Sam Altman on building OpenAI: from research lab to global intelligence utility
OpenAI inverted the normal startup playbook—most companies bolt on a research lab late; OpenAI started as one and had to build a product company on the fly. ChatGPT was launched as a low-expectation research demo to nudge developers toward chat-style API products. When it exploded, Altman's internal reaction was: "This is the good kind of emergency—we have to build a company and a product all at once." The core lesson he draws is that when something grows fast even in a rough state, you have a near-guaranteed hit.
Altman's consistent investing framework is emergent properties at scale: YC's batch-network effects, GPT scaling laws, and now agentic coding. He argues humans are cognitively ill-equipped to reason about exponentials, which is why skeptics have repeatedly misjudged how far LLM scaling would go. He points to an OpenAI model disproving an Erdős conjecture—something experts recently deemed impossible—as live evidence that LLMs can generate genuinely new knowledge.
His current strategic thesis is that AI is becoming a new utility layer—like electricity or the internet—and that the right analogy for selling it isn't "intelligence" (too abstract) but concrete outcomes, the way early electric companies sold "light at night" rather than kilowatt-hours. He flags the inference stack as critically underinvested relative to training, and predicts all frontier labs will need to become inference companies.
The two concerns Altman treats most seriously are: - Education atrophy: 3.5 years post-ChatGPT, he sees almost no systemic change in how students are taught or evaluated, which he believes will hollow out critical thinking at scale. - Power concentration: He estimates an ~80% chance AI development follows a broadly democratic distribution path, but calls the alternative—a handful of companies controlling a significant fraction of global wealth—intolerable, explicitly including OpenAI in that warning.
On the economic transition, Altman has softened his near-term jobs-doom view but is sharply focused on compute as the next scarce utility. His preferred policy response is ownership stakes over cash transfers—some form of citizen wealth fund that gives people a slice of the capital gains as leverage shifts from labor to capital.
Sam Altman traces OpenAI's unlikely path from research lab to the accidental launch of ChatGPT, argues AI is becoming a foundational utility like electricity, and warns that concentrated AI power—even in OpenAI's hands—would be catastrophic. He's broadly optimistic about outcomes but flags education, compute access, and wealth distribution as the defining near-term risks.
when something really starts growing and it's not very good, you have like a guaranteed hit on your hands
what they started marketing, selling to people was light at nights. You know, we are going to what you are getting from us is not electricity, it's light at night
As leverage in the world shifts from labor to capital, we should find a way to have something like a citizen's wealth fund in the country or in the world eventually, where people basically own a slice of capitalism.
everything about starting a startup has changed so much
OpenAI was like the strangest startup of the last maybe a couple of decades in Silicon Valley cuz it started as a research lab
with like an affordable amount of spend on tokens you can do what a hundred person incredibly great engineering team would do as a startup
there exists something today that just wasn't possible at all pre like automated coding era that is totally and non-obvious that will be you know, uh multi-trillion dollar market soon and that only four companies are working on right now
a huge part of the magic of what made YC work were uh was the sort of the network effects inside of the batch. And that was an emergent property at scale that just hadn't been discovered before
we did not evolve to be good at thinking about exponentials. People have a hard time imagining that scaling laws are going to continue exponentially
we launched in like I I don't know, something in the summer of 2020 the GPT-3 API
really the only business that people got to work in a significant way with GPT-3 was copywriting
more people were using they couldn't get the API to work for their business, but they were using their API key to just chat
we actually had GPT-4 done, but we had a new model we were ready to release in between called 3.5
our kind of internal belief at the time was that coding was how these models would control things on computers and robots were how these models would control things in the physical world
Codex got really good by early this year, but with 5.5 is when we saw this real inflection point where people are now like doing just incredible things with it
I think we are going to get like with the current pipeline, the current architectures, I think we're going to get over the line of when AIs can do incredible incredible work
the electricity companies, at least the ones I could find information about, they didn't talk about selling electricity cuz no one knew what that was or why they wanted it
As we move into a world where we all just have like this constant agent running for us, being useful to us all of the time, you may think about it as even one level up from tokens.
There are a lot of very smart people working on great training ideas, and we're going to have incredible models pretty quickly regardless.
