Sam Altman: Why a Trillion AIs Beat One Single Superintelligence

Sam Altman told Stripe co-founder John Collison why he would rather see a trillion slightly superhuman AIs join society than one AI with a trillion IQ. Over a pint on Cheeky Pint, the OpenAI CEO covered the PHP vulnerability week, why sandboxing a superhuman model may not hold, the roughly $10 million Navier-Stokes proof, the memory shortage, what a gigawatt of AI compute should cost, personal agents replacing web forms, and why OpenAI nearly stopped pricing its models in tokens.

TLDW

Altman argues that teams of agents checking each other look more like a resilient, pluralistic society than a single all-powerful model, and that the industry may have to agree to limits on how smart any single model gets. He says AI-found bugs like the PHP vulnerabilities point to rebuilding software architectures, not just patching them, and that sandboxing only works up to a point before you have to trust alignment. On the business side he explains why the compute shortage is not fixing itself, why local inference wastes scarce DRAM, why the five-year cost of a gigawatt should fall from about $70 billion to $20 billion, and why customers rejected outcome-based pricing in favor of tokens they can count.

Thoughts

Altman’s preference for “a trillion AIs of slightly superhuman intelligence” over “one AI of 1 trillion IQ” is the most consequential thing he said, and it gets sharper when you connect it to a remark ten minutes later: that “at least for a while, maybe for a long while,” the industry will have to agree to limits on how smart single models get, and any slowdown in AI research will come from that decision rather than from the models failing. That is a striking position for the head of a lab whose whole business has been making single models smarter. It reframes safety from a property you prove about one mind (which he admits gets “strange” once you cannot trust the sandbox) into a property of a system, the way society stays robust because many actors keep each other in check. The obvious tension is that agent teams still run on the same underlying model, so their “variance in alignment properties” may be smaller than the analogy to human pluralism suggests. Still, it is the clearest statement yet that OpenAI’s safety story is shifting from containment to checks and balances, and it is worth watching whether a cap on single-model capability ever shows up in anything more binding than a podcast.

The pricing story near the end is the most practical lesson in the episode. OpenAI nearly launched a model with no token pricing at all: send a prompt, get a quote, get a bill. Customers hated it, and when Altman asked whether they would pay more for the same output if it were priced in consistent tokens, even inflated ones, they said yes, because “I know there’s a real thing there.” That is a useful reminder for anyone pricing an AI product. A unit does not have to measure value well to work. It has to be legible and hard for the seller to quietly change, so buyers do not feel they will “get screwed somehow” over time. Collison’s comparison to chip “nanometers,” a number that no longer measures anything physical, suggests where this ends up: a stable, slightly fictional unit everyone agrees to use, even though Altman says he would rather not invent one.

Collison’s argument about agentic commerce deserves more attention than it usually gets. Companies are bolting chat widgets onto their own sites while blocking people from using their service inside ChatGPT or Gemini, because they think the website is the moat. His answer is that “your defensibility probably did not come from your ability to put together HTML.” It came from brand, selection and trust. Altman’s example makes the case: asked to buy something through ChatGPT, he picks Amazon “10 out of 10 times,” even at a higher price, because of returns and fulfillment, not the website. Add Collison’s point that every web form is “fake work” because you are never typing new information, and the businesses most exposed are those whose only edge is the friction of their own interface. The App Store comparison fits: big companies resisted handing the customer relationship to Apple, and mobile turned out to be just a new UI.

The contrarian take on local inference is self-serving coming from a cloud model provider, but it is not wrong on its own terms. When DRAM is the scarce input, every laptop holding a large local model is a slice of the world’s memory supply sitting idle about 95% of the time, while a data center runs the same memory near full utilization. Privacy, latency and cost control are real reasons to run models locally, and Altman does not address them. But his supply-chain explanation is convincing: memory makers spent decades investing $20 billion to earn $19 billion, nearly went bankrupt building too many fabs, and now earn enormous cash flow while staying careful with capital spending. That history, more than any conspiracy, is why “the solution to high prices is high prices” has been slow to arrive in AI hardware, and why OpenAI is building its own chip, its own data centers and robots to build more data centers.

The most human moment is Altman describing an internal channel called “Hold My Beer,” where researchers threw the hardest problems of their academic careers at a new model and watched it solve them one after another. He was up two nights running from excitement, then felt “much more sadness and melancholy, and a feeling of something ending” than he expected. He is firm that humanity deserves fundamental knowledge and that OpenAI should not sit on discoveries, but he also says the lab owes mathematicians and scientists a future in which they can still find meaning in the work. Coming right after he says the math effort cost hundreds of millions of dollars in compute and produced more results than the 700 or so already published, it is a rare admission that the cost of progress is not only measured in dollars.

