Altman tells OpenAI staff there are three endgames for AI: an IAEA for AI, a merger of the labs, or one firm that "melts everyone else's GPUs"
What's going on? OpenAI chief executive Sam Altman walked employees last week through three ways the race to build superintelligence could be made safe, The Information reported on 19 October, citing people familiar with his comments. The framing is stark, and it lands just as Washington sets up its first system for policing "superintelligence" developers.
The three endgames: First, an international agency modelled on the International Atomic Energy Agency that would enforce safety standards on frontier labs. Two people said Altman appeared to favour this option. Second, what he called the "Demis plan", after Google DeepMind chief Demis Hassabis's years-old idea of a CERN for AI: the big labs would merge into, or work through, a single body that supplies models to participating countries. Third, the "Eliezer plan", named after AI-safety theorist Eliezer Yudkowsky: a developer that reaches recursive self-improvement uses its economic power to starve every rival of compute, or in Altman's words "melt everyone else's GPUs". According to one person, he said "obviously I think that one's terrible" and included it only for completeness.
The bigger fight: who gets to govern frontier AI? None of the three ideas is new. The IAEA model and a joint project of the major labs both appear in OpenAI's May 2023 post "Governance of superintelligence", written by Altman, Greg Brockman and Ilya Sutskever, and the "pivotal act" idea comes from Yudkowsky's 2022 essay. What has changed is the setting. The White House's new Super Intelligence Force, chaired by Director of National Intelligence Jay Clayton, adopted its charter last week and now requires labs to report and fix rogue-model incidents. Anthropic is circulating its own draft federal bill that would let the Commerce Department halt a model only above 50 deaths or $1B in damage. OpenAI, Google and Anthropic are also reported to be building an industry "Standards Authority for Frontier AI" but disagree over who should test new models, while Meta's Mark Zuckerberg and Nvidia's Jensen Huang have pushed back on such frameworks. Hassabis himself has more recently backed a FINRA-style self-regulatory body rather than a CERN.
The subtext: Yudkowsky is said to have links to key Anthropic figures, and reports relaying The Information say some staff read the "Eliezer plan" label as a veiled dig at OpenAI's main rival. The two labs are racing each other towards IPOs, and OpenAI is trying to raise about $30B ahead of Anthropic's listing. Altman reportedly also told staff that the first two options need huge coordination between companies or governments, that OpenAI cannot simply stand down in the race, and that it needs to decide within a year which path to push.
What's confirmed and what isn't: The account comes from The Information's sources. OpenAI has not commented publicly. The only direct Altman quotes ("melt everyone else's GPUs" and calling the third plan "terrible") come from that report and the outlets relaying it. The idea that the label was a dig at Anthropic is staff interpretation, not something Altman said.
Why it matters: OpenAI's chief executive is telling his own employees that the realistic endings are a global regulator, a merger of the labs, or a winner-takes-all lab he calls "terrible". It shows how far frontier-AI governance has moved, from a policy footnote to a strategic choice that will shape OpenAI's IPO story, its rivalry with Anthropic and its stance towards the new oversight in Washington.
Sources: The Information (Stephanie Palazzolo, Amir Efrati, 19 Oct 2026); Yonhap and India Today (20 Oct 2026); OpenAI, "Governance of superintelligence" (May 2023); Eliezer Yudkowsky, "AGI Ruin" (2022); Axios (9 Oct 2026) and The Next Web (16 Oct 2026) on the Super Intelligence Force
