policy posture under the Trump administration, which has moved to reduce, not increase, federal AI oversight.
The FINRA model (Google DeepMind / Hassabis): An industry-funded, federally overseen standards body modeled on the Financial Industry Regulatory Authority, which polices Wall Street under Securities and Exchange Commission oversight. Under this model, the body starts with voluntary pre-release reviews and builds toward mandatory market-access rules over time. Demis Hassabis proposed this structure in a July essay, positioning it as more politically viable than the FAA model because it starts voluntary.
The key enterprise implication of the FINRA model: early membership is not meaningful. FINRA-style bodies begin as industry clubs whose standards carry reputational weight but no enforcement mechanism. The enforcement layer arrives later, potentially years later, after the body establishes credibility and gains regulatory recognition. An enterprise team that asks a vendor "are you a member of the AI safety body?" today would get a yes that means almost nothing under the FINRA model.
The IAEA model (OpenAI / Altman): A U.S.-led international forum modeled on the International Atomic Energy Agency, which certifies countries and organizations against agreed safety standards. Altman laid this out in a Financial Times essay, framing it as more appropriate for AI than for financial regulation because the technology is inherently global -- standards need to apply across national borders, not just US-domiciled labs.
The IAEA model is the most ambitious geopolitically and the most practically complex. It assumes international coordination that the September 14 news cycle already complicates: on the same day as the Huang-Trump call, Bloomberg reported that China has rejected the slowdown push entirely. A standards body that requires Chinese participation to be globally meaningful is a standards body that currently does not have Chinese participation.
The Trump-Huang moment and what it blocks
The All-In Summit call was not a policy announcement. It was a public alignment signal between the president and the CEO of the company that makes the chips every frontier AI model runs on.
What it closes off, specifically, is the antitrust path Amodei's plan required. The Amodei pacing plan published September 12 was explicit: for the three labs to coordinate on safety standards and development pace, "the US government needs to issue a narrow waiver for certain kinds of safety conversations." Without that waiver, coordination between Anthropic, OpenAI, and Google DeepMind on development limits looks like competitors agreeing to restrict output -- the kind of conduct that triggers DOJ or FTC scrutiny.
The Trump administration is not going to issue that waiver. The president who called AI alarms "a HOAX" and told Huang's audience that "data centers are the oil of the next 20, 25 years" is not going to facilitate a legal framework for AI labs to coordinate on slowing down.
That leaves the safety body working group with two viable paths. The first is the FINRA model: form a voluntary industry body without antitrust waiver, agree on standards, and let those standards carry market rather than legal weight. This is the most politically viable path but the one with the weakest enforcement mechanism. The second is the legislative path: if Congress passes the Senate duty-of-care bill that surfaced the same week, it would create the legal mandate that makes coordination involuntary and therefore outside antitrust scope.
The Senate duty-of-care draft that appeared in the same news cycle would put a statutory obligation on frontier AI builders to take reasonable care to prevent harm. That framing -- "duty of care," borrowed from tort law -- is different from the FAA blocking model and different from the FINRA voluntary structure. It would require frontier labs to demonstrate reasonable care, which creates a minimum standard without specifying how it must be met. It does not currently have the votes to pass; its significance for enterprise teams is that it signals what congressional Democrats are building toward.
3 vendor contract questions
The working group news and the Huang-Trump call together produce three concrete questions for enterprise teams evaluating AI vendors:
1. Which safety body model does your vendor support, and what commitment have they made to it?
Amodei, Hassabis, and Altman have all publicly stated their preferred models. But preference is not commitment. Anthropic has made an actual commitment -- the embedded evaluator pledge -- that is independent of whether the safety body forms. OpenAI has verbally agreed to embedded evaluators. Neither OpenAI nor Google DeepMind has made a structural commitment equivalent to METR-badge-level access.
Ask your vendor: what specific commitment has your company made to pre-release safety review, independent of which safety body model is ultimately adopted? A commitment that is contingent on a body forming is not a commitment. A commitment that is contingent on government action that has been explicitly opposed is not a commitment.
2. Is your vendor's safety body membership currently independently verifiable?
Under the FINRA model -- the most likely near-term outcome -- early membership is voluntary and reputational. There is no equivalent of a FINRA registration database that an enterprise compliance team can consult to verify whether their AI vendor is current on safety review requirements.
This means that if your vendor says "we are part of the AI safety body" before the body has a formal structure and registration mechanism, there is no way to verify that claim. Your due diligence checklist should note this gap and require: if you claim safety body membership or commitment, provide documentation of what that membership requires and what audit has verified your compliance with it.
3. What happens to the antitrust waiver your vendor's coordination strategy requires?
If you asked your AI vendor last week whether they would participate in industry coordination on development pace and safety standards, the honest answer required a government antitrust waiver to be legally safe. That waiver is now clearly not coming from the Trump administration.
One of two things follows: your vendor either participates in coordination without the waiver (accepting antitrust risk), or the coordination does not happen in the legally meaningful form Amodei described. Ask your vendor directly: what is your legal analysis of participating in safety standard coordination with other frontier labs under current antitrust law? What is the path to coordination that does not require an antitrust waiver the current administration will not provide?
The gap between what the labs say they want to do and what they can legally do without government cover is a material vendor governance question that no frontier lab has publicly answered.
For the full context on Amodei's pacing plan and the embedded evaluator commitment, see the Amodei pacing plan analysis. For the legislative context on the superintelligence ban bills and OpenAI board changes, see the superintelligence ban bills analysis. For structuring vendor contracts around these questions, see the agentic AI vendor contract clauses guide.
Related Reading
- Amodei's AI pacing plan: embedded evaluators and 3 vendor questions
- Superintelligence ban bills and OpenAI board shift: 4 governance questions
- AI agent monitoring gaps: OpenAI 30-min window and Anthropic 50% detection rate
- Anthropic recursive self-improvement: governance policy checklist
- Agentic AI vendor contract clauses: what to add in 2026
- Illinois SB 315: frontier AI safety law and vendor risk checklist
- Board AI governance reporting: quarterly template for 2026
- Trump-Xi AI Summit Sept 24: chip loophole, distillation, 3 vendor questions