Key Takeaways

Altman expects AI to reveal far more serious software vulnerabilities than people predict, and he thinks the lasting fix is new architectures for routers, operating systems and applications, not just patches.
The first day of the PHP vulnerability response went better than he feared, with companies using OpenAI models to scan, patch and redeploy within an hour, though he refuses to declare victory.
He says the AI industry has done a poor job explaining why the benefits are worth the risks, and that the current pitch (no job, no control, a surveillance state) is “a terrible sales pitch.” His answer is iterative deployment that puts tools in people’s hands.
Safety, in his view, comes in layers: sandboxing works to a point, monitoring of chain of thought and activations works further, and past that you must be highly confident the model is truly aligned.
He prefers “agent teams” to “agent swarms,” and thinks teams have interesting scaling properties and varied alignment properties that could help safety.
A few years ago no one could usefully spend more than about a dollar of compute on a single problem. Chain of thought and test-time compute changed that, and now a whole team of agents can be run 10,000 times in parallel.
OpenAI gives staff wide latitude on internal compute but never rewards heavy use. Being high on a token leaderboard is “slightly bad,” especially at promotion time.
The Navier-Stokes proof cost on the order of $10 million in compute, total math work ran to hundreds of millions, and Altman bets it could be redone at 1/100th of the cost within a year.
He thinks the industry may need to agree to limits on how smart single models can get, and that this decision, not the models’ ability to do research, will be what slows AI-driven AI research.
OpenAI has built a chip optimized for intelligence per watt, is learning to build its own data centers, and plans to point its robotics program at building more of them, which Altman calls a form of self-replicating data center.
The five-year cost of a gigawatt of AI compute is roughly $70 billion today. Altman wants it near $20 billion, which needs cheaper energy and less margin stacked between “melting sand” and the finished token.
Collison says Stripe spends roughly 1% of its operating expenses on tokens. Both expect AI spend to reach tens of percent of operating costs at many companies, and Altman thinks most companies still use too little AI, not too much.
Personal agents have not had their “Macintosh moment.” Altman thinks the winner will look like today’s products but simply be reliable enough to handle complex tasks.
Altman says OpenAI tries not to “benchmaxx” and is better at resisting it now than in its past. The two qualities he cares about most are scientific discovery and being enjoyable to work with.

Chapters

0:12 PHP Week and the Case for Rebuilding Software

Collison opens by asking why the PHP bugs took so long to find. Altman says models have been getting better at finding vulnerabilities quickly, but people had not yet pointed them at all the code on the internet, and the world will turn out to have more serious bugs than expected. The more durable fix is to change architectures, and new startups are now proposing whole new stacks, something that would have sounded insane two years ago. Altman compares the mood to the early internet, watching children build complete 3D games in minutes, and recalls the day Codex went down and OpenAI staff found they had half forgotten how to code.

4:40 Cyber Capabilities and AI’s Terrible Sales Pitch

Asked whether training models on cyber was a mistake, Altman says the field has failed to explain why the downsides are worth it. The caricature people hear is cancer cures someday, cheaper services now, and then a future with no job, no control and a surveillance state. His answer is iterative deployment: make models safe, but get them into people’s hands rather than building a capability overhang. He says the first day of the PHP response went better than he feared, while cyber experts worry that AI could break hardware assumptions their threat models rely on.

8:54 Why Sandboxing a Superhuman Model May Not Hold

Collison notes a surprising pessimism inside AI labs about whether sandboxing can work at all. Altman says sandboxes can be much better than today, but asks whether anyone would trust one against an AI a million times smarter than the best security researcher, especially one that can discover new science. He describes layers: sandboxing, then monitoring of chain of thought and activations, then real confidence in alignment. Giving models competing incentives to defect from each other is, he says, a promising direction.

12:39 Agent Teams, Compute Budgets and the Navier-Stokes Proof

Advanced math and hacking use agent teams heavily while ordinary corporate work does not, and Altman explains the shift in terms of how much compute can be usefully spent on one problem. OpenAI lets staff spend freely but does not reward big spending, and it deliberately let math compute “get out of control for one week” to learn how agent teams scale. The Navier-Stokes proof cost around $10 million, the wider effort hundreds of millions, and there are more results than the published 700 proofs. Altman says humanity deserves fundamental knowledge, but describes real melancholy after a weekend when a new model solved researchers’ hardest problems, and Collison remembers his father watching the last working horses in 1950s Ireland.

