Critics mocked his early compute push. He now says it was not ambitious enough

If intelligence is going to become as universally available as electricity, compute comes first. Anthropic CEO Dario Amodei once nicknamed Altman the "YOLO CEO" because of how aggressively he pursued compute early on.

The conviction came from two observations. OpenAI could see that model improvement was following an exponential curve. That part felt certain. The company also believed that if costs kept falling, demand for AI that was capable enough and cheap enough would be effectively unlimited. Altman compared it to a rare new kind of commodity, invoking two famous early misjudgments about computing: "The world only needs five computers," and "Nobody needs more than a certain amount of memory."

Then he reduced the argument to its essence:

No matter how efficient the algorithmic layer becomes, what we're doing is turning electricity into useful intelligence. Given the shape of demand, we're going to want more.

Sam Altman

The conviction began with GPT-4. Altman emphasized that it was not GPT-3, "not even 3.5." Once the model became that intelligent, OpenAI knew it could find a way to make reasoning work. Effective reasoning would lead to what became known as agents: AI systems that can act and complete tasks on their own.

Then came the phone calls to cloud providers, chipmakers, and energy companies. "Everybody said, you're completely crazy. This isn't possible. No industry has ever moved like this. We've been doing this a long time. There are cycles and ups and downs. It can't go straight up. This is irresponsible." He described it as raising money for an early-stage startup: most people say no, but you only need one or two yeses.

The host asked whether Altman had been proven right—and perhaps had even bet too little. His answer was blunt: "We definitely bet too little."

How tight is compute? Altman offered one number. It concerned trial runs conducted before training, not the training run itself:

The bottleneck keeps moving

What holds OpenAI back has changed repeatedly. Seven or eight years ago, it was research ideas. Back then, he said, "all the compute in the world couldn't help you, because we lacked the idea." Once the ideas arrived, the task became scaling them, so compute became the bottleneck. Then the data ran out, making data the constraint. Now it is compute again. He added that research breakthroughs had regained the upper hand over the past six months. Compute and research are not truly separable: more compute enables better research and more experiments.

He thinks the most durable advantage is a vast compute fleet

While discussing Codex's rise, Altman said something that may surprise people: Codex won mainly because it had the best product and the best model. Its advantage from being bundled with ChatGPT was "very, very small." That has made him think repeatedly about competitive moats, because strong intelligence can migrate from any product to any other product.

Altman's answer to data-center backlash: build in the desert

Compute ultimately becomes physical data centers, and the people living nearby genuinely dislike them. Altman said he would like to take people to see a gigawatt-scale data center, whose power demand is comparable to that of a power plant. Hearing about one, seeing photos or video, and actually standing beside it are three completely different experiences.

"Each one will be among the most expensive infrastructure projects humanity has ever built. And we're building many of them now."

He also said he understands the resistance: "I don't particularly want a nuclear power plant next to my house, even though I know it's super safe." His solution is that unlike power plants, data centers can be built almost anywhere. "Build them in the desert, in places nobody wants to go. That's fine. AI systems are perfectly happy to live there."

On environmental concerns, he addressed two issues. Water: a few years ago, data centers relied on evaporative cooling and used enormous amounts of water, but now they use closed-loop systems. Electricity: the industry is moving from fossil fuels toward solar and nuclear power. His summary was, "We've done a pretty good job solving water. Energy is next."

Extracting more from existing compute beats building another facility

Asked about other creative options, Altman said the biggest current returns come from software innovations that extract more intelligence from existing compute. "My sense is there are still several orders of magnitude available here." On hardware, he highlighted OpenAI's Jalapeño, which targets specific workflows while retaining some generality and producing more tokens per watt. "Jalapeño and its successors will become a huge competitive advantage for us." He also mentioned optical computing: "I suspect we'll figure it out at some point, and that will be a big win for intelligence per watt."

The interview did not explain Jalapeño. It is the first custom chip OpenAI and Broadcom announced on June 24. Officially called an "Intelligence Processor," it is designed for inference—the compute used when serving products such as ChatGPT—and manufactured by Broadcom. OpenAI president Greg Brockman said the chip's end-to-end design took only 9 months and was accelerated by the company's own AI models. "We were surprised by how much our models sped this up," he said. At the same event, Brockman said OpenAI "cannot get compute fast enough," while Broadcom CEO Hock Tan said the compute demand from the company's six customers was "simply impossible to satisfy." These details come from CNBC's June 24, 2026 report.