We have not invested enough in being able to deliver at scale huge amounts of cheap intelligence, so the inference part of the stack is underinvested in.
All of the frontier labs are going to have to become inference companies to a significant degree.
Yesterday we had one of our models disprove a conjecture, one of the Erdős problems that smart people had worked on for a long time.
A lot of smart scientists had even quite recently said that disproving an Erdős conjecture was not going to happen, and then the model just did it.
LLMs are capable of figuring out new knowledge and are capable of doing some intelligence tasks that humans just can't do.
The field was honestly held back by a generation of scientists who just were way too certain on what scaling was not going to produce.
Betting against LLM scaling at this point feels quite misguided to me.
I struggle to point to any significant systemic change that I've seen in the education system at large in the 3 and a half years since ChatGPT launched.
If we don't change how we teach, learn, or evaluate, there will be significant atrophy in people's critical thinking skills.
AI is just going to keep going, and if AI progress continues on the exponential that it's on for another 3 and a half years, what society is capable of will be completely different.
There is a real alignment failure in a very fragile world, and the best way to get to a world everybody winning with everybody's values represented is to push this technology out into the world.
I think it's like 80% we end up on the democratic path for AI distribution.
I've become much less of a even short-term jobs doomer; this may not even be as directly disrupted as I originally thought in the short term.
If the price of compute from a supply and demand perspective gets way out of whack, then there will be a very interesting fork about what it means to equitably distribute compute.
H100 prices and Blackwell prices—the spreads between long-term reservations and spot—are like 5x.
There's a gigantic computer shortage.
If we can make models sufficiently smart at a sufficiently low cost, demand for compute is kind of uncapped, so in some sense as long as we can continue to make progress on this there will be a shortage forever.
I am much more excited about people having some sort of ownership stake than a fixed monthly cash dividend.
if I can think of a really great startup idea, if it's like obvious enough to me, then it's probably obvious to a lot of people
stuff breaks uh at an accelerating rate and in an unpredictable way as you scale it
if we're going to get all this compute, why do we put it all into this one project? We're not going to run into something. Why not divide it up among all these all these projects?
even if we're totally right and intelligence is going to become this new utility that every company, every customer, every government just needs access to...I kind of don't think, at least right now, the right way for us to analogize that is we're selling intelligence cuz people are just like somehow not resonating
If we continue to teach and evaluate students as if we were in a pre-AGI world, it is going to lead to atrophy of learning how to think.
There is an attractor state where AI technology gets concentrated to a few companies, and they become a significant fraction of the wealth on Earth, which would obviously be terrible.
We are seeing compute shortages now, and I can imagine compute being like the most important utility that people need.
the normal course of startups is that you start a product company and then it like grows for a while and then growth slows down and then you start a research lab and you like bolt that on and you try to figure out the next thing to do
all of the most interesting ones uh have had something to do with emergent properties that scale, or scale continuing to provide returns far beyond what the consensus thinks will work
when you find a time that you can push something to a scale people have not tried before and it's already working in some interesting way at the smaller scale, more often than not, that seems to be a good idea
what is happening is we are we are in the process of creating a new utility. This doesn't happen very often. You know, electricity is a utility, internet's a utility. There's water, I guess. There's not a lot of these.
As a consumer or business, you will think in something closer to tokens or probably even one level up from tokens, and will not care about where the hardware is or what particular chip powers it.
we said, well, we'll build a chatbot around it. And we put that out and we still didn't think it was going to do that well. Uh there was it was really meant as like a research demo uh to convince other people that they should build chat-like products and pay us for the API
we set this goal that by September of this year we will use 500,000 A100 equivalent GPUs, like a lot of computing power as an AI research intern. And by March of 2028 that we will have a full end-to-end very talented researcher like figuring out completely new architectures
I think if we're going to become a new utility, we need to find a way to explain to the world what it means to have this like intelligence pipe that you can just do whatever you'd like with
“This is an emergency. This is the good kind of emergency, but we have to build a company and a product all at once.”
“betting against LLM scaling at this point uh feels quite misguided to me”
“the risk of keeping this concentrated in a handful of companies even though we would be one of these companies is not something we should tolerate”
Shared with Earmark · earmark-ai.com