Speaker A
Hello. Good to see you. Thank you. So I saw what's happening in Abilene with my own eyes.
Sam Altman discusses OpenAI's Stargate project, AI compute demands, partnerships, and the future of AI infrastructure and humanoid robots.
Key Takeaways
- AI demand is growing faster than anticipated, requiring massive infrastructure expansion.
- Stargate is a multi-partner initiative to build compute capacity beyond what any single company can provide.
- OpenAI is confident in its ability to generate value and sustain profitability despite large investments.
- Managing viral user growth requires innovative compute resource allocation and trade-offs in features.
- The future of AI includes humanoid robots and significant global competition, especially from China.
What the video covers
- Sam Altman explains the evolution of compute needs at OpenAI, especially after GPT-4 and ChatGPT's launch.
- Stargate emerged as a solution to scale compute infrastructure beyond what existing partners like Microsoft could provide.
- Altman details global trips to understand supply chains and partnerships with SoftBank and Oracle for building massive compute capacity.
- The project aims to meet surging demand for AI inference, which currently outstrips available GPU resources.
- OpenAI projects needing $500 billion in capacity investment over the next few years to sustain growth and innovation.
- Altman expresses confidence in OpenAI’s financial sustainability despite unprecedented scale and challenges.
- He describes the viral growth of AI usage and how Stargate helps manage demand surges and feature limitations.
- Competition from other AI models like Elon Musk’s Grok is acknowledged, but OpenAI’s edge is its infrastructure.
- The discussion touches on the impact of AI data centers on jobs and the future arrival of humanoid robots.
- Altman reflects on AI’s rapid progress and the importance of infrastructure and partnerships in shaping the future.
Chapters
- 00:00Introduction and Early Compute Challenges
- 00:35Scaling Compute Post-GPT-4 and ChatGPT
- 01:17Global Trips and Partnerships Formation
- 01:59Role of Microsoft and Other Partners
- 03:19AI Infrastructure and Data Center Design
- 04:02Demand Constraints and Compute Fleet
- 04:39Funding Needs and $500 Billion Investment
- 05:26Financial Sustainability of OpenAI
- 06:07Handling Viral Growth and Compute Challenges
- 07:04Competition and Future AI Models
Full Transcript — Download SRT & Markdown
Speaker A
Isn't it cool? It's happening. Take me back to the beginning, before Stargate. What did you start seeing that made you realize we need more compute, more power, more connectivity, that you were hitting a limit and needed to scale?
Speaker A
We used to think a lot about the compute we would need to train the models.
Speaker A
And what we didn't used to think about as much was how much people were going to use these models.
Speaker A
We did think about it. It just turned out people want to use the models much more than we imagined.
Speaker A
And a couple of years ago, maybe after the launch of GPT-4 and ChatGPT, it really started to hit like, “Oh man, this has gone from a lot of compute to like the biggest infrastructure project in history.”
Speaker A
And we started thinking about ways to do it. And we started trying to understand where the limits in the supply chain were.
Speaker A
And out of that emerged Stargate. There were other smaller things that we did. First, we have, you know, worked with other partners.
Speaker A
We've worked with Microsoft to build out, like, very gigantic amounts of compute. But, you know, this is like the next step in that evolution.
Speaker A
Right, I mean, you traveled around the world talking to people about this. How did you get to talking with Masayoshi Son and Larry Ellison?
Speaker A
And how did you all come together? I’ve known Masa for a long time. In 2023... I did sort of two long trips.
Speaker A
I think that was the year. It’s all a blur. It’s all a blur. I did two kind of long trips around the world, and a lot of it was to talk to developers and governments.
Speaker A
But a lot of it was also to just really try to, like, get my head around the supply chain.
Speaker A
This is not something I've done before. Like I had not before thought about, like what it was going to take to get compute up and running at this scale.
Speaker A
And there are a lot of, like, hard pieces. And on one of those trips, I met with Masa.
Speaker A
Masa has, like, thought for a long time about chip fabs in particular, but really kind of the whole thing that it takes.
Speaker A
And we got to speaking about what it would, you know, what it would take to do compute at this kind of scale.
Speaker A
And then it took us a while to figure out. I mean, it is a complex supply chain with a lot of partners and obviously a lot of capital.
Speaker A
And, you know, SoftBank's obviously your financial partner. Oracle's a technical partner. Why couldn't you get what you needed from Microsoft?
Speaker A
I mean, we do get a lot of great stuff from Microsoft, but I think this is more than any one company can deliver. We have...
