Let's talk about OpenAI. You have a really good relationship with OpenAI.
In December, we signed a huge deal with them. Well over $20 billion.
Andrew Feldman is the co-founder and CEO of Cerebras Systems, one of the leading companies in the development of artificial intelligence chips. Throughout his career, he has built relationships with some of the biggest names in the tech industry, including Sam Altman, Greg Brockman, and Elon Musk.
We solved a problem that had been open in the chip-making industry for 75 years. We made one the size of a dinner plate. There are lots of places where you don't want a big chip. You don't want it in a phone. You don't want it in a laptop. But for AI, big chips process information more quickly and delivered answers in less time. And by being big, we could use a different type of memory called SRAM. And that SRAM is thousands of times faster than what the GPUs use. Now, why this is important is because inference, which is how we use AI, inference requires us to move all the weights, a huge amount of data from memory to computer. And here you are with the memory for the GPU, which is slow at exactly that problem. It moves data very slowly.
And then for people to understand like the amount of difference, we just heard Sam Altman saying that the new GPT 5.6 it can get to 750 tokens per second. What would be the normal in a GPU?
So it's like 15 times bigger.
Well, right now we have a backlog of more than $24 billion of orders.
Not bad. We have seen extraordinary demand for our products. One of the things Sam has been really a visionary on. OpenAI sees this curve. They see the shape of the curve that it's an exponential and they weren't afraid to play that out a year or two or three. If this keeps going, we're going to need a lot more compute. We're going to need more data center. So they did Stargate and they made deals with Nvidia and with us and with others.
What does it take to make a data center? Because we've seen that some of them take forever, but then Elon built in like months something that was supposed to take years. So...
Elon is the greatest builder of stuff, of physical things I think the world's ever seen. He does things at a rate and with a creativity that others really can't match. You probably shouldn't say Elon can do it in 9 months.
What is Nvidia going to do about because what you're doing... and thank you for making time for us today. It's incredible what you have built with this company. But before we get into what's your opinion about the situation with AI and everything, I would like you to explain to me and to my audience in the most simple terms that you can what is it that Cerebras does that is so special that it makes it to be so fast. I mean it's completely different than other companies, right?
It is. We solved a problem that had been open in the chip industry for 75 years, for the entire history of the making of semiconductors. They were never bigger than your thumbnail. And we made one the size of a dinner plate. And there are lots of places where you don't want a big chip. You don't want it in a phone. You don't want it in a laptop. But for AI, big chips process information more quickly and delivered answers in less time.
Because they don't need to transfer between separate chips.
Exactly. Right. And by being big, we could use a different type of memory called SRAM. And that SRAM is thousands of times faster than what the GPUs use, which is called HBM type of DRAM.
So for what things is that good?
So there are two types of memory. One called DRAM. And sometimes some of it's called HBM that can store a lot but it's very slow to move it. And SRAM is very fast but per square millimeter can't store very much.
Okay. And so traditionally the way GPUs were made was they used this HBM and they had a chip that sat next to the GPU and that was your memory.
And it could store a lot but it couldn't move data quickly onto the GPU. Now why this is important is because inference, which is how we use AI, right? I mean a simple way to explain to your parents is you make AI with training and you use it with inference. Inference requires us to move all the weights, a huge amount of data from memory to compute, and then we do one calculation to generate a word, and then we have to do it again. So this workload called inference, it is extremely heavy in the moving of data and relatively light in the calculation. And so the proportion of time taken to deliver a solution is dominated by the movement of data. And here you are with the memory for the GPU which is slow at exactly that problem. It moves data very slowly.
And then for people to understand like the amount of difference, we just heard Sam Altman saying that the new GPT 5.6 it can get to 750 tokens per second. What would be the normal in a GPU?