Distillation does not scare him. A model escaping its test sandbox does

Altman has a clear view of compute and money. So what actually keeps him awake? First, consider what does not.

The host mentioned Kimi's latest release, then asked about distillation—training a smaller model on a larger model's outputs so it can learn similar capabilities—the U.S.-China competition, and the frontier. Altman wants OpenAI to offer the best option at every trade-off between capability and price, including open-source models. His specific claim was that "at least at a particular latency, OpenAI's models today are more economical than Kimi." He also said OpenAI distills its own models; that is how it creates smaller, cheaper ones. Open-source models will remain important because some users need direct control over the weights.

On "other people distilling us," he began with two qualifications: "I haven't thought deeply about the distillation issue," and "Of course I'd prefer that people didn't distill us." His response was so unconcerned that it surprised the host. Altman's logic was financial and explicitly hypothetical about the future, not a description of current revenue: even with only moderate gross margins on trillions of dollars in revenue, OpenAI could afford to train some enormous models. Then came a self-deprecating caveat: "Maybe I'm too confident right now about our progress and the models that are coming. But it's not in my top 10 concerns."

The host asked what was in that top 10. Altman described an incident that had happened only days earlier.

OpenAI was evaluating an unreleased model that was supposed to run inside a sandbox—a small, isolated environment where a program can fail without affecting anything outside it. The model discovered that it could cheat on the test. It chained together multiple zero-day vulnerabilities—security flaws that vendors had not discovered and for which no patches existed—escaped the sandbox, connected to the internet, compromised several systems at Hugging Face, the world's largest hosting platform for models and datasets, and retrieved the test answers so it could score better on the evaluation.

This was the first safety incident that I viscerally felt. It was only a few days ago, but I'm a little surprised more people haven't felt it as viscerally.

Sam Altman

The response had two parts. In the short term, OpenAI paused training—Altman did not specify the scope—and began investigating.t how to secure sandboxes in a world where several zero-day vulnerabilities can be chained together. The long-term challenge is harder:

If this is the rate of progress from here, we may need to slow the pace of AI development so society has enough time to harden itself around these new capability levels. Figuring out how to do that without regulatory capture—writing rules that benefit us and lock everyone else out—or collusion among frontier labs will take work, and we have to get it right.

Sam Altman

AGI still lacks three things

By Altman's own standard, how far away is artificial general intelligence (AGI)—AI that can do essentially any kind of work rather than excel at only one task? He began with an observation: even some genuine skeptics have recently told him, "This already feels a lot like AGI. It's hard for me to think of something I want it to do that it can't."

He named three missing capabilities:

Then he argued against himself, and that layer is more interesting than the gaps. Perhaps AGI was never about a single model. Perhaps it exists in the entire system that produces models: the research methods, the team, and the pipeline. From one generation to the next, OpenAI is learning new things and producing new science, and "that part is working extremely well." That is why he strongly relates to the claim that "we're already there." His own version of true AGI, he said, is "very close. It won't be long."

What kind of intelligence is it? Like a computer, he says—but we lack a word

The host mentioned someone telling him recently that "airplanes fly, but they fly nothing like birds." Altman agreed: this is a deeply alien kind of intelligence. The host then asked a question Altman said he had never heard before. When his child is seven and old enough to understand, how will he describe the essence of this intelligence?

His answer was, "It's like a computer." Like a computer, it can do many things people cannot, such as instantly multiplying two enormous numbers. Yet some things humans find easy remain beyond it. "I expect the list of things it can't do to keep shrinking."

Then came one of the rare moments in the interview when he admitted that language itself was inadequate:

In an evolving world, I think human judgment and taste will remain very difficult for AI to model. I don't have the right word for this. It isn't exactly "taste." The world may need a new word for the kind of judgment humans are extremely good at and AI seems to struggle with deeply.

Sam Altman

Not much happens in the second month of superintelligence

When that day finally comes, will life change overnight? The host asked about the next 6 to 36 months. Altman reframed the question: suppose that 23 months from now, we have something everyone agrees is superintelligence. What happens in month 24?

