Welcome back everyone. I'm John Furry, host of the Cube. We are here for our Future of the Data Center series. This is the large-scale systems that are changing the game, and we're interviewing the leaders who are powering that trend and continuing to invest in the fastest, latest, and greatest enablement for the software tsunami with agents and all the new AI-native applications coming. We've been reporting on that on SiliconAngle. Andrew Feldman, the founder and CEO of Cerebras, is back on the Cube. Last time we were together was in Paris at the RAISE Summit. Andrew, thanks for coming in. Thanks for coming into our NYC Cube studios here as part of the NYC Wired community. Appreciate it.
Thank you for inviting me. What a cool spot you have here.
Yeah, I mean, got the active trading, the energy. I love it here. The energy, of course, with the Palo Alto studio connecting Silicon Valley and Wall Street. Those environments are now merged, right? So you see money in tech. It's not mad money, it's tech money now, right? So it's like really now a tech-enabled scene here in this market. We've been covering the crypto trailblazers, which is essentially AI in, I mean, decentralized infrastructure, financial infrastructure that's being dislocated and innovated on. So that's one vertical. Life sciences, I mean, every single area is popping with new discoveries, new capabilities, and, you know, enabled by large-scale systems that can compute and infer around things that they couldn't do years ago.
That's right. New processes, new methodologies, new techniques, new ways to find insight in data. They already have guidance to which data they need to collect to gather new insights. It's really an extraordinary time.
I want to unpack that with you because you're at the forefront of this with speed and scale. But first, give us an update on Cerebras. What's new since we last saw each other in July? Give us a quick update.
Well, the inference market has continued to accelerate. It is white hot. And speed, which is the reason we exist, you know, as you remember, we're the fastest at inference in any model we do. Not close, 10x, 15x, 20x faster than the nearest competition. And that's important as we use inference to do more, as we use reasoning models, as we use agentic models, right? As we use inference to write code, the willingness of people to wait 10, 20, 30 seconds, a minute, 3 minutes, right? Nobody wants to sit and watch the little dial spin while they achieve nothing. And so by being able to deliver inference at these extraordinary rates, we found extraordinary demand.
So talk about the demand curve. We're seeing in the news huge financing and capex growth. More data centers are being built. Applications that are AI-native getting funded. We saw Nvidia going all in on OpenAI. I think they have like five major customers that represent a lot of their business. So obviously they have some big customers. But concentration, it's true. It points to though the demand and watching some of the, you know, the hot takes on TV. Obviously they're pretty obvious low-hanging fruit kind of obvious ones like, oh, the sector's hot, it speaks to the buildout. Great insight there. Super insightful, okay. I mean, it's hyped up, so hype and hot kind of synonymous in that layer of analysis. But when you look at the demand curve, the infrastructure enablement is still firing away. Can you take us through your thoughts on that? Because again, on the software side, there's a huge activity going on. Certainly open source on all the models, they're moving faster, so the models are like hitting a flywheel. People are decoupling the app layer, they're natively infusing AI and either native AI at or they're consuming AI to manage the models, and then the underlying infrastructure is kind of decoupled from that flywheel that's being optimized. You have layers above and beyond the models. That's the smart money. So the software side is in massive demand. They need horsepower.
I think one of the things that leads you to believe this is not a bubble is that all of us are looking for more capacity. We just opened a new data center in Oklahoma City. It's our fifth data center in the US. This is a 10-megawatt facility. And we immediately began expansion. We opened it, we expanded it, and we're looking for data centers around the world. And I think when OpenAI, when Anthropic, when other leaders are saying we need more compute, all right, when leaders in the IDE space, all right, Cognition, Cursor, or Lovable, when they say we need more capacity, right, what you're seeing is this sort of groundswell up out of the application, and it's washing over the makers of software. And there is just this giant sucking sound, and that sucking sound means we need to deliver more compute to do more inference. That compute needs to live in more data centers, and it just washes back down the supply chain.
You know, one of the comments that's come out of some of these interviews on these Future of the Data Center is, well, we've had x86 for a while. We got GPUs and XPUs. They've had that relationship for years. It's an interconnect. You obviously have a semi background as a founder. It's kind of just interconnect them, get the same. And then these NeoClouds are just resources. It's just resource capacity. Not so much a game-changing software perspective. What's your take on that reaction? Is it naive or is it outdated?
Both. Both. There are very few problems in this world where you can just add and connect for free. Right? When you break up a problem and you separate it, there's a communication overhead. The probability of miscommunication jumps through the roof. Right? And so as you fracture a problem and spread it out over thousands of GPUs, the problems multiply, the complexity multiplies, and that's why there's so few organizations that can train with 10,000, 15,000, or more GPUs. There may be a handful in the world. And this is complexity that at Cerebras we seek to eliminate by building the largest chip in the industry. We keep more information, more data on chip. We move it less often. We use less power, and we deliver answers in less time.
You know, we saw the news this week of Intel has a government stake in it. They'll come from the philosophy of go smaller, not bigger. Yeah. How's it worked for them in AI? I mean, explain, because it's not, they would probably throw shade on what you're doing. What's the net-net of going bigger with the wafer? Because you have cracked the code on speed first.
