Sam Altman admits his biggest mistake: going too conservative on compute and getting spooked by the financial markets; AGI will make humans even busier.

OpenAI CEO Sam Altman sat down with host Ti Morse for an interview on the podcast《Relentless》, compressing the company’s biggest bottleneck into two words: chips, then power. He admits he previously “placed far too little” on compute investments, because he was scared by the financial markets.
(Background recap: OpenAI’s compute spending is set to surge to $750B by 2030, but revenue is stuck around the $25 billion level)
(Background: FT broke the news that OpenAI delivered a knockout: a major ChatGPT overhaul introducing AI agents that can “do anything,” ending the era of pure chat conversations)

Table of contents

Toggle

• Transistors, then power
• The time he was scared by the financial markets
• The third wave is already on the way

OpenAI CEO Sam Altman recently appeared on the podcast《Relentless》for an interview with host Ti Morse. The entire conversation lasted about 70 minutes. Altman talked about entrepreneurship, ambition, and how painful it is to juggle being a father and working at the same time. He boiled the company’s biggest bottleneck down to two words: not enough chips, and then not enough power.

Transistors, then power

In the interview, Altman unusually laid out internal trade-offs. When GPT-3 started producing results, the company shut down the bot project it had been putting a lot of effort into at the time; more recently, as coding agents took off, what got shut down was Sora and the browser.

He emphasized that Sora wasn’t a failure. “It was supposed to be very successful—it's just that putting compute and effort into coding agents was more important.” It’s a decision a resource-constrained company would make: the product was loved by people, had momentum, and they poured huge amounts of money and a full year into it, yet they still had to stop. Altman said this kind of thing isn’t decided by holding a meeting and voting—it’s “slowly realizing that this compute, these people, have more important uses, and then making a very painful decision.”

At the same time, he believes the industry is “too focused on making algorithms generate better algorithms, but not enough on making data centers generate more data centers.” In simple terms: use the computing power of one data center to command a team of machines, building the next data center.

In his view, if the expectation about what compute will ultimately unlock is correct, the returns from doing this will be higher than using the same batch of compute for something else.

The time he was scared by the financial markets

The most direct self-criticism in the interview came when the host asked him: when was the last time he realized he wasn’t ambitious enough? Altman’s answer was compute investment: “I was seriously betting far too little on compute investment.”

When pressed on whether he could have known back then, he said yes. “But I was scared by things like the financial markets. That was obviously a mistake.”

He also gave a glimpse of OpenAI’s next form. After turning from a research lab into a product company, the company is moving toward large-scale infrastructure, where most of the value will fall on the infrastructure layer with relatively lower gross margins. As for how to take that step specifically, he said, “It’s not the time to talk about it yet,” adding just one line: he has already figured out how to align what he’s good at and passionate about with the company’s next 10x or 100x growth.

The third wave is already on the way

Altman split the growth of AI products so far into two waves: first chatbots, then coding agents—and the latter “is completely out of control right now.” He also believes the third wave will arrive soon: long-term, continuously operating agents—persistent agents—like a chief of staff or a coworker.

In this wave of coding agents, OpenAI is actually a fast follower. He said bluntly that at the time Codex was far behind Claude Code, and internally it was viewed as “a mission on the level of a kamikaze attack,” because the industry consensus was that once someone already had momentum in a category, hard-fighting to win was almost impossible. But the team built it. In his words, now most top engineers he knows use Codex.

As for where all of this is headed, he believes the real battleground at this stage is whether the world will move toward “AI authoritarianism” or freedom. He worries that someone will argue for leaving only a single model as the machine god, or having it controlled by a single company, using safety and economic shock as justification.

“Every time humans trade freedom for safety, in the long run it’s a net loss.” He described that he waited for this moment his whole life at one point: “We’re in the singularity right now.”

At the same time, he doesn’t think superintelligence (AGI) will put people into a “four-hour workweek” life. The reason is human nature: expectations will keep rising. We’ll always want more. We’ll invent new things for each other, and we care a lot about how we stack up against others. His conclusion is that after superintelligence, everyone will end up busier than they think they will be. “On the surface people will still complain, but in their hearts they’ll actually be happy.”