60. So it's like 15 times bigger.
Wow. And people doesn't really grasp the amount of difference it makes to have faster AI because people thinks like okay I can wait 20 seconds for an answer right but the reality is that when you mix it with things like vibe coding this can enable like fluid interfaces that create itself on real time right
That's exactly right. I think what's happened with AI is it's become really useful and we've begun to use it not occasionally but regularly. And when you use something regularly or constantly like in vibe coding, speed matters a great deal. I mean, how long will you wait for a website to resolve? Right. How come there's no market for slow search or how big is the market for dial-up internet? Right. And once we use it all the time, whether it's search or whether it's the internet access or whether it's AI, fast is the product, it becomes interactive. You use it more often. You stay longer. You use it for more problems. All of which is driven by speed. And you know, we're 15, 18, 20 times faster than anybody else at this work.
And it's not just that it's faster for the same things. Is that I think and that's to me is the most important part is that it opens up new ways of using the AI that were not possible before like we were saying like I remember Anthropic presented something for it made like a small demonstration of something that was creating software on the go so you would click on a window and it would create the next window on the fly and that's only possible if you have the speed.
That's right. Let's... I mean when I was growing up in the US, Netflix delivered DVDs in envelopes. Do you remember? Right. You used to, right? You'd get an envelope in the mail and when you were done watching the DVD, you put it in an envelope and you mailed it back to them. Right. When the internet got fast, they became a movie studio. Right? Fast internet opened up an entirely different opportunity. It didn't make them faster at delivering DVDs. It said you can bypass that whole thing. Bypass the store. And once you do that, you can become a content maker. You can become a movie studio. And speed does that. It opens up whole new things that couldn't previously be done.
But let's get real here because if it was everything so beautiful, then we will only use like fast chips. But there must be a downside. There must be something that they don't do so well.
Well, right now we have a backlog of more than $24 billion of orders.
Oh, wow. Congratulations.
Not bad. Right. People are trying and so you know we will... we have seen extraordinary demand for our products.
And is there any parts like maybe in the training where like these kind of chips don't work as well or they can't train as well?
They can train as well. Training is a different problem and has different characteristics and we use a different dimension of our advantage for training. But right now the whole industry is constrained by data centers, right? And we're building data centers and renting data centers and leasing data centers as fast as we can. And that's what everybody's focused on this minute because we're trying to move at the speed of AI and data centers are moving at the speed of real estate.
Yeah. And that is the next part which I think is really important. People when they think about AI, they normally think about software like the software layer, the agentic layer, whatever you want to call it. But the reality is that the real game is actually in the infrastructure.
The real game is in the infrastructure.
Exactly. And then you guys are building chips that allow maybe as well to distribute like AI in a better way because it's being faster. It means you can serve more customers and is it more expensive than normal GPUs for the amount of tokens?
Just to build on your point, right, the AI made Nvidia the most valuable company in the world. I mean if you think about its impact on infrastructure and it's made Broadcom one of the most valuable companies in the world. It's made all these hardware players and now the demand for memory is making the memory companies among the most valuable companies and it's extraordinary. Right, this push is having these enormous ramifications in infrastructure. Yeah. And based on that we went public. We carry a nice valuation and so infrastructure has been clearly a winner so far in the AI revolution.
Yeah. There's like the data point that I got not long ago that in America now in the United States there is more money spent in data centers than in offices. So...
It has become... it is unbelievable. Unbelievable.
Do you think it's going to keep on going like that? Like we're having lots of demand. I always put this example like most of people I think someone made a calculation in Spain the average per person is about a million tokens a month which is not much to be honest. I spend about six billion tokens in the last three months with my agents with OpenClaw and stuff. So when everyone is using AI like I'm using it, the amount of demand we're going to need for these things that...
You know I made that point on our earnings call that right now the demand for compute overwhelms the supply. It overwhelms the supply for our chips. It overwhelms the supply for Nvidia chips and it overwhelms the supply for data centers. And only a tiny fraction, less than 1% of the world's heavy users right now, internet has about a 40% penetration. I mean what happens when AI has a 3% penetration or a 5% penetration, right? It will be extraordinary demand for compute dwarfing what we have today.