Speaker A
Microsoft will do a lot of compute for us, a lot, a lot. I'm very happy about that.
Speaker A
Why call it Stargate? It began as a codename and sometimes codenames stick. But it means something, right?
Speaker A
The design of one of the very early layouts of a data center looked a little bit like Stargate from the show.
Speaker A
A futuristic wormhole? Not like that, at all. You announce this at the White House the day after the inauguration.
Speaker A
How did this get to President Trump's desk? And what was that moment like? Well, the president is super interested in infrastructure, and he's made, you know, a big priority, which I think is wonderful, about permitting the energy production and the facilities
Speaker A
for the data centers. So that was kind of how the conversation started. We're in the middle of a step change for AI.
Speaker A
And clearly, what we saw just in the last week from you alone. You're seeing what's next on the roadmap.
Speaker A
How does that inform what the design of the data centers of the future need to be?
Speaker A
I think there's a lot of details we're learning about how to design these data centers, what you want, really at all levels from, like, the chip to the sort of architecture of the whole data center.
Speaker A
But the main thing that's been on my mind, and I think on many people's minds, is just how much inference demand there is.
Speaker A
We are crazily constrained. We have a gigantic compute fleet, like gigantic, gigantic, and yet still, if we had twice as much, we would be able to offer much better products and services.
Speaker A
There's no shortage of demand. So for me, there are all the technical lessons about what we've learned and how we want to build this, but mostly we just want a lot.
Speaker A
Explain the math to us. How does it all add up to needing $500 billion?
Speaker A
Because folks out there don't get the accounting. Who's putting in what, how much funding is really secured?
Speaker A
How does it add up in terms of like what do we need that for?
Speaker A
How does the funding come together? Like explain that number, 500 billion. Oh. Well, that covers the capacity we think we need for the next few years, given our growth projections.
Speaker A
The interesting question is, if we knew how to get $1 trillion right now, which we don’t,
Speaker A
would we be able to deploy that profitably in the next few years? And I'm not sure about that, but I feel confident we can, like, make $500 billion in value back.
Speaker A
Let's talk about that because you just raised even more money, generally. Into OpenAI. Right, into OpenAI.
Speaker A
How confident... it's a lot of money. Tens of billions of dollars. How confident are you that OpenAI is going to be a financially sustainable and profitable enough company to justify all of this investment?
Speaker A
It's looking like we're doing really well. I mean, like, we have to, we are definitely doing something unprecedented. But, you know, it seems like a... I like, I feel confident in the bet.
Speaker A
It doesn't mean something can't go wrong. You tweeted that the GPUs were melting. Yeah, I didn't realize people were going to take that literally.
Speaker A
I mean, I get that it was a joke, but it’s a metaphor. But like the metal is not.
Speaker A
GPU shortage, bro. That the team can't sleep. You're so busy. What happens when a launch goes viral?
Speaker A
And how does Stargate solve for this? Well, first of all, this level of viral, right?
Speaker A
This is an unusual thing. This last week. I don't think this has happened in the history of tech to any company before. Like the...
Speaker A
I don't know of any. I've seen viral moments, but I have never seen anyone have to deal with an influx of usage like this.
Speaker A
You added a million users in like an hour, or something like that? Yeah, I mean more some hours, but yeah, it was like a... it was like unprecedented, wild.
Speaker A
And also like making an image is not a... it's not exactly like a low compute task the way we do it with the new image gen, so we had to do a lot of very unnatural things.
Speaker A
We had to, you know, borrow compute capacity from research. We had to slow down some other features because it's not like we have like hundreds of thousands of GPUs sitting around just like spinning idly. So with if we had more
Speaker A
GPUs, we would be more able to handle demand surges like this. And also we wouldn't have to put such restrictions on...
Speaker A
I was just this morning making the list of what we have, what we'd like to launch in the next few months.
Speaker A
And I was like, you know, I had the list over here, and I had like my rough math of when we're getting GPUs here and where I thought we could get the efficiency gains.
Speaker A
And I was like, this is not going to work. Like, we're going to have to, like, pull some of these things, make some trade-offs, limit some things...
Speaker A
But this is not all of this is not going to fit here. So it's that directly linked?
Speaker A
Yeah, I mean you can find efficiency gains and we tried to do... you can limit features and usage and we do that too.
Speaker A
But you know, more compute means we can give you more AI. And you can use it for images or writing software or whatever you want.
Speaker A
Elon's got his own data centers, and xAI just bought X, which means he's got access to all of that additional data.
