Sam Altman pitched three endgames for AI safety, report says

OpenAI CEO Sam Altman laid out three possible endgames for AI safety in an internal presentation — an international watchdog modeled on the IAEA, a merger of the major AI labs into a single entity, or one company so dominant it could shut down everyone else's compute, according to a report from The Decoder. Altman himself called that third option "terrible."

The detail that stands out is the third scenario. In it, a single firm becomes so powerful that it can effectively "melt" the GPUs of every other player — a blunt way of saying it could destroy rivals' ability to run advanced AI at all. That Altman even articulated this as a possible path, while explicitly labeling it a bad one, gives a rare look at how the CEO of the company racing hardest toward artificial general intelligence frames the stakes.

What are the three paths Altman described?

According to the report, the internal presentation sketched three routes to a safe AI future:

• An international watchdog modeled after the IAEA, the United Nations nuclear inspectorate, which would monitor and verify AI development across borders.
• A merger of the major AI labs into one entity, consolidating frontier research under a single roof.
• A single company so powerful it could shut down everyone else's compute, a scenario Altman called "terrible."

The three options differ sharply in who holds the power. The first puts it in a multilateral body. The second concentrates it in a combined industry giant. The third leaves it with whatever firm wins the race outright.

Each path answers the same question differently: who do you trust to control systems powerful enough to be dangerous?

Why does this matter?

The presentation is internal, not policy, but it comes from the CEO of OpenAI — the company whose ChatGPT launch triggered the current AI race. When the person running one of the world's leading AI labs privately maps out endgames that include global governance, industry consolidation, and unilateral compute dominance, it signals how unsettled the question of AI oversight still is at the very top.

The IAEA comparison is the most concrete of the three. The International Atomic Energy Agency exists because nuclear technology was deemed too dangerous for any single nation to manage alone. Invoking that model implies AI may sit in a similar category — a technology whose risks cross borders and demand verification regimes rather than market competition alone.

The second option, a merger of the major labs, would end the competitive pressure that currently drives rapid deployment. It would also raise obvious antitrust questions in the United States and Europe.

The third option needs little elaboration beyond Altman's own word. A world where one company can disable everyone else's hardware is a world with no outside checks on that company at all. Altman calling it "terrible" is the strongest quoted signal in the report about where he draws the line.

What does this tell us about how AI executives think?

The framing itself is revealing. Altman did not present AI safety as a problem solved by better technical alignment or faster research, at least in this presentation. He presented it as a structural problem — a question of how the industry and its power are organized.

That is a governance-first view of AI risk, and it matches a broader drift in how frontier labs talk about the technology's future. The scenarios assume AI will become powerful enough that market structure, not model quality, determines whether outcomes are safe.

It also puts OpenAI's own position in context. A merger of labs, or a single dominant company, are both scenarios in which today's leading labs — OpenAI among them — hold enormous structural power. An IAEA-style watchdog would instead constrain them. Which option wins out shapes not just safety but the distribution of the AI economy's gains.

What happens next?

An internal presentation is a starting point, not a decision. But the report lands at a moment when governments in the US, Europe, and elsewhere are actively debating AI oversight mechanisms, and when the compute needed to train frontier models is concentrating in fewer hands. The three endgames Altman sketched — multilateral oversight, consolidation, or unilateral dominance — map almost exactly onto the live policy debate, and the industry's own CEO has now put all three on record.

By Sophie Lindqvist