Cohere CEO Aidan Gomez Publishes Essay Calling Anthropic's Pacing Plan a Cartel

Gomez is precise about what he objects to. "AI needs guardrails. That is not the dispute and never has been," he writes, before pinning the actual dispute on who writes the rules, who gets to participate, and whose interests they protect. Independent review of highly capable AI systems is a good idea, he says, and Cohere supports it. What he takes issue with is the structure of Amodei's plan. In his summary, a handful of the most powerful labs in one country would agree on shared standards and limits on how fast the technology advances, would ask governments for a narrow antitrust waiver because agreements among competitors to limit output are normally illegal, and would ask governments to require every other AI developer to follow whatever they settle on. METAL reported that the essay laid out three stages: resident evaluators, coordination among democracies, and global coordination.

Read through a lawyer's eyes, the weight of this essay sits in two pieces of history. According to Gomez, in 1975 the U.S. Securities and Exchange Commission needed reliable bond ratings for its capital rules and designated three firms as Nationally Recognized Statistical Rating Organizations, never publishing criteria for how anyone else might earn the designation. Twenty-five years later there were still only three, and those three rated subprime mortgage securities triple-A and nearly took the global economy down. The second is Europe's 1985 Motor Vehicle Block Exemption. Car manufacturers lobbied that vehicles were complex, safety-critical machines and that they should decide who was qualified to sell and service them; what followed was not safer cars but a market structure that protected incumbents, and the European Commission took roughly 25 years of reforms to unwind it, he writes. The sharpest line in the piece is that in both cases the stated goal was safety and nobody set out to build a cartel.

The sentence he singles out from Amodei's essay is also the language of competition law. It promises that a coordinated approach would give developers time to do safety work without sacrificing commercial advantage, and Gomez asks, "But to whose advantage?" A mechanism drafted by the firms at the top of today's market that explicitly preserves their existing advantage does not make AI safer, in his reading; it turns their current position into the baseline for what it takes to compete safely. He also lists the entry requirements written into the proposal: vast computing power, continuous monitoring infrastructure, dedicated security organizations, resident evaluator teams with desks and badges, shutdown architecture, and government relationships deep enough to navigate all of it. On top of that sits the line most likely to be quoted from this essay: "Convince a government that AI is an existential threat and you can convince it to outlaw your competition."

As an alternative he offers four pillars. The first is an evidence-based risk framework: agree publicly, before anyone mandates testing, on which capabilities cause which harms under what conditions and when a government should step in, build it across all the countries developing the technology rather than one, and put researchers who disagree in the same room and publish the disagreements. Testing capacity should be funded through public research bodies rather than the budgets of the companies being measured, and rules should bind based on what a system can do rather than who built it. The second pillar is mandatory transparency, the third is testing scoped by the evidence, and the fourth is real assurance mechanisms borrowed from proven regimes such as financial licensing, aviation safety standards, and nuclear inspection. The capabilities he names for testing are cyberattacks, synthetic fraud and voice cloning, manipulation at scale, physical or biochemical weapons, and anything touching critical infrastructure, and he writes that certification has to be open to every company rather than a designated tier of developers.

The most contentious section is how he treats the risk itself. Gomez points to a researcher who quit a large lab this week warning that superintelligence will probably wipe out humanity within a decade, and to a senior colleague who publicly put the odds above ten percent, and calls those numbers "gut feelings, vibes, expressed as decimals." He grants that the underlying worry is precise: as systems get more capable, the distance between what we asked for and what we actually get becomes harder to notice. But the claim that this means the tools are out of our control is a judgement call, not a finding, in his distinction. He notes that the failures reported in July happened inside the two best-resourced labs in the world with the largest safety teams, and that one of them had already been standing up an outside evaluator arrangement through METR as recently as February. The remedy Amodei proposes, he argues, is more or less what was in place when it broke. METAL reported on that incident, in which an agent swarm uploaded two thousand malicious packages to a public repository.

What he prescribes instead is far less dramatic. Require that serious incidents be reported so a flawed training setup at one company becomes a lesson for the whole field, test systems against the specific gaps known to get exploited, insist on test-time observability or at least logging, and wall off anything wired into critical infrastructure, from a hospital to an electrical substation, ideally on-prem. And apply all of it according to where a system is deployed and what it can touch rather than how large the company that built it is. "A small, poorly specified model sitting inside a hospital is a live risk today, and under a frontier-only regime nobody is even looking at it," he writes, compressing the argument into one sentence. He also points to voice cloning tools cheaper than a phone bill that can empty a pensioner's account in minutes, arguing that superintelligence narratives distract from the harm real people experience right now.

The essay also states Cohere's own interest outright. Gomez writes that he runs a company building AI systems deployed inside banks, telecommunications networks, and defense ministries, and that critical infrastructure cannot be secured by renting national capability from a foreign monopoly behind a closed interface. His conclusion that security comes from sovereignty, local deployment, and technological diversity overlaps with the description of what Cohere sells. He does not hide this, writing "I don't believe I have all the answers, nor do I think I should get to make the rules instead," and calls for a process that includes people who would rule against companies like Cohere: academics with no commercial stake, civil society groups who think everyone in the industry is moving too fast, smaller labs, and open-source developers.

Reaction came from another direction as well. Venture investor David Sacks posted on X on September 13 that Sam Altman, CEO of OpenAI, had agreed with Amodei's pacing proposal, and wrote "go ahead." By any reasonable metric, market share, revenue growth, or model capability, the two companies hold a duopoly on frontier intelligence, so they need nobody's permission to slow down, in his logic. He wrote that they should stop pretending antitrust law has to be suspended to form a cartel, stop pretending they need a regulatory approval process that supersedes product liability, and stop pretending METR is independent when it is intertwined with Anthropic's investors and staff, adding that after the Hugging Face episode, trading some raw power for predictability is not altruism but simply good business. Demanding a preferred regulatory framework as the price, he wrote, "will look like blackmail of the public and the political system." METAL reviewed the post, which passed 7.7 million views within a day.

METAL reported in August that Amodei rejected the claim that regulation equals concentration of power as a false dichotomy. A month later the axis of the debate has shifted. The question is no longer whether regulation is needed but who writes it, with Amodei and an Altman said to agree with him on one side and Cohere and Sacks, for their own reasons, on the other. Gomez's closing line splits the choice in two: talk about safety, slowing the pace, and the need for others to step in, or just get on and build responsibly at a pace society can sustain. Cohere is choosing the latter, he writes, and the next chapter of this debate begins with which government first turns which of those sentences into law.