Powering High-Performance AI: Cerebras’ Andrew Feldman

Welcome to The Software Report, a leading information source on software companies and solutions. In this episode, we speak with Andrew Feldman, co-founder and CEO of Cerebras Systems, a company developing high-performance AI compute infrastructure for training and inference. Cerebras is known for its WaferScale engine, currently the largest processor built for AI, and for its CS-3 systems, which can be clustered into some of the world's largest AI supercomputers. Enterprises, research institutions, and government agencies use Cerebras technology on-premise and through the Cerebras Cloud to accelerate model development, real-time inference, and large-scale experimentation. Before founding Cerebras, Feldman co-founded C-Micro, where he pioneered energy-efficient microservers later acquired by AMD. Earlier in his career, he held senior roles at Force 10 Networks and Riverstone Networks through its IPO. I'm your host, RJ Lumba. We hope you enjoy the show. If you like the episode, click to follow. Andrew, thank you so much for taking the time. It's a delight to be with you.

Well, thank you for having me.

I was thinking about where to start, and you've recently raised a sizable round, I think over $1 billion. Fidelity was one of the leads.

That used to be a big round.

Yeah, I think it still is. So maybe what we could do is just hear about where the company is today in its life cycle. Maybe give us a sense for the scale of the business currently and kind of the, the reason or the rationale for the capital raise.

Sure. You know, we have about 730 people. We have offices here in Sunnyvale, in Toronto, in Bangalore, and a little one in, in Dubai. We are growing unbelievably quickly. Yeah, I'd say we have order 100 open recs right now. We've just hired key leaders in sales organization in the Middle East. We just hired one in Asia. We are in the midst of the inference boom. We are awash in demand because we are delivering the fastest inference in the industry, not by a little bit, but by more than an order of magnitude. And that's being deployed in coding, in deep research, in sort of this collection of different areas that are just booming right now.

Mm-hmm. When we think about the client base or maybe the end user, where is it most prevalent? I mean, can you break it down by either sector or use case?

Sure. We have large enterprise customers like GlaxoSmithKline and Mayo Clinic and governments and research institutes. And we have customers all the way down to individual engineers and startups who are served by our self-service cloud offering. Growing extremely quickly right now are IDEs or coding agent makers. They're booming right now. And also we have business in the sovereign clouds in building sovereign infrastructure for nations. And that business is going well too. So it's an exciting time. There's a huge amount of demand in both inference and in training. It's a good time to be building hardware, it turns out.

Mm-hmm. And was this, I guess, demand that you're currently experiencing, was it anticipated when you founded the company? I mean, the original idea was that you would be doing exactly this?

I wish. I wish I were that good or that my marketing was that good. I think we had some idea in 2016, late 2015, we had no idea AI would be this big. I had no idea.

Mm-hmm.

But we knew it was on the horizon and we knew it would be big enough to be interesting. And we knew that if we worked hard and were creative, we could build a processor that was vastly faster at AI than anything that had been built at the time.

Mm-hmm.

And what we decided to do was to see if we could build a, a chip the size of a dinner plate. And by doing that, we could solve some of the hardest problems in AI compute. We could keep data on the silicon, mm-hmm, could move it less, therefore use less power. We could put fast memory onto the silicon, and therefore our movements to memory and from memory were thousands of times faster and used less power. And we knew that if we did these things, we would be in a, a leadership position. Now, we, we didn't know that the market would be this big, that all these great things that have happened in the last 2, 2.5 years would happen. But we knew that by making bold decisions, by solving problems that others were afraid to tackle, that others hadn't been able to solve in the entire history of the compute industry, we would put ourselves in the middle of something big.

Mm-hmm.

When you start, that's sort of what you aim to do.

And now to get a little bit more context around this, that chip that you just displayed there, like in what, percentage of the data centers out there could you find your chips?

You'd have to work pretty hard right now. Even though our business is ripping and our numbers are really big, look, NVIDIA's the market share leader. They have a lot of share. But look, we have now data centers in Santa Clara. We have data centers in Stockton, in Dallas, in Minneapolis, in Oklahoma City, in Montreal. We're opening data centers in Toronto. We're opening data centers in Norway, in France, in Vancouver. So we're going, but I think, you know, the honest answer is, is we're still a modest market share.

And is there a way to make a direct comparison as to how fast or more efficient?

You absolutely can. You can run the exact same open source model on an NVIDIA processor or server and run it on us and What the benchmarks will show you and what production use will show you is we're between 15 and 25 times faster.

Mm-hmm.

All right. That's not nothing. You get more than an order of magnitude faster than NVIDIA.

And is the constraint then more capital than, like, if you threw more capital at this? I mean, the market's going to be there, right? If you're that much better.

I mean, we just raised more than a billion on that hypothesis.

Right.

And We are getting more data centers, we're hiring more engineers, we're expanding manufacturing. We are growing about as fast as any hardware company in history has grown.

And you're no stranger to the hardware space. Maybe if you could tell us a little bit about your background and kind of the prior companies you were leading.

