OpenAI and Anthropic Want to Slow Down AI — and It Builds Them a Moat

The striking thing about the sudden enthusiasm for slowing down artificial intelligence is not that the people building it now want brakes. It is that the brakes they are describing are ones only they can comfortably afford. When the leaders of the most valuable AI labs agree, almost in unison, that the field is moving too fast, the safety case is real — and so is the business case sitting quietly underneath it.

The trigger was a call from Anthropic’s Dario Amodei to “pace the frontier”: slow the rate at which model capabilities improve, install independent evaluators inside the labs, and get the leading players in democratic countries to agree on common safety standards. He framed the public backlash against AI as, in his words, a crisis of trust. Within a short window Sam Altman said he agreed, and that OpenAI had been discussing exactly this; Elon Musk offered a two-word endorsement, “Dario is right.” The coverage treated it as an overdue conscience arriving all at once.

Read the same plan as an analyst rather than a safety officer and a second picture appears. Embedded third-party reviewers, certified alignment testing, published risk frameworks and incident reporting are not only safeguards; they are fixed costs. And fixed costs are the most reliable competitive weapon a large company owns, because they fall on everyone equally in dollars and on no one equally in pain.

Consider the arithmetic. The largest US cloud and AI infrastructure providers have committed to something on the order of 660 to 690 billion dollars in capital spending for the coming year, with Amazon alone near 200 billion and Alphabet, Meta and Microsoft each in the hundreds of billions or approaching it. Against balance sheets like these, a compliance regime — a team of embedded evaluators, an annual safety framework, a penalty that California caps at a million dollars per violation — is a rounding error. For a startup trying to reach the frontier, it is the entire runway.

This is no longer hypothetical. California’s SB 53 already defines a “large frontier developer” by two gates: models trained above a very high compute threshold, and more than half a billion dollars in revenue. Those firms must publish safety frameworks and report critical incidents on tight clocks. In Europe, the AI Act’s Article 55 requires providers of the most capable general-purpose models to run adversarial testing, mitigate systemic risk and report serious incidents. The floor is being poured; the question is who is tall enough to stand on it without noticing.

The clearest tell is who is not asking for brakes. The plan never names its natural losers, but they are easy to identify: the labs whose entire model is giving the weights away — Meta, France’s Mistral, China’s DeepSeek and Alibaba. As one dissenting analysis put it, these are “the four names the essay never prints,” the makers of the free models a paying customer might otherwise download instead of renting an API. You cannot certify the alignment of a model anyone can modify in an afternoon, so a standards regime aimed at the frontier lands hardest on open weights. And the open camp is winning the very users the incumbents would rather keep: one Chinese open model, Qwen, counts downloads in the billions where Meta’s Llama counts in the hundreds of millions.

None of this is new. It is the oldest move in any regulated industry — the established player welcomes the rulebook it can afford, because a cost that is trivial on its books is a wall on a rival’s. Economists have a plain name for it, regulatory capture, and it does not require bad faith. A safety argument can be entirely sincere and still function as a barrier to entry. Here, both things are true at once.

For Europe the stakes are sharper than they look. The continent’s one credible frontier challenger, Mistral, is precisely an open-weight lab, and its Spanish and wider European customers — the startups building products on cheap, modifiable models — are the ones who lose most if the global compliance floor is written around four American incumbents. A rulebook designed in San Francisco and ratified in Brussels can protect the public and pull the ladder up behind the leaders at the same time.

So welcome the safety turn: the risks are real and the evaluators are overdue. But watch what comes next. If the final rules settle exactly where the four giants can pay and nobody smaller can, the slowdown will not have paced the frontier so much as fenced it.

Tags: OpenAI, Anthropic, AI regulation, Open-Weight Models, Mistral