Okay. First, Lip-Bu is an extraordinarily knowledgeable and seasoned executive, and he was one of our early seed investors. So he knows exactly what the benefit is. I think that the notion that there's a one-size-fits-all in technology or that there's a single silver bullet is always garbage, right? It's just never the truth. The interesting parts about AI are that the challenges that make the AI work difficult for a processor are several. First, you need a lot of very simple calculations. You get more than generally fit on a small standard-size processor. That means you need to tie them together. So that means you've got this crazy communication pattern of moving intermediate results. Final thing is you need to get to memory, and we are two and a half thousand times faster at getting to memory. And getting to memory is the speed factor in inference. I mean, SRAM was the holy grail.
SRAM is the holy grail. And by how much can you put in there?
Well, that's why you make a big chip. That's why your chip's 56 times larger than the largest GPU, so you can stuff it to the gills with the fastest type of memory on Earth. And this is this architectural decision we made, these inventions. This is where the advantage comes from.
Yeah, we talked about that big bet on our last segment. So, folks watching, check out that interview from the RAISE Summit with Andrew. He went into great detail on the speed there. So, I'll just put that out there. Talk about the data centers because inference is where the money's being made. Everyone I talk to, inference is where the money is. Now, let's unpack that. Certainly on the infrastructure side, you're building inference centers. My word, not yours. I'm just maybe that's the wrong word, but that's fine. They're inference massive data centers filled with machines that are running inference. And so software writers, whether it's written by a machine or a human or both, they want to deploy software that's intelligent, right? They're going to need to have that capability. Talk about where that goes because I think this is where I'm connecting the dots with Cerebras. Because okay, I'm a software. I don't really care who's got what. That's right. I just want the fastest. No one says I want the slower product. That's right. Nobody's ever said, can you make my results appear slower on the screen? Right. No one says that. Yeah. Totally. So take me through that enablement because I think this is where I think the crossover is on the money-making on both sides, by the way.
So what happened in early 2024 was for the first time AI sort of moved from being a novelty to being valuable, right? In the sense that before that, the emphasis was on training. There was a little bit of inference, but we couldn't do useful things with it. And so it was cool. You could do this or that, but could it be useful? And starting in the middle of 2024, what we saw was sort of the rise of AI being an extremely useful tool. And it could write code, all right, it could generate text, it could deliver thoughtful answers, it could do deep research, it could survey vast amounts of information, synthesize it, summarize it, and deliver it back. And so what's happened is software writers want to take advantage of that technology, right? They want to bolt on to AI. They want to use AI, use a chat service inside their application. And what's happened is it's become dead simple to move, right? It requires no special coding. You can move from a competitor to Cerebras in 10 keystrokes. All right. And once you do that, you're getting results 10, 15, 20, 50 times faster. That's what's sort of driving this is sort of adoption across all app makers are including some form of AI and they want it dead fast and that's where we come in.
I want to get your thoughts on the AI hype and to reality because I think, yeah, if you go back to 2024 into 2025, every company, we're an AI company. I mean, no one's going to say they're not an AI company because that's like, I mean, unless you like clearly aren't an AI company, right? Construction equipment, right? That probably has some IoT AI in too. But then now we're starting to see the rise of AI-native startups, people who are thinking AI-first, who are building AI into the products from day one, not bolting it on. That's right. What does AI-native mean to you? What does that definition mean? Because we saw cloud-native, cloud-native really powered the SaaS market. It did. Is AI-native real? Because agents have workflow knowledge and you mentioned things like moving data around. I'm paraphrasing, moving data, access to memory, the bigger chip does all that. That handles a lot of that velocity. It does. And it's nonlinear, right? It's also not predictable, is generative.
So I think what AI-native means is that AI and AI's capabilities are in the very architecture of your product. That they are not an add-on. The product is built around those capabilities. Right? And you are thinking about in every step how the product can benefit from AI, not how AI can help the product. If you're bolting on customer service or a chatbot, that's great. It can do really interesting things for you. It can save you a ton, but that's not central to your product, right? When the very value proposition of your product is sort of cornerstoneed with AI, then you're an AI-native product. Otherwise, you're sort of an AI drag.
You know, one of the things about New York that I love, and some people might not understand, right, today couldn't have been it, that was not it, is that they're results-oriented. And Silicon Valley is a little bit of a long game on tech. As a founder and as a technical founder too and CEO, you're in like a great machine, hype machine. We are. How do you balance that? Because you got to balance results. Okay. Speed, you certainly deliver the speed side. And then also play the long game knowing what the bet is and what you guys going to see through. How do you handle that? Because it's easy for some founders that may not have the experience to say, I got to get the shiny new toy. I got to be on that high horse and look at the deals I just closed. Yeah. A lot of people, you have to squint through that. So we're trying to get that. How do you view that, handle that? And what's your philosophy?
Well, as you know, this is my fifth startup, and over the past 27 years, there's some big highs and some deep, deep lows, right? And what I tell people in the startup world is you can get kicked in the stomach three or four times before lunch and have it still be a reasonable day, right? You are being whipsawed. And I think you add that to the fact that no great company, no company that changed the landscape did it quickly. All right. Change, fundamental change takes years, and it takes real R&D, and it takes the type of R&D that is unafraid to fail, that is working on problems that other people have refused to take upon themselves, that either they've been too scared or they didn't think they could solve it. Those are the problems we're interested in. We talk about it inside the company as doing fearless engineering. And when you solve problems that matter, that are hard, that others haven't solved, that produce real benefit to your customers, it doesn't take three weeks. This isn't a 10-day sprint. These are years of extraordinary work. And it produces massive results like our innovations in chip design and in chip packaging, which lead us to be an order of magnitude, an order of magnitude and a half faster than companies that are 50 or 100 times larger than we are.
