Three companies that spent 2026 trying to bury each other on price, benchmarks, and talent have been meeting since July to agree on one thing: who should write the rules for frontier AI. The answer they are circling is themselves.
The Information reported on September 13 that Anthropic, OpenAI, and Google DeepMind have held regular working-group meetings since July to design an industry-led AI standards body, one that would set shared protocols for technical testing and pre-release auditing of the models each of them builds. PYMNTS confirmed the outline: a self-regulatory organization, run by the labs, defining what counts as a dangerous capability and how it gets tested before release. The framing everyone reached for was safety. That framing is doing a lot of work.
The antitrust exemption is the tell
You do not need special legal permission to publish safety research. You do not need it to open-source an evaluation harness, share a red-team methodology, or agree on how to name a benchmark. Labs already do all of that in public.
What you do need permission for is coordinating on the pace of development and the bar a model must clear before it ships. That is the thing US antitrust law is built to stop competitors from doing together, and it is exactly what Dario Amodei asked for. His September 12 essay “We Must Pace the Frontier” proposed cooperation among leading developers on safety practices, and, per Reuters reporting summarized across the coverage, floated targeted antitrust exemptions to make that cooperation legal.
Read the request literally. The three firms that already sit at the capability frontier want a legal carve-out to jointly decide how fast the frontier moves and what a model has to pass to be sold. Strip the vocabulary of risk off the top and it is a standards cartel with a compliance gate attached. The gate is the point. Whoever writes the pre-release audit bar decides which models are allowed into the market, and a bar written by incumbents tends to describe the incumbents.
Three visions that land in the same place
The public disagreement among the three is real, and it is instructive because all three roads end at the same toll booth.
Amodei wants something close to an FAA for AI: a body with the authority to block a model’s release before it ever ships. Demis Hassabis has proposed a FINRA-style organization, industry-funded, starting with voluntary pre-release submissions for independent testing of “dangerous capabilities involving areas such as cybersecurity, biological risks and deceptive behavior,” with a board that includes “industry, government and open-source voices,” and an explicit path from voluntary to mandatory market-access requirement. Sam Altman has pushed an IAEA model: a US-led international forum that certifies countries, companies, and standards, and he has said the labs should build it themselves without waiting on Washington.
An agency that can veto releases. An industry body whose voluntary reviews harden into a condition of selling. An international forum that certifies who is allowed to play. The powers differ on paper. The practical output is identical: a checkpoint, defined by the three largest vendors, that every other model must pass before it reaches a buyer. The argument between FAA, FINRA, and IAEA is an argument about the shape of the gate, not about whether there should be one or who should hold the key.
Gomez said the quiet part
Aidan Gomez, Cohere’s co-founder and CEO, was blunt. He mocked the plan on X as “some great ideas from a cartel” and published an essay putting the objection in plain terms: “should a handful of select, market-dominant AI companies from Silicon Valley get to define the rules and safety standards of a generational technology for the entire world? All while simultaneously determining how fast this technology progresses?”
His structural point is the one worth keeping. If a few large labs set the definition of a dangerous system, the evaluation thresholds, and the release pace, those choices will reflect the capabilities those labs already have. A model built differently, trained cheaper, or shipped with open weights does not get to argue the rulebook. It just fails to clear a bar written by the companies it is trying to compete with. An antitrust waiver, Gomez argues, would entrench that.
This is not a fringe complaint. Andrew Ng made the same warning from the Ai4 stage last month, describing the danger as extinction fear merging with burdensome licensing, a combination he called a bad idea. The people raising it are not the ones who dismiss AI risk. They are the ones who notice that “we must be careful” and “we should decide who is allowed to ship” are two different sentences that keep getting spoken as one.
The two governments in the room
The self-regulation pitch has a convenient backdrop: neither government the labs care about is going to hand them a clean framework soon.
Beijing rejected the premise outright. After Amodei’s essay warned that a Chinese lead in AI would pose grave danger and called for continued restrictions on advanced chips, Foreign Ministry spokesman Guo Jiakun answered on September 14: “Fearmongering, confrontation and vicious competition will only disrupt the process of global AI governance and serve the interests of no one.” State media called the essay a “Cold War playbook” for AI. An international certifying body led from Washington is a non-starter for the second-largest source of frontier and open-weight models on the planet, which means any “global” standard the three labs write is regional from birth.
Washington is not aligned either. Altman’s own position is that the labs should build the body without US government backing, and reporting points to a possible executive order creating a self-regulatory organization rather than a real regulator. Into that gap stepped Microsoft. Satya Nadella called for decentralized governance on September 14, said AI “cannot be controlled by a few entities, but must have broad representation across ecosystems, countries, and sectors,” and moved to publish Microsoft’s first MAI model Code of Conduct for public review. Microsoft is the largest AI infrastructure company backing the opposite of a three-lab club, which tells you the fight is about market structure, not only ethics.
What a rulebook written by your vendors means for buyers
Here is the operator lens, from twenty years of watching who controls the standard end up controlling the roadmap. When a small group of dominant suppliers writes the certification everyone else must meet, the certification stops being a neutral safety floor and becomes part of the supply chain. For anyone deploying AI in production, three things follow.
First, the set of models you can legally and safely put in front of users would be pre-filtered by a board your primary vendors sit on. That is a dependency, not a reassurance. Treat any body with the power to gate model releases the way you would treat a sole-source component: something to be second-sourced around, not trusted by default.
Second, the open-weight and sovereign options that make your lock-in defense possible are the ones most likely to sit outside an incumbent-approved list. If your second-model strategy depends on Mistral, a Cohere deployment, or a self-hosted open model, a standards regime that quietly narrows “approved” to “what the three labs already ship” narrows your negotiating position at the same time. That is the moment vendor concentration stops being an abstract risk and starts showing up in your renewal pricing.
Third, watch the mechanism, not the mission statement. A standards body that publishes its methods, seats independent and open-source members with real votes, and lets outside developers contest thresholds is a genuine safety floor. One that keeps the rulebook, the board, and the pace behind a legal exemption from the companies being audited is a moat wearing a lab coat. The Hassabis proposal at least names open-source voices on the board; the test is whether they get a vote or a viewing gallery.
None of this means the risks the labs describe are invented. Autonomous cyber capability is real, and the same firms have shipped models paused at a critical-cyber threshold this year to prove it. The question Gomez keeps asking is the right one and it is not about whether AI needs guardrails. It is who writes them, who participates, and whose existing capabilities the definition of “safe” happens to match. Right now the answer being negotiated in private is: the three companies with the most to gain from writing it. For a buyer, that is not a safety story. It is a market-structure story, and the smart move is to keep a second model qualified while the rulebook is still being drafted.