And you think it's going to keep on growing like...
As long as we keep solving hard problems.
Okay. And what do you think when people talks about that there is an AI bubble that Nvidia is overpriced because it feels like they have the orders they have people the same as you like you have lots of like demand. So...
Nvidia is an extraordinary company. I think you bet against them at your own risk. They've proven again and again over the past decade that they're one of the great companies in history. That's an extraordinary run. Right. 2013 they were probably 20 billion in market cap and now they're four or five trillion. I mean what an extraordinary run. Historically bubbles were in infrastructure when people said if we build it they will come. Like in the late 90s we built out too much fiber in the networking world and there wasn't demand for it. Six or eight years later the demand caught up and it made sense. But at the time they were building ahead of demand, right? In AI right now, we're all building behind demand. We're trying to catch up to what people want, not the other way around. That doesn't feel like a bubble, right? It doesn't feel like a bubble when users want to use something and they want to buy more of it and there isn't enough to sell them.
Yeah, it's super crazy that if you try to rent a GPU from like four years ago, three years ago, it's more expensive than four years ago. How can that be? It can be because there's so much demand and supply isn't moving at the right rate. And you know Sam Altman and he said not long ago he made like a case that OpenAI started selling compute in advance like three years in advance because he said that we will go into a phase where there will be not enough compute for everybody and do you share that opinion do you think we're going to get into a place where demand is going to grow but we are taking time to build data centers?
For sure. One of the things Sam has been really a visionary on is OpenAI sees this curve. They see the shape of the curve that it's an exponential. And they weren't afraid to play that out a year or two or three and to say, 'Wow, if this keeps going, we're going to need a lot more compute. We're going to need more data center.' So they did Stargate and they made deals with Nvidia and with us and with others and they made deals with the memory companies, right? And they weren't caught by surprise. They saw this coming and I think you should listen to them when they talk about sort of the way this will unfold over the next several years because they have the best data.
Here because many companies they saw this as a risky move. Actually I heard Dario from Anthropic say many times that...
And then they got caught with not enough capacity and had to rent it at a very high price from Elon.
Yeah. Right. Very high price.
That was a very good business from Elon. Right.
That I mean what a brilliant businessman he is.
Mhm. Yeah. But that was the point. No. So basically OpenAI they overinvested not overinvested in the sense of not using it but at the time when no one was thinking we'll need so much they made commitments and people thought it was crazy and that was very risky and then demand catch up and we still in a very early stage.
We are still in the very early stage.
So it looks like we're not going to be able to build enough of these things.
We're not going to be able to build enough data centers, certainly enough memory, enough data centers for the next 18 months to two years for sure.
So for the people that is listening to us, what does it take to make a data center because we've seen that some of them take forever but then Elon built in like months something that was supposed to take years. So first...
Elon is the greatest builder of stuff, of physical things I think the world's ever seen. I agree. He does things at a rate and with a creativity that others really can't match. So you probably shouldn't say Elon can do it in 9 months. So, first you need a big piece of land and you need power either from the grid or you need it from a source like gas, which is what he's using. You need to build a building. You need if you use grid power, you need power switches and regulators and you need permits from the local community so they don't go crazy and get angry at you and...
Which is happening. You need chillers and pumps and all these things that have lead times. They all have to be brought together and you're building multiple buildings the size of soccer pitches and you're putting enough power into them to power a small city and you're building many of them at once, right? This is what it takes. And if you're a good builder from scratch, it might take two years.
Wow. Yeah. And how much of that brings to the table compared to what we have nowadays like in the sense I don't know...
I think in the US there's tens of gigawatts being contracted for and people are trying to build out. 100 megawatts is if I remember correctly enough for a city of 100,000. So cities of tens of millions of people's worth of power is being built out in data centers.
But proportionally to what we have nowadays like...
I don't know exactly how much we have to...
