Karmactive Team
OpenAI CEO Sam Altman told staff there are three ways the race for superintelligence ends — and one of them involves a single AI developer accumulating enough computing power to render every competitor's GPU clusters obsolete. The internal address, reported by The Information, frames physical hardware control as the deciding factor in who survives the next phase of AI competition.
Altman described three scenarios in an address to OpenAI staff. The first follows an IAEA model: an international safety agency with inspection authority over frontier AI development, similar to how nuclear facilities are monitored. The second, which Altman referred to as the "Demis plan," involves a CERN-style joint laboratory where leading AI developers collaborate on frontier research under shared governance. The third — the "Eliezer plan" — describes a single self-improving AI developer that accumulates sufficient hardware compute to make every competitor's GPU clusters obsolete. Altman reportedly described the third scenario as "terrible," saying he included it for "completeness" rather than as an endorsed outcome. He said he personally favors the IAEA model because it allows established developers to maintain their intellectual property while subjecting all frontier labs to external inspection.
The practical consequence for cloud infrastructure providers, semiconductor manufacturers, and enterprise AI buyers is that competitive strategy in AI is shifting from algorithmic design toward physical chip supply. If a single developer achieves recursive intelligence gains while simultaneously commanding leading semiconductor supply lines, rival cloud providers could face hardware capacity that no longer keeps pace with state-of-the-art model training. That scenario, according to the reported remarks, is why Altman favors international oversight mechanisms over either a collaborative lab merger or unchecked competitive dynamics.
The GPU Bottleneck Behind Altman's Three-Way Split
The three endgame scenarios each reflect real constraints in how frontier AI models are built today. Training leading frontier AI models can require tens of thousands of GPU accelerators running simultaneously for weeks or months, depending on the model's scale. The global supply of the highest-performance chips is concentrated in a small number of fabrication facilities, primarily in Taiwan. The companies with the largest existing compute investments — including OpenAI, Google DeepMind, Microsoft, and Meta — have structural advantages over later entrants, as hardware cannot be rapidly deployed regardless of budget.
Altman's reference to "melting everyone else's GPUs" describes what happens if a self-improving AI achieves recursive capability gains faster than competitors can acquire matching compute. The leading system could use its existing resources to solve chip design problems faster than human engineers, produce more efficient training architectures, and accelerate its own improvement cycle until no other organization has the raw computational capacity to keep pace. AI compute infrastructure and the GPU supply chain and how semiconductor manufacturing shapes AI development cover the hardware dynamics in detail.
The IAEA analogy Altman prefers is not a new proposal. AI safety researchers have referenced nuclear inspection regimes as a governance model for years. What Altman's framing adds is specificity about governance at the hardware level rather than the software level.
The "Demis plan" references a CERN-style collaborative framework associated with DeepMind CEO Demis Hassabis. More recent reporting notes Hassabis has expressed support for a FINRA-style regulatory approach.
Data centers in North America and Europe are already facing power grid capacity limits that restrict how many high-density GPU racks can be installed in new facilities. Energy demand from AI training workloads is now a factor in grid planning for multiple US regional transmission operators. AI data center energy demand and power grid pressure explains how electricity infrastructure is shaping facility deployment decisions.
AI safety proposals referenced in the leaked address draw on Eliezer Yudkowsky's frameworks around recursive self-improvement and Demis Hassabis's collaborative lab model. Altman's stated preference for the IAEA approach reflects his long-held view that international oversight is preferable to either lab consolidation or unchecked competition.
What are Sam Altman's three AI endgames?
Altman described an IAEA-style international agency that inspects frontier AI compute installations, a CERN-style collaborative lab where leading developers share frontier research infrastructure, and a winner-take-all scenario where one self-improving AI captures enough hardware compute to make every competitor's GPU cluster obsolete.
OpenAI has not released a formal policy document based on the internal address. Coverage will be updated as official statements are released.