Speaker A
He claims Grok is the smartest AI on Earth. How do you think it compares?
Speaker A
I don't use it much. I think it's probably a very good model. I mean, obviously there are lots of competitors out there.
Speaker A
There's a lot of good models. I think good models will become very bountiful, very plentiful.
Speaker A
So what's your edge going to be? The best infrastructure layer.
Speaker A
You know, we have way more people use ChatGPT than use any other, AI service. Way more, and I think you'll see with a lot of the new features roll out that that will become, more of an advantage over time.
Speaker A
So once this is all built, what's the grand vision? Give people tools to let them do whatever they're going to do better.
Speaker A
You know, it's this is in some sense, this is like it is different this time, but in many other senses, this is just like another tool and another piece of technological history.
Speaker A
People will use it and unleash their creative energy and make all sorts of stuff that you and I love, or some stuff that we don't, or stuff to just entertain themselves.
Speaker A
Personally, the area I'm most excited about is AI for scientific discovery. I think that will you know, we're not there yet, but we're not far away.
Speaker A
I think 2025 will be a world where we have agents do a lot of work, but work of the kind of work things we already know how to do.
Speaker A
I'm hopeful that 2026 will be a big year of New scientific progress. Do you envision multiple Stargates on every continent?
Speaker A
Like, where else are we going to see you break ground? You will see them on other continents.
Speaker A
Everywhere? I don't know about everywhere. You went to visit one. I did. These are hard things to do.
Speaker A
Well look this is an audacious bet on the future. What are the risks? I mean, maybe, you know, people stop wanting to...
Speaker A
pay for AI services, and then we have a difficult financial position. How do you feel about one chipmaker Nvidia having so much power over the industry?
Speaker A
The future of the industry? They make an incredible product. And if you make an incredible product and people want it this is what happens to you.
Speaker A
I do want to talk about the jobs thing because obviously we saw so many people working there.
Speaker A
There is this lofty promise that AI data centers are going to create thousands and thousands of jobs.
Speaker A
Meantime, AI is destroying jobs elsewhere. And I feel like even, you know, there is serious anxiety out there.
Speaker A
People are scared. Totally. Even among the best engineers and technology, people are scared. What do you say to those people?
Speaker A
AI is for sure going to change a lot of jobs. Totally take some jobs away, create a bunch of new ones.
Speaker A
This is like the kind of this is what happens with technology. And in fact, I think if you look at the history of...
Speaker A
the world, technological driven job change or whatever you call it, when like one class of jobs go away and another one pop up, That's very consistent, it happens at a...
Speaker A
...it's punctuated. But like that's just been happening for a long time. And the thing that is different this time is just the rate with which it looks like it will happen.
Speaker A
The thing I think the world is not ready for like people have maybe abstractly thought “Okay, it's going to be a better programmer than me.
Speaker A
It's going to be, you know, better at customer support and whatever.” I don’t think the world has really had the humanoid robots moment yet, and I don't think that's very far away from like a visceral like, “Oh man, this is going to do
Speaker A
a lot of things that people used to do.” So yeah, it's coming. And we we have always tried to just be super honest about what we think the impact may be.
Speaker A
Realizing that we’ll be wrong on a lot of details. What happens when the humanoid robots get here?
Speaker A
I mean, obviously do a lot of jobs, but I the point I was trying to make is, I think, like the first day you're like, walking down the street and there's like seven robots that walk past you doing things or whatever.
Speaker A
It's going to feel very sci fi. DeepSeek appears to have found a more efficient way to power AI.
Speaker A
Was that a moment of rethink for you? And are you doing anything differently now?
Speaker A
I think the DeepSeek team is very talented and did a lot of good things.
Speaker A
I don't think they figured out something like way more efficient than what we figured out.
Speaker A
But do you think there is a more efficient way? Oh I'm sure, yeah, I'm sure. We will We have made incredible efficiency strides year over year, and I'm sure we'll keep doing that in the future.
Speaker A
So if that's the case, why are you building all this? If we had an AI that we could offer at one tenth of the price of current AI, I think people would use it 20 times as much and we would still need
Speaker A
twice as much compute to satisfy the then current demand. So let's talk about that, Jevons Paradox.
Speaker A
Like how do you think Jevons paradox applies here, that, you know, technological progress means paradoxically you're you're going to be using more resources.
Speaker A
It’s never seemed like a paradox to me at all. Like, you know, we talk about supply demand curves and elasticity and all sorts of other things.
