Sam Altman on Mistakes, Money, and What OpenAI Is Actually Trying to Do

Sam Altman agreed to speak with prompt/power on the condition that we would publish his responses in full, without paraphrase. This is that conversation, lightly edited for length and clarity.

The last eighteen months have been eventful. What did you get wrong?

Several things. I underestimated how quickly the competitive landscape would change — not just from the companies you’d expect, like Google and Anthropic, but from open-source. Llama 4 is a genuinely good model. A year ago I would have told you that the gap between frontier closed models and the best open-source alternatives would widen. It narrowed.

I also underestimated the enterprise sales cycle. We built GPT-4 and assumed that enterprise customers would come to us because we had the best product. Some did. Many didn’t, because enterprise procurement doesn’t work like consumer adoption. It requires trust, compliance frameworks, procurement relationships, and integration support. We’ve built that capability over the past year, but we were late.

OpenAI has made several unusual strategic moves — the capped-profit structure, the for-profit conversion discussions, the Microsoft relationship. How do you think about the tension between the mission and the business?

I think about it constantly. The honest answer is that the tension is real and it’s not fully resolved. Our mission is to ensure that artificial general intelligence benefits all of humanity. Our business requires capital at a scale that requires commercial success. Those two things are usually compatible and occasionally in tension.

The capped-profit structure was an attempt to institutionalise a constraint on the commercial imperative. Whether it’s the right structure — whether it’s the structure we’ll have in five years — I genuinely don’t know. What I’m confident about is that running OpenAI as a purely commercial operation would be wrong, and running it as a purely non-commercial operation would mean it ceases to exist. We’re navigating somewhere between those two failure modes.

What does OpenAI look like in ten years?

I think about this differently than I used to. I used to think about it in terms of products and capabilities — what can the models do, what applications have we built. Now I think about it in terms of the relationship between AI systems and human institutions. In ten years, AI will be deeply integrated into the institutions that organise human life: healthcare systems, educational systems, legal systems, economic systems. The question is whether that integration is healthy — whether it’s making those institutions more capable and more equitable, or whether it’s concentrating power and eroding accountability.

I want OpenAI to be on the right side of that. I don’t always know exactly what that means. I’m sure we’ll make more mistakes. I hope we make fewer important ones.