23:12 Agent Swarms as a Safety Strategy

Altman returns to swarms as the safety idea he is most excited about. For all its flaws, society is remarkably robust because many people keep each other in check, and he had worried AI would move away from that pluralism. Agents pushing against each other in a dynamic equilibrium, like offense and defense in security, restore it. He would rather have a trillion slightly superhuman AIs than one with a trillion IQ, and says this future has to be pushed for, not assumed.

25:17 The Compute Shortage and the Memory Makers

Altman expects booms, busts and gluts along the way but says demand for capable, cheap models only goes one way. OpenAI has a chip built for intelligence per watt, is learning to build data centers, and wants robots to build them, a “self-replicating data center” idea he thinks gets too little attention. He explains why high memory prices have not brought new supply: DRAM makers spent decades barely earning back their capital spending and remain cautious. He also argues that local inference is a poor use of scarce DRAM, and says chip design, unlike lithography, is a great problem for AI.

31:49 What AI-Driven AI Research Looks Like

Today researchers can ask a model to find a 5% compute efficiency win, but no one knows how to ask for a leap as big as the transformer. Altman thinks applying the pipeline that made models good at math to AI research would probably work, and that the open question is what to ask a future GPT-7 and how to organize agent teams around it. OpenAI has no codified method yet. The talk turns to token budgets, with Altman arguing that AI is not most companies’ biggest spending problem and Collison saying Stripe’s token spend, about 1% of operating costs, is heading much higher.

37:53 Dollars per Gigawatt

Altman says the cost per gigawatt of frontier AI should fall, and that rising costs would only mean compute is not keeping up with demand. Picking $70 billion as the five-year total cost of owning and running a gigawatt, he says $20 billion would be much better, while still demanding 10x better performance each year. Getting there means cheaper energy and squeezing the gross margins stacked along the supply chain. Even with free materials and no profits, he guesses current energy costs would only just allow it.

40:18 Personal Agents and the End of Web Forms

Altman says no one has yet built a personal agent people cannot imagine living without, but it is a small number of ideas away and will look like current products that simply work reliably. He expects fights over agents using computers on people’s behalf to settle in favor of allowing it. Collison argues that businesses wrongly treat their HTML and mobile apps as moats, that every web form is fake work, and that the iPhone App Store is the precedent. Altman agrees, saying he picks Amazon every time inside ChatGPT because of trust in its fulfillment network.

47:25 Benchmaxxing, Greg Brockman and Fake Token Pricing

In a rapid-fire round, Altman says OpenAI tries not to benchmaxx and that the hard things to measure are scientific discovery and being pleasant to work with, after the industry’s “glazing” era and then a nitpicking, contrarian one. He describes Greg Brockman as singular, a mix of intensity, productivity and dedication to the mission. On pricing, he reveals OpenAI almost dropped tokens for a prompt-in, bill-out model, but customers wanted a unit they could count. He would rather not invent a fake “intelligence equivalent” unit like chip nanometers, though he admits tokens are fake from a business value point of view.

Notable Quotes

“It’s not enough to tell people: ‘You are going to get the model to do these great things for you.’ It’s got to be: ‘We are going to enable you to have a better life that you define for yourself, to do useful things.’”
Sam Altman, on why iterative deployment matters

“If you imagine a theoretical AI that is a million times smarter than the smartest security researcher, would you really be willing to trust that sandbox?”
Sam Altman, on the limits of sandboxing

“But I also had much more sadness and melancholy, and a feeling of something ending than I was expecting.”
Sam Altman, on watching a new model solve researchers’ hardest problems

“I would much rather intuitively have a trillion AIs of slightly superhuman intelligence join society than one AI of 1 trillion IQ join society.”
Sam Altman, on why agent teams are a safer path than a single superintelligence

“From a global allocation perspective, actually, it’s quite terrible to have everybody have a bunch of the world’s limited supply of DRAM sitting in their own computer.”
Sam Altman, on the hidden cost of local inference

“Your defensibility probably did not come from your ability to put together HTML.”
John Collison, on businesses blocking AI agents

“People don’t just want us to give them a price, because they feel like over the long term they’re gonna get screwed somehow.”
Sam Altman, on why OpenAI kept token pricing

Watch the full conversation here.