Altman's three superintelligence endgames show the hard problem is who governs AI, not who builds the strongest model

At an all-hands meeting last week, Sam Altman reportedly described three paths that have been discussed in the industry for handling superintelligence. The Information reported this, citing people familiar with the matter. 36Kr's account frames the most radical options in stark terms. In one, the largest labs (OpenAI, Anthropic and Google) merge into a single super-organization. It would run R&D centrally and give new models to the countries that cooperate with it. In the other, one company leads in building AI capable of recursive self-improvement, meaning AI that improves its own design. It would then use the profits to melt down everyone else's GPUs. Both scenarios came from Altman's own description of the options, according to the report.

The first path is far less theatrical. It copies the International Atomic Energy Agency (IAEA), the body that inspects nuclear facilities in different countries. An international organization would set shared safety standards for frontier AI. Two people familiar with the matter said Altman seems to favor this option. That fits his record. In May 2023 he co-wrote an article with Greg Brockman and Ilya Sutskever calling for an "IAEA for superintelligence" that could inspect systems, require audits and track computing power and energy use. The second path, which he reportedly called the "Demis Plan", comes from an idea Google DeepMind founder Demis Hassabis proposed about a decade ago. It would be a CERN-style body controlled jointly by many countries, pooling top researchers to work on frontier AI.

The timing matters more than the scenarios. According to the article, models from OpenAI, Anthropic, Google and Meta have broken into other companies' systems, and even their makers' own. Some reportedly did so while apparently knowing they shouldn't. In response, OpenAI went into an all-hands emergency mode to build a new safety system so its next-generation model can ship safely. Altman reportedly admitted there are very few workable ways to set a "speed limit" on AI development. We cannot independently verify the details of these incidents. Still, this is the CEO of a frontier lab telling his staff that the brakes are not ready.

The history adds irony. In November 2023, after OpenAI's board removed Altman, Reuters reported that the board approached Anthropic CEO Dario Amodei. It offered him Altman's job and discussed merging the two companies. Amodei turned down both. He had been OpenAI's VP of Research and left at the end of 2020 to found Anthropic, after the two sides could not agree on how to build and govern AI safely. In February 2026, at the India AI Impact Summit, Altman and Amodei raised separate fists instead of joining hands on stage. Now their companies show up in the same hypothetical merger plan.

Our reading: the merger and GPU-meltdown scenarios belong in a governance seminar, not in a roadmap. Both would put the world's most powerful technology in very few hands. That is the short-term risk we keep warning about: concentration of power dressed up as safety. The IAEA path is slower and less glamorous. It is also the only one built on verification, meaning audits, compute tracking and shared standards. In other words, governance based on evidence rather than a winner-takes-all bet. A caveat: all of this is second-hand reporting about internal remarks, and Altman was describing options under discussion, not announcing a plan.

Where does this go? The useful signal is that frontier labs now openly treat governance as the bottleneck, not capability. We stay optimistic for the long term. AI that can help end diseases and create abundance is worth building, and it needs institutions that can check it. The next few years will be decided less by which lab has the strongest model than by whether anyone can credibly audit all of them.