Because what I think is like as I said before people is using like a million tokens I'm using billions of tokens there's a difference of like 2,000 times so especially with agents we have seen a massive explosion of token consumption like from 100 to 400 times more usage and that was from Microsoft and everything which brought us all the problems comes with a spike of pricing. But at the end of the day, it's like we need a ton more.
We need a ton more power and I don't know if we're going to have all the power in the US. I mean, it took 70 years for us to build it in, right? And the grid is old and it's out of date and to build new power plants is challenging. It takes time. It takes time and it takes a lot of time to work through the process in the US. I don't know how it is in Spain or in other but my guess is it's not going to be more efficient in Europe.
Definitely Europe permits are more complex more complicated. China has done a really good job of improving their grid, bringing more power on using hydro and using solar and building new grids and they saw this coming and that's been something worth watching. I'm loving this because as you're talking I'm getting like so many things I want to ask you because I want to pick your brain on these topics that has always been like kind of like a blur for me because I don't know but I don't want to forget any so I take notes. Then basically the electricity is going to be the biggest bottleneck.
Electricity is the biggest bottleneck. Electricity for us and for Nvidia and for AMD and is delivered through a data center. Right. A data center is just like a giant hotel for compute, right? It needs power. It needs cooling. And so you build this giant space, you put all your compute there, you feed it power, you cool it and you send it data and you send back answers.
So basically we are trading electricity for intelligence.
Intelligence. That is what we do in AI at the fundamental level is we turn electrons into intelligence.
And that's what basically your...
That's what our chip does.
Exactly and you don't do only the chip right do you do the...
We don't. We later today we can show you around. We do the chip and we do the system and we deliver the rack and we even build out our own data center so you can get to us via the cloud.
So when we see these big investments from Elon we're going to spend that much billions on a data center. How much of that goes to companies like yours and Nvidia or...
What Nvidia published a report saying a gigawatt is about $40 billion worth of their hardware. And that's a lot. 40 billion. And a data center right now in the US costs somewhere between 10 and 14 million a megawatt to build. So a gigawatt is a thousand megawatts. So a thousand times 10 million, 15 million, that's what it costs to build a gigawatt.
A lot. But most of it goes to the provider of the chip and the...
That's right. Some an important amount goes to chip. The building lasts 20 or 30 years, right? The chips six or seven.
Six or seven years. Okay. That's been one of the biggest... it's up for debate.
Okay. Because that's been one of the biggest concerns from people when they talk about the bubble or about how companies like OpenAI are going to make money and how companies like Nvidia or yours are going to make money because obviously if the chips are lasting only for six years the amount of capex you cannot make the money back. What's your take on that?
I think you probably want to be conservative in your estimates. These chips are behaving differently than others have historically. Historically we used five or six years for the depreciation. These chips look more like maybe six or seven. And you can see that there's still an active market for H100s and there's even a market for A100s which are now six, seven years old. And so what we don't know is what happens maybe when supply isn't as constrained for the new chips. And that's still unknown. And I think that the challenge when you're in a whole new area is when you can look at the past and learn from it and when you have to think that something's going to be different than the past. And that's a real challenge right now.
Right. Not long ago, a couple of months ago, we were able to go to China and we were visiting as well the Huawei. We were able to meet the people from the PR from Huawei. Obviously, they didn't let us in like you guys by the way. Thank you for showing us around. What do you think about China in the chip manufacturing because I have the feeling that people is sleeping on China and people doesn't see Huawei as a competing for Nvidia for Cerebras and actually it looks like they're doing a lot like they're very far behind maybe but...
I think sleeping on China is a big mistake. I think they've made it a national priority. They used government resources to support their AI software. They used government resources to make giant data sets available to their AI software users. They have used government resources to build out infrastructure like power. I think sleeping on them is a huge mistake.
How far behind are them with the chip industry?
I think with software, with the AI models, it's probably six or eight months, maybe a year. With chips a little longer but they are, you know, it's a lot smaller than it was 10 years ago.
Yeah. Or right the question isn't where they are today it's what the trajectory is.
And they are working really hard to be better at building chips.
