The Information reports that Sam Altman told OpenAI employees last week about three possible endgames for safe AI, and The Decoder summarized the account on October 20, 2026. The three: an international watchdog modeled on the International Atomic Energy Agency (IAEA); a merger of the leading labs into one entity, which Altman reportedly called the "Demis Plan"; and a single company that accumulates so much economic power through self-improving AI that it could withhold or switch off everyone else's compute, reportedly the "Eliezer Plan." Altman reportedly called that last option "terrible" and raised it only for completeness.
This is a secondhand, anonymously sourced report about an internal talk, so the useful question is what we can check. Below: the claim, its sources, which parts line up with public record, and what is missing. For context on how labs structure their own accountability, see our posts on AI lab public benefit structures and the OpenAI safety leadership changes.
TL;DR: claimed vs verified
| Claim | Source | Status |
|---|---|---|
| Altman presented three governance endgames to staff "last week" | The Information, citing people familiar, via The Decoder | Unconfirmed by OpenAI; original not reviewed |
| Option 1: IAEA-style body with audits and compute monitoring | Same report | Consistent with OpenAI's public 2023 position |
| Option 2: "Demis Plan," labs merge, CERN-like | Same report | The nicknames are only in the report |
| Option 3: "Eliezer Plan," one firm cuts off others' compute | Same report | Reported as a thought experiment, not a proposal |
| Remark read as a jab at a competitor because of Yudkowsky's ties to Anthropic | One source, per the report | Interpretation by an anonymous source |
| OpenAI will choose a path within a year | Same report | Unconfirmed |
| Trump administration preparing a first framework for superintelligence developers | Same report | Not checked independently here |
What the report says in detail
By The Decoder's account, the first path would create an organization that enforces safety standards, inspects systems, requires audits, and monitors compute and energy use. It would depend on international cooperation, including from countries building frontier models such as China. The second path, the "Demis Plan," draws on a proposal from Demis Hassabis, then head of Google's AI lab DeepMind, for a CERN-like institution. In Altman's reported version the big labs merge into a single entity, or operate under its umbrella, and supply cooperating countries with models.
The third path, the "Eliezer Plan," is named for AI safety theorist Eliezer Yudkowsky. One company becomes so economically dominant through self-improving AI that it can deny compute to everyone else and, at the extreme, shut other systems down. The report says Altman raised it only for completeness, and that one source read the remark as a jab at a rival given Yudkowsky's ties to Anthropic. That reading is the source's inference, not a statement from Altman.
Altman also reportedly complained that too few viable proposals exist for a "speed limit" on AI development, while OpenAI cannot afford to lose ground to rivals. The report adds that both OpenAI and Anthropic are reportedly working toward initial public offerings.
What is on the public record
The IAEA idea is not new for Altman or OpenAI. In May 2023, OpenAI leaders publicly argued for an international authority for superintelligence, comparable to the IAEA, with inspection and audit powers, in the company's post Governance of superintelligence. The Decoder notes the same. Separately, a search for coverage turned up reporting that Altman proposed an IAEA-like international organization again at a New Delhi summit in February 2026, citing concerns that open models could help create pathogens; we did not verify that report against a primary transcript.
Nothing we found on OpenAI's site or Altman's public posts uses the labels "Demis Plan" or "Eliezer Plan." They may be internal shorthand. A search for the phrases returned no earlier public use. That does not mean the report is wrong; internal nicknames rarely appear in public.
What cannot be verified
- The talk itself. There is no recording, transcript or slide deck in the public domain. The Information cites people familiar with the matter.
- Altman's exact words. "Terrible" is attributed to him by The Decoder's summary. We have not seen the original wording.
- The decision timeline. "Within the next year" is a reported plan, not an announcement.
- Whether this reflects OpenAI policy. An exec laying out options in an internal meeting is not a company position.
- The competitor-jab reading. Reported as one source's interpretation.
We could not open The Information's original, which is paywalled; the details here are from The Decoder's account of it. If you need to cite the claim, cite The Information and note that OpenAI has not confirmed it.