This is my 5th startup. Immediately prior to this, I started a company called C-Micro and we were pioneers in building extremely energy efficient servers. And we were acquired in 2012 by AMD, and we all worked there for a couple years. The founding team here at Cerebras all worked with me at C-Micro, so we'd all had years of working together.

Mm-hmm.

And in fact, we have, what, nearly 100 people here who've worked with me before over the past 25 years at one of those startups.

Mm-hmm.

And so, uh, we have a world-class, I mean, I, I'd say there are probably 8, maybe 10 great chip teams in the world, and, and we have one of them. Those guys have been with me since the late '90s. We have a collection of truly extraordinary people who've chosen to come back and be part of the team, not once, not twice, in many cases, 3, 4, 5 times.

Do you have a philosophy around leadership? You're able to retain talented people that are probably highly coveted throughout the market that you play in.

One of them told me I was the least bad boss they ever had. So I, I Seek to be less bad than others at leadership. No, I, I think a couple things. One is they can ask me anything, how much I make or how much stock I have, anything. And I'll tell them in person or in an all hands, I'll tell them. Number 2, it's been shown over years that after you're sold and everything is public in legal documents, that I've done exactly what I said, that I didn't seek to enrich myself, that I put their interests first and that I, I really was working every day to make them and their families better off. I think the final thing is I have a nose and an interest in hard technical problems, and I'm not afraid to raise money in pursuit of them. And I think extraordinary engineers don't want to do what they did last time. They don't want to work on trivial things. They want to work on really hard things and really interesting things. And so if you can bring together interesting problems plus really smart colleagues and management that has integrity and is, is actually really working every day for what's better for the team and, and their families, got a good chance of keeping good people.

Mm-hmm. Given your current scale and how fast you're growing and the amount of capital you just brought in, do you find that you're spending a good portion of your time recruiting and dealing with, you know, just the key executives and ensuring everyone's aligned?

Yeah, I think right now I'm spending a lot of time working on data centers since they're a limiting factor in our growth. I'm on the road, I'm talking to people who are either owners of data centers, who are developers of data centers, who own the key inputs to production of a data center, whether that's transformers or generators. We just signed an agreement with the government of Guyana to build a giant data center in Guyana, which is south of Venezuela and north of Suriname, if that helps. It didn't help me at first. We are talking to people who own gas rights and mineral rights. We're talking to people who have access to geothermal in the Nordics, people who have, and organizations that have access to low-cost hydro. Around the world. We are constrained by power right now and data center space. And so that's what we're chasing. Now, until we were about 400 people, I interviewed every single person we hired. I think, uh, I still interview everybody who's over a certain comp. I think if you're going to come in and be a leader, it's, it's nice to know the CEO. And that's not just managers, that's world-class individual contributors. That's people who are going to have a big impact and seek to have a big impact. I'd like to get to know them.

Mm-hmm. Referring back to the data centers and how that's the limiting factor, it sounds like you're playing a key role in the orchestration of the construction, meaning you're one of the parties that's part of the team that has to get the data center off the ground.

Yeah. You know, you used to, in our business 10, 15 years ago, you'd pick up the phone and call Equinix and ask for some capacity or, or Digital Realty Trust. Or some guys you knew who were regional players who had 2 or 3 data centers. And you'd say, look, I need 20 racks. And they'd say, all right, I'll get you a quote. And day later you get a quote. No more. Right now you're talking to people and they're talking about delivery in the first half of 2027. You're talking about understanding which generators and how they're getting their power and whether you're on the grid or behind the meter. You're trying to think about what level of redundancy you want in your data center, because that impacts cost. The data center has become the factory floor for making AI. And this isn't BS marketing. It's heavy equipment. You're bringing in an enormous amount of power. I mean, we hear about these gigawatt facilities being built, and I think Pittsburgh is fifth or sixth largest city in the US, and the whole city consumes 3 gigawatts. And so we're talking about in these deployments, sometimes data center deployments that use the same power as millions of people in one set of buildings, in one patch of land in Texas or in Louisiana or Mississippi or in the UAE. And so these are extremely complicated structures. And I think those of us who would love to be in the business of just renting space, ended up having to integrate upstream and being more involved in their construction and their sourcing in order to get 'em.

Is there such a land grab, so to speak, such that if you were not to be actively involved, someone else—

Oh, for sure.

So you gotta be there first.

You gotta be there first. You gotta be creative. I mean, the guys who've been really creative are some of the Bitcoin mining companies. Early in the game, they were experts at sourcing power. They didn't really know how to build data centers for AI, but they knew how to get cheap power. And so Crusoe came out of that heritage. TerraWolf came out of that heritage. FluidStack came out of that heritage. Hut8, a whole group of these guys are really good. Mara came out of this heritage and now are looking around saying, we really understand the power landscape. Well, I had to learn it. They had to learn how to put up a, a tier 1 data center. That's something I knew. And so we're coming at the problem, us from above and them from the ground up. And there was a lot of learning for both sides.

Is the whole Mid-Atlantic area, is that already kind of like old news? That's like saturated. There's no more space there.

