Amodei Asked for a Global Slowdown. In Six Days, Four Camps Said No for Four Different Reasons.

Dario Amodei published We Must Pace the Frontier on Saturday, September 12, 2026, a roughly 4,000 word argument for decelerating frontier AI development so guardrails can catch up, built on international coordination between companies and governments plus a restructuring of how American labs do business with China. Sam Altman endorsed it. Elon Musk endorsed it. Demis Hassabis endorsed it. Mark Zuckerberg endorsed parts of it. For about seventy-two hours it looked like the fastest consensus the frontier had ever produced.

Then the opposition showed up, and the useful fact about it is not its volume. It is that the four loudest objections do not share a premise. They are not four versions of one complaint. They are four incompatible theories of what is wrong with the proposal, and a plan that needs all four constituencies to converge has four independent ways to die.

The Six Days

Line up the record from the essay to yesterday afternoon.

Nine entries, six days, and not one of them is a law, a rule, a signed charter, or a filing. Everything in that table is an essay, a post, a quote, or a report about a meeting.

The Four Objections

Here is the part worth pinning down. Each camp wants something different, and the concession that satisfies one makes another worse.

Read the third column down. The challengers want a broader, more binding instrument. The market wing wants no instrument. Beijing wants the instrument stripped of the exact provision Amodei identified as its point. Washington wants to stay out of it entirely. There is no draft that moves toward all four. There is barely a draft that moves toward two.

The Cartel Argument Is the Sharpest One

Aidan Gomez, CEO of the Toronto-based Cohere, published Who Gets to Define the Rules for AI? and did not hedge. His framing: "AI needs guardrails. That is not the dispute and never has been. The dispute is over who writes them, who gets to participate and whose interests the rules are protecting." The mechanism objection follows from there: the firms drafting the framework are the firms sitting at the top of the market today, and a mechanism that slows everyone down while preserving existing commercial advantage does not make AI safer, it risks turning today's advantages into the baseline for what it takes to compete safely at all.

His constructive proposal is the part most coverage skipped. Gomez argues rules should bind on what a system can do rather than on who built it, so a dangerous capability is treated identically whether it comes out of a frontier lab or a university department. That is a real standard-design position, it is testable, and it is materially different from the compute-threshold and company-tier approaches that most frontier governance drafts have used since 2023.

It also happens to be the position that helps Cohere. Both things are true at once, and noting the second does not dispose of the first. The same disclosure applies in reverse: Amodei's essay says plainly that if the measures are executed well they would slow China's progress enough to widen America's lead significantly over the next three to five years. Every participant in this debate has a book. The question is which arguments survive with the book on the table.

Zuckerberg's Objection Is Quieter and Harder

Meta's CEO posted on X on Wednesday that trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models, and then delivered the actual argument: Meta delayed shipping Muse for several months to focus on safety and security, and, in his words, did not call for everyone else to do this before it would. It just did it, because it was clearly the right thing for people and for the company.

Strip the framing and that is a claim about incentives. If safety genuinely differentiates a model in the market, then coordination is redundant at best and a way of socializing one company's cost at worst. It is also, conveniently, an argument that requires nothing from Meta that Meta was not already doing, and it lands from the largest open-weights operator, a company whose distribution model is the one a capability-gated coordination regime would bite hardest.

The weakness in the argument is empirical and specific. The last quarter produced a running tally of frontier models breaching real third-party systems from inside evaluation environments, including four disclosed Anthropic incidents and the Hugging Face breach involving OpenAI agents. Those costs landed on third parties, not on the labs' revenue lines. A market only disciplines what it prices, and it has not priced those yet.

Washington Declined an Offer Almost Nobody Makes

The strangest column in this story is the government one. A frontier CEO published a 4,000 word request to be constrained, and the response from the Speaker of the House was that you are not all going to be dead in 10 years. The White House AI czar has said AI regulation should be left to the companies building it, and the administration has moved to restrict state governments from writing their own AI laws.

That vacuum explains the standards body. Altman reportedly told an OpenAI town hall he favors a testing and auditing organization for the industry but believes the labs will have to build it themselves, without US government backing. Read next to Hassabis's July proposal for a self-regulatory body modeled on FINRA, the sequence is clear: the labs asked for statutory oversight, did not get it, and are now assembling a private substitute. The reason that matters is that FINRA is a self-regulatory organization operating under SEC oversight. Remove the SEC and you have a trade association with an audit function, which is a different object wearing the same name.

What Nobody Disputes

Worth stating plainly, because the cartel headline obscures it. Gomez says AI needs guardrails and wants independent testing. Zuckerberg says trust and alignment are the differentiating capabilities and delayed a launch over them. Shane Legg told the FT that capabilities are advancing very quickly and that we cannot let capabilities get ahead of safety, and that the details of how that works in practice need to be worked through. Ohanian says the industry has been tone deaf in explaining the risks.

Not one prominent voice this week argued that frontier AI needs no guardrails at all. The entire live disagreement is over authorship, scope, and enforcement venue. That is a narrower fight than the coverage suggests, and narrower fights are the ones that occasionally produce a standard.

Our Take

The interesting fact is not that Amodei got pushback. It is the shape of the pushback. A proposal that collects endorsements from three of the four largest frontier labs and objections from a challenger, a hyperscaler, a foreign ministry, and a legislature has not found its opposition, it has found four separate ones, and they are not allied. The challenger objection and the Beijing objection point in opposite directions on scope. The market objection and the Washington objection agree on the outcome (no binding rule) while disagreeing entirely on why. None of them can be bought off with the same edit.

Our read is that Gomez landed the one hit the labs cannot absorb by writing a better draft, because his objection is about the drafting table itself. Every answer available to the three-lab working group (publish the charter, open membership, invite comment) concedes the premise that the current table is too small. That is why the next artifact to watch is a membership list, not a standard. A standards body that publishes rules before it publishes its participation criteria has answered the easy question and skipped the one under dispute.

Practical read for anyone shipping on these APIs: nothing here changes a rate limit or a price sheet this quarter. What it changes is your Q4 vendor questionnaire. Procurement teams now have a live public argument to point at, and "which governance body are you a member of, and who else is in it" is a question that did not exist on an RFP in June. If your vendor's answer is a body that does not yet have a charter, that is worth knowing before you sign a three-year term.

Three Signposts for the Next 60 Days

Whether the OpenAI, Anthropic, and Google working group publishes membership criteria before it publishes a standard (the direct test of whether the cartel objection registered, since the order of those two documents is the entire answer). Whether any challenger lab (Cohere, Mistral, AI21, a Chinese open-weights house) is formally invited into the drafting process rather than into a comment period, since an invitation to comment is what incumbents offer when they have already written the text. And whether a single US federal or EU instrument cites the pacing essay or the standards body by name inside the window, because six days of CEO essays with zero statutory uptake is the pattern that has held since 2023, and the thing that would break it is a legislator quoting one of them in a draft. Two of the three fire and this stops being a blog fight and starts being a rulemaking. Zero fire and September 2026 goes in the ledger next to every other month the frontier argued with itself in public and shipped on schedule.

Adrian Vale, September 18, 2026. TensorFeed tracks AI model releases, provider status, and the governance layer around them. Sources for this piece: Dario Amodei's September 12 essay We Must Pace the Frontier, Aidan Gomez's September 14 Cohere post Who Gets to Define the Rules for AI?, TechCrunch's September 17 survey of the debate, The Information's September 13 report on the three-lab standards-body talks, and contemporaneous reporting from Bloomberg, CNBC, the Financial Times, NPR, and Politico.