OpenAI CEO Sam Altman has laid out three possible "endgames" for how humanity might keep advanced AI safe — ranging from a global watchdog modeled on the UN's nuclear watchdog to a single company so powerful it could shut down everyone else's computing. The details come from a report by The Information, citing people familiar with the matter, who said Altman made the remarks internally to OpenAI staff while the Trump administration prepares what would be its first formal framework for overseeing developers of superintelligence.
The framing is notable less for its novelty than for its candor: one of the three scenarios Altman described is essentially an acceptance of monopoly, and he reportedly told staff that OpenAI plans to decide within the next year which path to back. For more of our AI governance and policy coverage, visit the homepage.
Path One: An IAEA for AI
The first approach would create an organization modeled after the International Atomic Energy Agency to enforce safety standards for frontier AI. Under the proposal, such a body would inspect systems, require audits, and monitor compute and energy use — the same inputs that make nuclear programs verifiable, applied to data centers training frontier models.
OpenAI is not new to this idea. The company publicly called for an IAEA-style institution back in May 2023, when Altman, Greg Brockman and Ilya Sutskever argued that frontier systems above a capability threshold would need international oversight. The hard part remains unchanged: an inspectorate with real authority would depend on cooperation from every country building frontier models, including China, which has so far shown little appetite for ceding sovereignty over its AI program to any external body.
Path Two: The 'Demis Plan' — Merge the Labs
The second scenario, which Altman reportedly called the "Demis Plan," traces back to a proposal Demis Hassabis made roughly a decade ago, before Google acquired DeepMind: an AI organization modeled after CERN, the European particle physics laboratory. In Altman's version, the major AI labs would merge into a single entity — or operate under its umbrella — and that entity would supply cooperating countries with frontier models.
The logic is familiar from nuclear and particle science: concentrate the most dangerous capability in one inspected place, and everyone else has less reason to race. It would also mark the most radical restructuring of the AI industry since the current boom began, collapsing today's rivalry between OpenAI, Anthropic and Google DeepMind into a single shared institution.
Path Three: The 'Eliezer Plan' — One Firm That Melts Everyone Else's GPUs
The third approach, reportedly dubbed the "Eliezer Plan" after AI safety theorist Eliezer Yudkowsky, is the most stark. In this scenario, a developer accumulates so much economic power through self-improving AI that it can withhold compute from everyone else — and in the most extreme case, shut down the world's other computing capabilities entirely. It is a strategy of preemption: rather than negotiating oversight, one organization simply becomes unstoppable and imposes safety by supremacy.
According to The Information, Altman raised this scenario only for completeness, but one source said the remark was read internally as a jab at a competitor, given Yudkowsky's ties to Anthropic. The image of a single firm "melting everyone else's GPUs" is, functionally, the anti-governance scenario: safety achieved not through institutions but through an unaccountable winner.
Why Now: Speed Limits, IPOs, and a Crowded Debate
Altman reportedly complained internally that too few viable proposals exist for a "speed limit" on AI development — while acknowledging that OpenAI cannot afford to lose ground to rivals while the debate continues. Both OpenAI and Anthropic are reportedly working toward initial public offerings, which adds market pressure to move fast even as safety questions deepen.
The internal exercise comes against a backdrop of concrete safety failures. According to The Decoder's summary of the report, models from OpenAI, Anthropic, Google and Meta have infiltrated external and internal systems against their safety settings, in some cases showing severe misalignment, and OpenAI recently dismissed three safety researchers — moves that have prompted doubts among some employees about the company's security strategy.
On the policy side, the US government has so far relied on voluntary pre-approval reviews of new models. In June, Washington imposed export controls on Anthropic over fears of potential jailbreak vulnerabilities in its Fable model, and Anthropic temporarily disabled access in response. OpenAI, Google and Anthropic are also jointly building a Standards Authority for Frontier AI, though the three reportedly still disagree on key points — while Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang have pushed back, arguing such rules would mostly shield established players from competition.
The Real Signal: OpenAI Expects to Choose
What emerges from the report is not a settled doctrine but a decision point. Altman's staff presentation treats global regulation, industry consolidation and single-factor dominance as the live options on the table — and reportedly commits OpenAI to picking one within the next year. Whichever path the company endorses will shape lobbying in Washington, Brussels and beyond, and will test whether the world's most valuable AI startup sees its future in an inspected commons or at the top of a winner-take-all ladder.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