Yeah we seen the last DeepSeek report where basically was trained for Nvidia so therefore they kind of finding their way.
That's right. I think there are lots of different views on how we should manage our export policy and good arguments on both sides. But what's clear is their chips are not at the cutting edge. They're behind. They're closer to the cutting edge than they were 10 or 15 years ago. And you shouldn't sleep on them for sure.
Or we will pay the price.
In Europe over the last I think 18 months we made the decision like the governments of Europe they made the decision that we needed more data centers. They call it gigafactories. Like recently Spain announced 800 million euro investment on a gigafactory which is not much in terms of America but...
We play in a different league. It's not the Champions, it's the... but...
But we have a saying there that we say with some of the people I normally talk about in the same time China built I think 11 massive data centers and in this time Europe is still fighting for where do we put them in Europe.
In the US the problem is different. It seems like you're getting the public against data centers.
Yeah. What do you think about that?
Well, first I think that for big public projects, our systems of democracy are slow. In Singapore, in China, in the UAE, they've been able to move extremely quickly and vault ahead, right? Our process of inclusion has some pros, but it also has some cons, some negatives. And for these big infrastructure projects that cut across communities that have many constituencies, we're not moving fast. And that's true. Not as fast as we should in the US also.
But beyond that, we've been seeing more of a resistance from public. It's mostly based about what they think it's bad for them.
So the industry at the beginning made some bad decisions. And it began maybe with the Bitcoin miners who were looking for the cheapest power and were looking to maybe push off some of the cost on local communities by finding loopholes. And as a result and then some of the AI data centers the early developers did a bad job of including the community in the discussions. And as a result, they got a bad reputation and people are upset. But there's nothing inherent about a data center that should be bad for the community. We should be able to build data centers that include the community. They're good for the community, that create thousands of construction jobs, that increase the value of land and the tax base for these communities. They don't need to use a lot of water. In fact, all the data centers in the US use less water than the almond industry in California.
Almonds. What's an almond?
Almendras. See, almendras. Right? Now I understand. Now I understand Barcelona.
I think Satya Nadella said not long ago that a data center, one of the modern data centers, takes the same amount of water than a restaurant.
That's right. A small restaurant. And so they were the leaders. So one of the leaders in this one they said look we can build data centers and be good neighbors and here are some things and Brad Smith put out a publication, a note that said here are some guidelines that everybody should follow. And they weren't a whole lot more fancy than be a good neighbor. And when data centers are built, it's entirely possible for the industry to be good neighbors. We can make the communities better off. We can help their prices for power go down, not up, and we can increase the value of their property. We can pay taxes so their schools and local communities are better off.
But should it bring employment? And bring employment because these are put in areas where there's a lot of land and unused power resources because we're in a situation where obviously there is a lot of discussion let's call it about what AI is going to do on employment and for me it feels you know I have we had similar thing in Spain with the nuclear plants yes there was a time where we wanted to build nuclear plants that was when I was young and people started banging against it of course that was like big in Chernobyl and then people was really absolutely against it but it doesn't fit in my head that people can compare it or they can think about there something because that's not that kind of blow up.
And also France did it and they built some of the lowest cost power in Europe is clean.
As far as I know people that I know that lives in cities where they build nuclear plants they are very happy.
That's right. They're very good jobs, very qualified jobs, they pay really well and the whole area had a boom on being really good. So, especially when you take off the danger that obviously a nuclear plant brings and you talk of the data center which is like computers in a rack.
That's right. These are computers in a rack.
They shouldn't use resources. They take I mean a big one, you know one of the ones that I'm familiar with is one of the bigger ones. I mean they're 7,000 jobs running three shifts a day.
On a data center. 7,000. That's sort of 24,000 jobs per day are made at this data center, right? That's a lot of money to a community. And those people go out and they have a coffee and they get a pastry, right? And they buy lunch and they do things. And so but at the beginning, I think we earned some of the ill will in the industry by not being good neighbors.
And paying the price now.
