The Economic Incentives Behind the Push to Pause AI

On September 12, Dario Amodei published an essay called We Must Pace the Frontier, which says the AI industry should slow down how fast it improves its models. Sam Altman agreed the same day, and Elon Musk, whose SpaceX owns xAI, answered with "Dario is right." Everyone is now arguing about whether the danger is real. I think you should look at the motives either way. That includes the researchers who quit this month, at least one of whom walked away from unvested stock. I cannot speak for them and I am not going to speculate, beyond saying that even they might have reasons of their own. For the companies there is no need to guess. I see at least four economic incentives, and every one of them points the same way as the safety argument.

The first one is that the top models have become too expensive to sell. A bigger model needs more compute, so it costs more to run and the price goes up. OpenAI has tried hard to bring prices down, but the top model GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens, against $4 and $20 for GPT-5.6 Sol. Claude Fable 5.1 is $10 and $50 as well, twice the price of Opus 5, with a tokenizer that produces about 30% more tokens for the same text.

Developers like me use these models all day, but on subsidized subscriptions. Companies paying full API prices were already under cost pressure even before the current top models existed. Uber gave its engineers Claude Code and burned through its entire annual AI budget in four months. The money was gone by April, before Fable 5 and GPT-6 Astra came out, and in June the company capped spending at $1,500 a month per employee per coding tool. AI may make these companies more productive, but if the productivity costs more than it brings in, they do not buy it, and I think a lot of them already stay away from Fable for that reason. I have the same problem on a smaller scale, paying $800 a month in subscriptions and still pushing everything I can to cheaper models. So there is no one standing in line to pay full price for an even bigger model.

The second incentive is government money. A full halt, where nobody builds better models, will never happen in the US, because China keeps building them and falling behind would be a national security problem. The essay itself says pacing "does not mean halting model training or technical progress," and the next day Amodei called China the "toughest dilemma". So I guess the idea is that those better models still get build, but don't get released to the general public. They will be exclusively used by government institutions who will pay any amount of money for the privilege.

The third incentive is that both companies need to go public. Anthropic filed in June and is expected to list on Nasdaq as soon as next month, possibly at a $2 trillion valuation. OpenAI filed too, and on the Saturday the essay came out, Altman told Fortune that "right now would be an ill-advised moment to go public" and ruled out 2026. Both need the stock market to lock in the valuations their investors paid, and I do not think the venture market can keep writing checks this size. OpenAI's audited 2025 numbers leaked in June and were verified by the Financial Times. It lost $20.9 billion in operations on $13.1 billion of revenue and spent $19.2 billion on research and development, i.e. training new models. The subsidized subscriptions are thriving at both companies. A year in which you do not spend tens of billions training a model is a very good year. Slowing down saves that money, and having the government pay for it saves it too.

The fourth incentive is insurance. The one concrete thing Amodei committed to is outside evaluators. They get desks, badges and laptops, and "access to workspaces, tools, and permissions mostly comparable to what internal risk assessment teams have." Altman called that "a great idea" and said OpenAI will do the same. The labs are saying they cannot fully control what they are building, and the fix is to bring in outsiders who know nothing about those models, let them look around, and have them say it all looks fine. When one of these models causes billions of dollars in damage, then guess who will be blamed.

My bet is that new models keep getting trained, that they are made public more slowly or not at all, that the prices at the top stay high, and that the first customer for new top models will be the US government.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com