AI giants demand a development slowdown to save their business models

By Samuel Abiola · 2026-09-15

The essay and the S-1

On 12 September 2026, Dario Amodei published an essay calling for a measured slowdown in AI capability development. He framed it as an ethical obligation. The plan runs to three steps, each aimed at slowing how fast AI systems gain power.^[1] That same day, Sam Altman told a Fortune interviewer that taking OpenAI public in 2026 would be ill-advised given AI safety concerns.^[2]

Anthropic had already submitted a confidential draft Form S-1 to the U.S. Securities and Exchange Commission on 1 June 2026, positioning itself for a public listing once regulatory review concluded.^[3] AllMind analysis of that filing and associated Polymarket data put the probability of a first-day market capitalisation at or above US$2 trillion at around 60 per cent.^[4] The company preparing for one of the largest technology listings in history was, on the same afternoon, urging the industry to slow down.

Reading the balance sheets

The financial picture at both firms makes the slowdown argument considerably more legible. OpenAI expects to lose US$14 billion in 2026 on projected revenue of roughly US$25 to 30 billion. Cumulative cash burn through 2030 is forecast at US$218 billion. No positive free cash flow is expected before that year.^[5] These are extraordinary figures even by the standards of venture-backed technology, describing a company whose spending on compute infrastructure and staff is outpacing revenue growth by a substantial margin.

Anthropic's position differs in degree. Both companies have raised private capital at valuations that presuppose continued dominance of the frontier. OpenAI's February 2026 round valued it at US$852 billion. Anthropic's May 2026 Series H put it at US$965 billion. That capital funds the compute that drives the capability race. A slower race is a cheaper race, and that logic runs directly through the balance sheets of both firms.

Who benefits from the moat

The obvious concession comes first: Amodei may be entirely sincere. The risks of rapidly escalating AI capability are serious enough to have attracted credible independent researchers well outside any financial orbit. Amodei has argued for slowing the rate at which AI models are made more powerful across several years and in various forums.^[1] Sincerity, though, does not dissolve a conflict of interest; in this case it compounds one.

Regulatory caps on frontier development, the natural policy outcome of a successful pacing argument, would entrench the incumbents best placed to absorb compliance costs. A smaller competitor burning through seed funding to close the capability gap would face those same obligations on a fraction of the capital base. A well-designed pacing regime, whatever its intent, distributes costs regressively across the industry: proportionally heavier on challengers, proportionally lighter on firms already at scale.

Altman's remarks about 2026 being an ill-advised moment for an OpenAI IPO arrived on precisely the day Amodei published his call for a slower frontier.^[2] Whether the timing was coordinated or coincidental, two CEOs of the world's two most valuable private AI companies were publicly pumping the brakes on the same afternoon. Safety language, on a day like that, does a great deal of work that financial language cannot do publicly.

The structural question

Bad faith is the wrong charge to lay here. The point is narrower. Commercial interest and stated ethics have converged, and that convergence deserves scrutiny rather than applause. Policymakers evaluating calls for AI governance should weigh the expertise of the people in the room against their structural interest in the outcome of any regulatory decision.

The AI governance debate will be shaped substantially by the firms closest to the technology, and that is probably unavoidable. What is avoidable is the assumption that safety advocacy from incumbents is purely about safety. FutureSearch projects OpenAI will not reach positive free cash flow before 2030,^[5] and every quarter of slower compute scaling between now and then reduces the gap between projected burn and actual spend.

Samuel Abiola

Samuel Abiola writes about inequality and social policy. He works close to the research and is interested in what the numbers mean for the people counted in them.