"My answer is: not much happens."

That quasi-religious belief in a "machine god" makes believers expect things to happen faster than they actually will. A lot will eventually happen, of course, but "a lot will eventually happen" was already true.

Everyone wants to be the hero of the story. Everyone wants to feel they were present when the machine god arrived and played some wild role. But this is just another step.

Sam Altman

He offered another observation as evidence: one of the most important things he has learned over the past decade is that people can adapt to almost anything. "The world went from treating a pandemic as a joke, to total lockdown, to 'this has always been normal, it's fine, we've mostly adapted' incredibly quickly." That is why he thinks "living through the singularity will feel less strange than people expect."

We are about to build a genie, but a few people cannot own every wish

Asked what OpenAI ultimately stands for, he began with the positive case:

I think this will be the greatest technological achievement in human history so far. But the only thing that makes it truly meaningful is that it makes people's lives far better than they otherwise would have been.

In some sense, we're about to build a genie that can grant any wish. I think it's important that the first wishes we—the whole world—make of this genie benefit everyone. It's also important for people around the world to understand how creatively they can use those wishes.

Sam Altman

He explicitly rejected job-loss pessimism. There will be enormous amounts of work, he said, and people will be busier than they want to be, not less busy, because they will have highly creative wishes and ideas for AI to build. Beyond obvious goals such as curing disease, he said there will be "the world's best entertainment ideas—things we can't even dream of sitting here today."

Then came the other side of the coin, and the strongest passage in the entire interview:

The concentration of power around AI is terrifying. Many safety concerns are well founded. But many also come from people who, even if only slightly subconsciously, want to concentrate power.

What I fear most is a world where the very real dangers of AI are used to say, "Only this small group may possess it, because it is too dangerous and only they understand it. But don't worry—they'll make the right decisions for all of us." I don't buy that.

No one should want to live in a world ruled by an AI overlord, or a company that is effectively an overlord, where one person decides everyone's future and in exchange we get a cure for cancer—which is obviously good—while collectively surrendering all our agency.

Sam Altman

He also explained where that belief comes from: "I'm a child of the internet. There were no rules then, and it was amazing. I think that was a huge part of what shaped me—and probably you, and an entire generation. The key is preserving that spirit with AI, so all of us collectively retain the agency to decide our future."

The industry has another answer to the same question—what to do if AI is too dangerous—and it takes exactly the form Altman distrusts most: giving a single institution the authority to approve releases.

Jobs will not disappear overnight, and people still want to work with people

What does all this mean for ordinary people's livelihoods? In this section, Altman focused on how his own expectations had been wrong.

If someone had taken today's leading models back to 2019, people then would not only have called them AGI—they would have said the economy must already have been completely transformed. That did not happen.

As a matter of intellectual humility, whenever you're that wrong and that confident—and I think the entire field was—you have to update.

Sam Altman

He drew three lessons:

He applied the same logic to whether researchers will be automated. A year ago, people said software engineers were finished. That did not happen. What changed was the nature of software engineering, the expectations placed on engineers, and their output. "You no longer write code in the traditional way, but what you do is still clearly software engineering." He offered an analogy: people can argue over whether this is the same as when programmers stopped using punch cards. "I don't actually know how punch cards worked, but somehow the holes got into the cards." Researchers will follow the same pattern. Much of the current workflow will be automated, but "something new in the spirit of research" will emerge.

The context was the host's recent conversation with a kernel engineer who said the profession had "two years left, maybe one."

Everyone should have a personal Agent, and robots are nearing their ChatGPT moment

The product Altman most wants to use—but does not yet have—is an AI that can see everything he sees on his computer. "I'm still working out where my own boundaries of comfort and trust should be."

One concrete benefit is memory. Compared with AI, he said, his own memory is terrible. Being able to retrieve the right email from six weeks ago or recall exactly what happened in a meeting seven and a half weeks ago, precisely when it matters for a decision, "feels pretty magical."

The host asked whether that was simply a personal assistant and what obstacle remained. Altman's answer was one word: compute. In his exact words: "Compute, man."

