Sam Altman Says He'd Rather Have a Trillion AIs Than One Superintelligence
On Stripe co-founder John Collison's Cheeky Pint podcast, the OpenAI CEO also talked about robot-built data centers, wasted memory in local AI and falling compute costs.
Published
Last updated / Fact-checked on
OpenAI CEO Sam Altman argued against building a single, all-powerful AI on the October 29, 2026, episode of Cheeky Pint, the podcast hosted by Stripe co-founder John Collison. "I would much rather intuitively have a trillion AIs of slightly superhuman intelligence join society than one AI of 1 trillion IQ join society," he said. The 53-minute episode is titled "Why Sam Altman is against a single superintelligence."
Image: Illustrative photo of two people chatting over beer. Clem Onojeghuo via Unsplash / Wikimedia Commons — CC0 1.0
Many AIs, not one
Altman tied his preference to how human societies stay stable. "For all of the deep flaws of society, it is unbelievably robust," he said, attributing that to people "keeping each other in check and collaborating" and to "these values of pluralism and democracy." He said he had been "sort of sad" that AI seemed to be moving away from that model but is now excited about agent teams.
He prefers the term "agent teams" to "swarms." He said such teams likely have "very interesting scaling law properties, but also variance in alignment properties," and should be studied for safety as well as capability.
Later in the episode, discussing AI that does AI research, Altman said that "at least for a while—maybe for a long while—we are going to have to agree to some limits about how smart we let these single models get."
Robots that build data centers
Asked about the compute shortage, Altman listed OpenAI's efforts:
| Effort | What Altman said |
|---|---|
| Custom chip | "Incredible performance" in units of intelligence per watt |
| Data centers | "We are learning how to build our own data centers" |
| Robotics | A program "we're going to point at making lots of data centers" |
He said people talk about recursive self-improvement as AI doing AI research, but there is "relatively little talk about self-replicating data centers." He described using AI to control robots that run fabs and build turbines, solar panels and data centers. "I think we have that within grasp," he said.
Collison pushed back that data centers are bottlenecked by permits and components, not labor. Altman replied that robots could make more components and agreed that this would play out "over a long period of time." He gave no timeline.
Why local AI wastes memory
Altman disputed the idea that running open models on personal computers will relieve pressure on data centers. From a global allocation view, he said, it is "quite terrible" to have the world's limited DRAM sitting in individual computers and being used "5% of the time." In a supply-constrained market, he said, cloud deployment maximizes utilization.
He said the market isn't fixing the shortage quickly. "I think the invisible hand is failing right now," he said. Memory makers, he explained, historically would "invest $20 billion of CapEx to make $19 billion" and nearly went bankrupt in past busts. So they remain cautious even as profits surge.
Compute should get cheaper
Altman said the cost of a gigawatt of AI compute should fall. Using an illustrative $70 billion for five-year total cost of ownership, he said, "if that number were 20 instead of 70, we should be much happier." Getting there would require cheaper energy and squeezing out the margins stacked along the supply chain.
Why it matters
These are views from the head of one of the largest AI developers, not company commitments. Altman offered few specifics on the robotics program or on what limits on single models would look like. Still, three points stand out for readers following AI:
- Safety strategy. OpenAI's CEO is publicly favoring many cooperating agents over one dominant system.
- Physical bottlenecks. Memory, power, permits and components, not just model quality, limit AI growth.
- On-device AI. It is not automatically the efficient alternative to the cloud.
More is in our Artificial Intelligence hub.
How we reported this
We checked every quote against the official episode transcript on Cheeky Pint's podcast host. We restored the exact wording of Altman's main quote, moved his remark about limits on single models to its correct context, and clarified that the $70 billion figure is a five-year total cost of ownership.
Published and fact-checked October 30, 2026.
Sources
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.