Why it matters even if only partly true
- The three options map the real design space. An international inspector, a pooled lab, or a dominant winner are the three shapes most governance debates keep returning to. Hearing a lab CEO name them side by side, even privately, shows how the people building frontier models frame the choice.
- It collides with the IPO and race pressures. The report ties the safety talk to a competitive reality: neither OpenAI nor Anthropic wants to slow down. That tension is the same one behind Anthropic's public benefit fund and the reaction to it and the WSJ report on Anthropic's Long-Term Benefit Trust.
- It lands as US policy takes shape. The report says the Trump administration is preparing its first framework for overseeing superintelligence developers. For what the administration has already said about model review, see our coverage of the Trump AI framework and open-model exemption.
- The compute-control scenario is the one to watch. A world where one firm can deny compute to others is not only a thought experiment for safety theorists. It is a description of concentrated infrastructure, which is why Naval's "god leash" argument and the OpenAI leadership safety departures belong in the same conversation.
How each endgame would work, and where it breaks
The IAEA model. The nuclear agency works because uranium enrichment is physically hard to hide and countries accepted inspections in exchange for access to civilian nuclear technology. The AI analogue would rely on the one thing that is physically visible: large training clusters and the energy they draw. That is why the reported version includes monitoring compute and energy use. The weak points are obvious. Algorithmic progress can reduce the compute needed, open-weight models spread without a data center, and the participation of China is a precondition the report itself flags.
The Demis Plan. A CERN-style institution pools talent and equipment under shared governance, which makes sense for particle physics where one collider serves everyone. Applied to labs that sell competing products and answer to investors, a merger raises antitrust and national-security questions, plus the problem of who controls a single entity that supplies models to cooperating countries. It also concentrates risk: one set of weights, one point of failure, one point of capture.
The Eliezer Plan. This is less a governance proposal than a warning about an outcome. If self-improving AI gives one developer compounding economic power, that developer could in principle starve others of chips, power or cloud capacity. Altman reportedly raised it for completeness and called it terrible. It matters mainly because it describes what happens if neither of the other two paths is built.
What this means for builders and buyers
For most readers, none of these endgames changes this quarter's roadmap. What it changes is how to read lab statements. When a lab says it wants oversight, ask which of these shapes it means, who would be inspected, and whether its own compute would be inside the scope. When two labs both prepare listings, as the report says, ask how safety commitments are written into governance documents and who can enforce them. And when you choose a model vendor for a critical workflow, treat policy risk as a real input: an inspection regime, a compute license, or a merger would all hit pricing, availability and data terms.
How to follow the story
- Watch for OpenAI or Altman responding on X or in a blog post; an official post would be the primary source.
- Watch for a second outlet to confirm the three-option structure and the nicknames.
- Watch for the US framework text, which would show whether any of these ideas are in play at the policy level.
- Treat any viral quote as unverified until it appears in a recording or an on-the-record statement.
If you run AI agents today, governance debates about superintelligence feel remote, but the near-term controls are concrete. AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions, which is the practical end of the same safety question.
Questions readers are asking
Is this a leak of OpenAI policy? No. It is a report of an internal presentation of options. OpenAI has not announced a choice.
Did Altman endorse shutting down rivals' compute? According to the report, no. He reportedly called the option terrible and raised it for completeness.
Why name a plan after Yudkowsky? The report says the label refers to the scenario of a dominant developer cutting off others' compute, associated with Yudkowsky's arguments. Whether the nickname was meant as a dig is, again, one source's reading.
This post reflects reporting available on October 20, 2026. OpenAI had not commented in anything we found; we will update if it does.
Related reading
- AI lab public benefit structures compared
- Aleksander Madry returns to OpenAI
- Anthropic Public Benefit Fund reaction
- Anthropic Long-Term Benefit Trust: claimed vs verified
- Trump AI framework and open-model exemption
- OpenAI leadership safety exodus
- Naval's god leash and superintelligence alignment
- Sources: The Decoder, OpenAI, Governance of superintelligence