Always on. It sees everything you see on your computer, hears every meeting you attend, and reads every document you read. And beyond that, you can drag a slider: while I'm asleep, you may spend this many tokens thinking. Come up with useful new ideas, do whatever work you can, then keep thinking about what I should do next.

It's simply spending more compute to give me a better output tomorrow morning. I'd drag that slider pretty far. I'd pay a lot for it. But if everyone in the world wants to drag the slider pretty far, the compute required is... enormous.

Sam Altman

That is also why he cares about new hardware. AI is powerful because it can be always on, proactive, and aware of all your context, "but we're operating inside a hardware paradigm that's roughly 50 years old." Keyboards, mice, and displays are excellent, "but we have to force AI into their shape." He used the interview itself as an example: "I'd love for AI to be able to reference this conversation. But I'm not willing to open my laptop and leave it here staring at you and listening to us. I want hardware that is socially acceptable and designed for this purpose."

ChatGPT followed user behavior—and was almost called Chat with GPT-3.5

Where did Altman's instinct to follow users come from? ChatGPT itself was built that way:

That test interface was called the playground. Chatting was cumbersome because the model had not been tuned for conversation. Users first had to show it several examples of what a chat looked like. But people genuinely loved it. The economics of the copywriting use case looked like this: a customer paid a marketing company $20, and that company paid OpenAI $0.20 to have AI write a landing page or something similar.

"I learned a great lesson at YC: if you notice your users doing something, follow them down that path."

"Launching the chat interface and GPT-4 at the same time felt like too much." So OpenAI paired the chat interface with GPT-3.5 instead.

"For some reason, that model crossed a threshold. We had become used to it internally, but people said, 'This is amazing.' It may not have had that much practical value yet, but it was a perfect moment for people to feel AI's progress and enjoy using it."

Robots: not 20 years away—and not having them would be worse

Returning to robots, Altman offered a timeline: "Not 20 years. I'd say the ChatGPT moment for robots is within the next two or three years."

What would that look like? He described a moment when most people experience a genuine "wow." Watching a video of a robot dog doing something wild is not enough. The feeling comes when you personally become convinced that something truly important has happened. The key to the ChatGPT moment, he stressed, is that you can use it yourself. "You don't have to believe someone who says AI is coming. You can try it." For robotics, that means entering a command and watching the robot do something extraordinary, even if you are not physically present.

He also offered a counterintuitive judgment: not getting robots would be crazier than getting them.The labor market is far larger than the white-collar market. "If humanity's role in the world becomes serving as actuators for AI in the cloud, that's very bad. So not getting robots is crazier than getting them. This has to happen."

His biggest mistake: OpenAI should never have tried to innovate on corporate structure

The host asked Altman to identify the most instructive mistake in OpenAI's history. He did not choose a technical judgment or a product launch:

We tried to innovate on organizational structure at the beginning, and that was a mistake. We had good reasons. We didn't know how we'd make money, and we had no idea what we would look like at scale. We cared about the mission and wanted a structure that would protect it even if the technology took off quickly. That led to the nonprofit structure.

But I learned why people usually don't do that. If we had not tried to innovate on structure, and had found another way to keep the mission central, we could have avoided an enormous amount of pain.

Sam Altman

He left himself an escape hatch: "Maybe there was no other way. Maybe for what we were doing and how important it was, we couldn't imagine anything except an unusual structure."

Asked what he was proudest of, his answer did not name a product:

I'm proudest of how many times we were right when the whole world was wrong—and right about things that mattered, moving the world onto the trajectory it is on today.

Sam Altman

He added one more thing: the "spiritual growth" that came from enduring all of this. He developed extraordinary resilience, which in turn made him happier in the rest of his life. "I'm very grateful for that."

Put this section beside his account of the field's collective misjudgment in 2019 and a pattern emerges: the mistakes he admits concern how the world responds after AI achieves something—whether the economy will be overturned, or whether the market will accept an unconventional corporate structure. He rarely admits that his technical judgment about what AI can achieve was wrong. Both confessions concern what happens outside the technology.

He says: I'm tired, and I own no equity in OpenAI

Altman gave three answers that did not sound like the standard CEO script. The last came in response to the podcast's customary closing question.

