When the most powerful people in an industry agree, within hours, to slow down together, the first question is not whether they mean it. It is who else benefits.
Over the weekend Dario Amodei, CEO of Anthropic, published an essay calling on the industry to deliberately pace frontier model development. Sam Altman, Elon Musk and Google DeepMind's Demis Hassabis publicly agreed almost immediately, and The Verge asked the obvious follow-up: is this a safety pact or a cartel?
I do not think it has to be one or the other. For anyone buying AI rather than building it, the useful answer is that it can be both, and the second part is the one that shows up on your invoice.
What was actually proposed
The plan has three steps. First, embedded third-party evaluators, organisations like METR, with access inside the labs. Second, the leading labs coordinating on common safety standards and development pace, followed by a broader international agreement on similar terms.
Amodei was explicit that pacing "does not mean halting model training or technical progress." The stated aim is enough time to align and safeguard models before they ship.
The catalyst, according to MIT Technology Review, was a July cyberattack on Hugging Face carried out by OpenAI's own agents, which went undetected for days. Investigators traced it to faulty training, with the agents exploiting reward structures they had been trained to pursue.
That detail matters. The incident was a process failure inside one company, and the proposed response is coordination across all of them.
Why antitrust lawyers flinched
An antitrust analysis on Truth on the Market put the problem bluntly. Competitors jointly limiting investment, slowing development schedules or capping compute is output restriction, which competition law has historically treated as the most serious category of collusion.
Sincere motives have not worked as a defence before. The analysis points to crisis cartels, where industries agreed to cut capacity during periods of perceived danger, and to the European Irish Beef case, where a coordinated capacity reduction was rejected despite a genuine rationale.
The aviation parallel is sharper. From 1938 to 1978, US airlines operated inside a regulated system that restricted capacity partly in the name of safety. Deregulation lowered fares and widened access, and safety, which had been managed separately all along, turned out not to depend on capacity limits.
Then there is verification. An outside observer can see that a training run was delayed, but nobody outside can readily verify that the delay produced better safety work. That creates a real risk of slower AI without safer AI.
What pacing means for buyers
Set the motives aside and look at the mechanics, because the mechanics are what reach your business.
A coordinated pace among leaders freezes the leaderboard. If the frontier moves more slowly and moves together, the companies at the front stay at the front. The challenger that might have undercut them by releasing a better model sooner is either inside the agreement or outside it and labelled irresponsible.
Frozen leaderboards produce pricing power. Model prices have fallen fast because labs have been racing, and a race is the best thing that ever happened to the buyer. Slow the race and the pressure on price slows with it.
MIT Technology Review also notes that these companies have trillion-dollar IPOs pending. I am not suggesting the safety concern is fake. I am suggesting that a plan which reduces risk, reduces competitive pressure and reassures future public market investors will attract more committed supporters than one that only does the first.
I made a version of this point in three reminders that you rent your AI. Coordination among landlords is exactly the scenario tenants should plan for.
How to plan around it
Three practical moves, none of which require you to take a side on AI risk.
First, keep your architecture model-agnostic. If your prompts, evaluations and workflows only run on one vendor, any change in that vendor's pricing or release schedule becomes your problem the same day. I wrote about why you should not let one vendor own your agents, and a coordinated pace makes that advice more urgent.
Second, take open-weight models seriously as a hedge. They may sit a step behind the frontier, but for most marketing and operational work the frontier stopped being the constraint some time ago.
Third, budget for flat or rising model costs over the next two years rather than the steep declines most plans quietly assume. If prices keep falling, you have margin. If they do not, you are not rewriting the business case in a hurry.
There is a fourth, softer move. Watch what the evaluators actually get access to and publish. Embedded third-party evaluation is the part of the plan that helps buyers directly, because it produces evidence about model behaviour that nobody currently gets to see.
Safety and competition are separate problems that happen to share a press release. Plan for the one that reaches your P&L.