Speaker A
And then people are like, “ah, but I'm going to trick you with Jevons Paradox.” This is just the way the world works most of the time.
Speaker A
But you do think there will be more efficient ways? What are those? Where is it better chips?
Speaker A
Is it.... All of the above. We will have better chips. We will have better energy sources.
Speaker A
We'll have better algorithms. We will optimize. I mean, this is just this is like what industry does.
Speaker A
We'll optimize everything. There's no question that China is going to be a formidable player.
Speaker A
Sure. In AI, How do you think about OpenAI's chances to win a global race?
Speaker A
You know, probably I should have some deep thoughts about that. We're doing the best we can.
Speaker A
Like we just wake up every day thinking how we can be a little bit better.
Speaker A
I don't know how to, like. I don't know how our actions would change based off of, like, some deep answer to that question.
Speaker A
President Trump, President Trump, you said he cares about infrastructure, you said he cares about infrastructure, President Trump is in power at a pivotal time for AI development.
Speaker A
What do you think his mark on this moment will be? I think he will get to make some of the most important decisions anyone in the world has gotten to make related to AI.
Speaker A
And, you know, I'm optimistic he's really, he’ll really do the right thing there. But I don't like... he has an unenviable job, Right.
Speaker A
The velocity of invention. We were talking about a launch. It's the velocity of innovation here is just mind blowing.
Speaker A
The speed at which you're releasing new things, the speed of what's coming up. What I saw rising from the red dirt in Abilene.
Speaker A
How do you personally grapple with the the pace of it all? That's one thing that's really impressed me about President Trump by the way, is his ability to just like.
Speaker A
understand the whole industry and all the changes and quickly seem to have a very good intuition and make good decisions about it.
Speaker A
While things are changing so fast, has has really been quite impressive to me. But what about you personally?
Speaker A
Yeah, yeah, I was going to say I... I'm in it every day and so I don't like, It's like watching your own kid grow day to day.
Speaker A
You just see every change. And so it's, like, not as striking. It does feel like it's going very fast.
Speaker A
It’s certainly... Certainly think if I could, like, transport myself back three years ago, it would seem like unimaginable progress.
Speaker A
But, you know, day to day you can kind of get used to anything. Well, since you mentioned kids, you just had a baby.
Speaker A
Has that reframed anything for you? I mean, like, look, I don't think I have anything non cliche to say here, but it is the best, most amazing thing ever.
Speaker A
And it totally rewired all of my priorities. I remember in like the first hour I like felt this neurochemical change and it happened so fast.
Speaker A
I was like, oh, I get to like observe this. Like I am being like neuro chemically hacked.
Speaker A
But I'm like noticing it happening. I'm totally fine with it. That's great. But like everything is going to be different now.
Speaker A
What about how everything that you're doing here and building all this, what it means for humanity?
Speaker A
Has it reframed any of that? A lot of people have said, I'm very happy you're having a kid, because I think you'll make better I think you'll make, like, better decisions to whatever degree you have...
Speaker A
you got to, like, make decisions here for humanity as a whole. I really wanted to get it right before, and do the best I could.
Speaker A
I still really want to now. It's somehow. Yes, it's somehow it does feel like it's different, but I can't articulate how.
Speaker A
but I can't articulate how. As you keep moving forward, As you keep moving forward, if ChatGPT were to discover a Stargate, a futuristic wormhole that we could all travel through.
Speaker A
What's on the other side? Like, what do you see that the rest of us don't?
Speaker A
If ChatGPT could, like, transport us to the future? Yeah. Through a Stargate, a wormhole.
Speaker A
Please humor me wormhole metaphor... I have no idea. I mean, no, not. No one knows, right?
Speaker A
Like it's gonna... It's going to discover, I believe AI will help dramatically, like unbelievably, dramatically accelerate the pace of science and human discovery and our understanding of what's possible.
Speaker A
But like, do I think I could have sat here in 1905 and told you what we were about to discover in physics, and that 40 years later we would, like have an atomic bomb?
Speaker A
Definitely not. I and I think I am way too self-aware of my own limitations to sit here and try to say I can like tell you what's on the other side of that wormhole.
Speaker A
I have no idea. But, net net good, yes? Happy? Better? A better world? Up and down.
Speaker A
Not better in every way. But yeah, I think. Up into the right? Yeah, up and to the right.
Speaker A
With some choppiness.
Topics:Sam AltmanOpenAIStargateAI infrastructureGPT-4ChatGPTcompute scalinghumanoid robotsAI investmentAI competition
