OpenAI CEO Sam Altman has reportedly presented employees with three possible paths for keeping advanced artificial intelligence under control, including a scenario in which one AI company becomes so powerful that it can deprive its competitors of the computing hardware needed to develop their own models.
According to a report, Altman discussed the three proposals with OpenAI employees the previous week. The options range from international oversight of AI development to merging leading AI laboratories into one organisation, with a third scenario concentrating control in a single company. Altman said he did not support the third option, describing it as a terrible outcome that he included to complete the discussion.
1. An international agency to oversee AI
The first proposal would establish an international organisation modelled on the International Atomic Energy Agency (IAEA), which monitors nuclear activities and helps verify compliance with international safeguards.
Applied to AI, such an organisation could establish safety standards, monitor the development of powerful systems and coordinate oversight across countries.
Altman appeared to favour this approach. OpenAI has previously advocated for an international body capable of inspecting advanced AI systems and tracking important resources used to develop them.The challenge would be securing cooperation among governments and AI companies, particularly as countries compete to develop increasingly capable models.
2. Bringing AI laboratories under one organisation
The second option, which Altman reportedly called the “Demis plan,” draws on an idea associated with Google DeepMind CEO Demis Hassabis.
It envisions leading AI laboratories merging into, or working through, a shared organisation similar to CERN, the international research centre jointly operated by participating countries.
Under Altman’s reported version, the organisation would develop advanced AI models and make them available to countries working with it.
The approach would seek to reduce the risks created by competing laboratories racing to develop increasingly powerful AI systems. However, it would require an extraordinary level of cooperation among companies that currently compete for talent, investment, computing resources and customers.
3. One company controls the computing power
The third scenario, called the “Eliezer plan,” refers to AI safety theorist Eliezer Yudkowsky, who has argued that the development of superintelligence could pose an existential threat.
In this scenario, one AI company reaches recursive self-improvement, meaning its AI systems can help improve their own capabilities. The company then becomes so powerful that it can prevent rival developers from catching up by controlling or destroying their access to computing resources.
Altman reportedly described this possibility as a company becoming powerful enough to “melt everyone else’s GPUs.” GPUs, or graphics processing units, are essential hardware for training and running many advanced AI models.
The idea echoes Yudkowsky’s argument that the first developer to achieve superintelligence might carry out a “pivotal act” to prevent others from building dangerous systems. That concept has included disabling the computing infrastructure other developers rely on.
However, the report does not suggest that OpenAI is planning to destroy competitors’ hardware. It describes a hypothetical future that Altman reportedly presented as an undesirable outcome.
OpenAI faces a difficult safety question
The three scenarios illustrate a central problem in AI governance: how to reduce the risks of increasingly powerful systems without allowing control over the technology to become concentrated in the hands of a single company.
International oversight and shared research would depend on cooperation between governments and competing developers. The third scenario would instead rely on one company gaining overwhelming power, an outcome Altman reportedly said he did not want.
Altman believes OpenAI needs to determine which direction to support over the coming year, even as it continues competing with other leading AI developers.
None of the three proposals has been established as a final plan for OpenAI. For now, they represent possible approaches to a problem the industry has yet to resolve: how to develop more powerful AI while keeping it safe and accountable to humans.