If you're in Virginia, it's very hard to get space. If you're in Lenoir, North Carolina, these facilities ate up everything on the grid, right? I think what people are looking at right now, there's still some opportunities in Ohio, particularly in the refurbishing of heavy manufacturing facilities. In West Texas, it's all gas.

Mm-hmm.

So you're behind the meter using propane or some derivative of, of natural gas. That's also happening in Utah and Wyoming. In the Nordics, you're using geothermal. In northern New York, you're using falling water, which is a great source of power, very cheap and renewable. So everybody's trying to build on the advantages they have.

Mm-hmm. Would like to learn a little bit about you, where did you grow up and were you always kind of an engineer or a computer scientist?

I am neither by training. I grew up on the Stanford campus. My parents were, were professors.

Okay.

I went to Stanford as an undergrad and went back to business school. I got interested in technology by mistake and late while I was at business school and ended up writing a business plan for some friends and we ended up starting the company and we sold it for to $290 million about a year later. And actually those guys are still with me to this day. They're still at the company.

Mm-hmm.

And what we learned was that I'm not a technical visionary, but I'm pretty good at listening to visionaries and thinking about how to match that to solving a real problem, solving a problem that customers need. And so my career sort of was through product management, and marketing through business development, trying to understand what customers needed and how I could convert the work of my brilliant friends into a product and into a customer who was really happy.

I think it's incredible that you're able to kind of retain this circle of folks around you, and maybe it's not necessarily you actively retaining them. It's all more naturally occurring, and it's because of those principles you mentioned.

I think many of the traditional sort of ideas of leadership don't work in Silicon Valley, and they don't work because I've got hundreds of guys who with a phone call can work anywhere.

Mm-hmm.

And I think you can't lead without people choosing to follow. Right? I mean, and that's a really interesting point.

Mm-hmm.

I think especially as they get a little more senior and maybe have worked elsewhere and seen some political environments, seen some problems that were ordinary or pedestrian. And even if they did good work, didn't feel like they were gonna move the dial. That when we get to work together, there's a, a sense that we wanna work on hard problems, we wanna do big things, and that the people around them respect them and the, the people who they work for are also working for them, right? I don't think there's ever a minute where these guys think that I'm not giving my best for them and their families.

Are you primarily hybrid, remote, in-person?

Primarily in-person. Okay.

And you probably find that makes a big difference.

Look, I think the problem really is, is that remote is the death of serendipity and that meetings don't actually end like that, right? Like they do in Zoom. What, what actually happens is people who are interested form a little group and they step out and they work on the whiteboard for 15 or 20 minutes. And, and real progress happens in these little groups right after a meeting. And while it's certainly true that in some phases of a project, quiet, uninterrupted time is really important, it's also true that there are large portions of time where having lunch with your colleague and continuing the discussion sitting next to somebody in hardware when you're in software or software when you're in hardware or hardware when you're in manufacturing advances the learning.

Mm-hmm.

And I think particularly for a hardware company, and I know many of the companies that are all remote or software and particularly open source, but when you have really tight engagement between hardware and software, mm-hmm, where you don't have clean handoffs, but they're negotiated, right? Not in a negative legalistic way, but in a way that as you're inventing technology, sometimes it makes more sense for the hardware to go a little further, and sometimes it makes sense for the software to solve this problem.

Mm-hmm.

Those are clearly best done in person, and that's been our experience.

We're coming up on time. I have two final questions. One is, can you tell us about a person who has had a profound influence on you?

Yeah, I, I think there are lots of people who've had profound influences on me. I think, uh, Pierre Lamonde, who's one of the great venture capitalists of his generation, now at the age of 95, uh, still calls me and tells me what I'm not doing right. He showed me how to be a good CEO in a lot of ways and taught me a great deal about what hard work really is.

Mm-hmm.

Mark Leslie was, I think, not only one of the great CEOs, uh, he was CEO of Veritas. And then was a great board member and thoughtful and fair. I learned a great deal from listening to him. Obviously my parents, but that's an answer that while often true is less interesting than maybe than some external forces.

Last question, can you tell us about a charity cause or other endeavor that you're passionate about?

I think AI will have a huge impact on education. I think we have done a horrible job at bringing technology to education. I think in many ways, maybe holding YouTube apart, which was a fundamental change in which you could get access to your passion and hobbies on the long tail, right? But for sort of K through 12 education, I think we haven't made much progress since Aristotle was teaching Alexander the Great, right? What institution looks similar 2,000 years later, 3,000 years? We don't use data to teach, right? There are a class of mistakes. Some students make that if you taught it to them differently, they would do better. And we never say that. We never say, you say, oh, you got it wrong, here's the red mark, here's your test back, and they throw it away and you move on. There's no closed loop here. Whereas if instead what you got was, you know, 8% of your class made the same type of mistake, and this type of mistake has been shown to be remedied really effectively with this other approach, Right? Suddenly you're using data to inform how you teach. I think AI has a huge opportunity there.

Mm-hmm. Andrew, I want to thank you again for taking the time. It's been a great conversation.

Well, thank you for having me. I really appreciate you making a little bit of time for us today.