And now we're paying the price a little bit. But I think there's some areas in Canada in some of the Nordics in the north where they're inviting people to come build data centers. They've got a huge amount of geothermal. They've got very low-cost power. They're short of jobs and these data centers are coming in and being really good for the community and that's what we ought to aspire to.
And what do you think it has to change for people to understand it? Like is an education process. People has to get it.
Once you infuriate people, it's very difficult to put them in listening mode, right? Once you make them angry, it's difficult to put them in listening mode. I think there's an old saying about from an owner of a casino and it's the worst thing about having someone come and beat the casino is seeing them go next door, right? That the worst thing for these communities is for the developers to say, 'Great, we're going to go to Saskatchewan. We're going to go to Sweden. We're going to go to rural Texas' because then all the money leaves and it goes somewhere else.
And especially when we are going to see that with AI being so important data centers are going to be a matter of national security.
That's the separate issue is that it's probably important to have some control in within your borders of AI.
Yeah. Because basically we seen that with the chip industry we have been delegating to other countries like Taiwan but now we are paying the price of it and it's a challenge you guys are trying to bring it back to...
We are trying to bring it back to the US and I think that's a strategic imperative. And some of us have been arguing for this for 10 or 15 years that there's this is you know in previous wars your manufacturing base was valuable, right? And we saw what happened when during COVID when a ship got stuck in the Suez Canal and you know there's a three-month delay in getting a dishwasher, right? And that was just a delay. The supply chain for chips is complicated and multinational and leaves nations vulnerable.
How is it right now for you guys like where do you get your parts?
Taiwan. This chip is designed by us and manufactured in Taiwan and it then comes back to the US where we manufacture the whole system around here about 20 minutes from here and then we ship it around the world for delivery.
And there is a thing that we supposedly do very well in Europe that it goes even beyond the chip manufacturing and there is a company I think is in Sweden or is it in Switzerland? I don't know what is it. Or something like that.
That's what they're in I want to say Holland, ASML.
Exactly. And they are extremely unique, right?
Yeah, they make something called a stepper. Which is how you etch the transistor into silicon. And they're the only company in the world who figured out how to make it. And for 20 plus years they've had a monopoly. It is such a hard technical problem that many companies have tried and failed and so TSMC buys from them. Samsung buys from them. GlobalFoundries buys from them. Everybody buys from them to make this piece of equipment that is required to make chips.
So that's basically the first step, the first stone that...
That allows for everything. That is inside the factory at TSMC is their equipment.
Okay. Wow. And I heard that China seems that they're starting to compete on that.
They're going to try. Nobody's ever been able to build.
So far nothing. But lots of people have tried. Turns out to be hard.
What would it mean if China was able to do that breakthrough?
Then they could build fabs in China and the control on the shipping of that factory equipment would no longer delay China's sort of progress towards being equal in chips. So that would be probably the most important thing that China should do if they want to catch up.
They been trying for a long time. It's a hard problem.
It's a really hard problem.
And what should Europe do now? We heard a lot about the governance of AI not being able with the latest news about Fable being taken away and export controls from America. Europe panicked a bit. We called this G7 meeting where they were like being very friendly to Trump being like okay let us use the AI that you're having because at the end of the day we have nothing that we can... you know I think sometimes and maybe you can correct me if I'm wrong but in Europe the first thought when a new technology arrives is how we could regulate it. That's the hobby we go there.
Yeah. And I think the result is in America the first thought is how we can make money from it. Right. And probably both are not exactly right. And that we could think a little more about how it could be safe and in Europe you could think a little bit more about how you could innovate and move it forward. I think that it is important that European companies and European nations create environments where AI flourishes. I mean Google's AI came from an English company, right? It's not like it can't happen in Europe. There was an English company outside of a great university, right? And the way you create this is by having universities where people can innovate.
and have professors who are excited about new technology and where you have investors who want to bet on companies that are young and innovative.
It is a hard part in Europe.
It's not in your culture.