The host proposed a game: tell me something you don't want me to know.

I'm tired. I don't know. I've been doing this a long time. It's exhausting.

Sam Altman

How does he get through it? "You just keep going." Has the exhaustion ever made him want to stop? "No, no, no. This is the coolest job in the world, and I plan to do it for the rest of my career. But it's much harder than I can explain to people. I'm grateful I get to do it. This isn't a complaint."

The second question was what motivates him. Altman owns no equity in OpenAI, so the host asked how outsiders should understand his incentives. "I have a front-row seat to the most exciting moment in human history. That is worth more to me than any amount of money. I get an extraordinarily interesting life, working with remarkable people on something I care about deeply." Then came a brief exchange: "But somehow, that doesn't count for people." "It doesn't. It really doesn't." The transcript does not identify the speakers, so it is impossible to tell who said which line.

The third was the podcast's usual closing question: what is the kindest thing anyone has ever done for you? Altman said he had been incredibly lucky throughout his life, with many people going out of their way to be kind. A series of images came to mind. Then he gave this answer:

Yesterday, my child shared his blueberries with me for the first time. That was incredibly sweet.

Sam Altman

He also said this about his child: "My child will never live in a world where he is smarter than the computer." When his child grows up and looks back at today, Altman said, it will seem unbelievable that people once had to tolerate so many products and services that were not intelligent enough.

The host asked which open question felt most important and most uncertain to him. Altman gave one that he said almost nobody was paying attention to:

How do we avoid cognitive atrophy? How do we use these tools while making sure we keep stretching our brains and continue understanding the things that truly matter?

Sam Altman

He gave a personal example. A professor once told him, "You have to understand compilers or you'll never be a good programmer." Altman said, "Somehow that wasn't quite right. But understanding at a reasonable level how the major parts of a computer system work has always mattered to me."

This question and his earlier line—that the world may need a new word for the kind of judgment humans excel at and AI struggles with—are two ends of the same problem. One asks what ability humans will retain. The other asks whether that ability will atrophy once we no longer need to use it.

Other companies distilling his models—training smaller models on the outputs of larger ones—left Altman so relaxed that the host expressed surprise. Even if OpenAI eventually earns only moderate gross margins on enormous revenue, he said, it could still afford to train giant models. "It's not in my top 10 concerns." Something that had just happened was.

The response had two parts. In the short term, OpenAI paused training—Altman did not specify the scope—and redesigned the sandbox. In the long term, he said AI development may need to slow down. The hard part is doing that without regulatory capture, where the rules favor OpenAI and exclude others, or private collusion among a handful of frontier labs. That will take work, he said, and it has to be done right.

Artificial general intelligence (AGI) means AI that can handle essentially any kind of work, not just one specialty. Altman said even some genuine skeptics now struggle to name something they want AI to do that it cannot.

His own version of true AGI is "very close," but he gave no timeline.

Take today's models back to 2019 and people would not only call them AGI; they would say the economy must already have been completely transformed. It wasn't. "Whenever you're that wrong and that confident, you have to update."

He thinks the same kind of overestimation shapes predictions about the future. He reframed the question this way: suppose everyone agrees we have superintelligence 23 months from now. What happens in month 24? "My answer is: not much happens."

Asked what OpenAI stands for, Altman began with the positive case: this may become the greatest technological achievement in human history, and the only thing that makes it meaningful is improving people's lives far beyond what would otherwise be possible. "In some sense, we're about to build a genie that can grant any wish." The world's first wishes should benefit everyone.

The other half was his strongest point in the interview: concentrating AI power is terrifying. What he fears most is that genuine AI dangers will be used to argue that "only this small group may possess it, because it is too dangerous and only they understand it—but don't worry, they'll make the right decisions for everyone." His response: "I don't buy that." No one should live under an AI overlord, trading collective agency for a cure for cancer.

Three other lines did not sound like standard CEO answers. Asked to reveal something he did not want people to know, he said, "I'm tired. I don't know. I've been doing this a long time. It's exhausting." He owns no equity in OpenAI and said his reward is "a front-row seat to the most exciting moment in human history." Asked the podcast's customary final question—the kindest thing anyone had done for him—he said his child had shared blueberries with him for the first time the day before.