Today we have the pleasure of speaking with Andrew Feldman, founder and CEO of Cerebras Systems, the company behind the largest chip ever built and the fastest AI inference engine in history. Led by a veteran team of infrastructure builders, Cerebras is redefining what's possible in high performance compute, quietly becoming one of the most technically ambitious and resilient companies in the world.
Andrew, thank you so much for spending time, man. I know you're extremely busy. You're probably one of the hardest working AI founders I know or any founder I know. Like, I don't know how you stay married or stay sane.
I think I got lucky with a wife.
You messed up. Welcome to the club. Thank you. There are so many great things about Cerebras that I love because as you know when you find a company that has a story that's just like totally surprising then you have to, you want to tell that story. The stories of like, you know, all the usual suspects, OpenAI and all these other guys, drama or no drama, they're fairly linear. Like Anthropic's linear, OpenAI is sort of linear. Cerebras is not linear. And so I thought what we would do today is talk about your journey because it's such an unlikely story. I mean, you know, starting with the fact that you started this company 10 years ago before anybody knew what the AI was going to even be. I mean, we already knew what AI was but they all thought about it as Big Blue and like chess, finding a cat.
So, let's get into it, man. I want to, what I would love to do is talk to you, start with like, you know, who you are. Of course, I've known you for 20 years. I hired you forever ago for, you know, who, and I've always thought the world of you. But what I'd love for you to do is share with the audience just a little bit about who you, who are you, what your, a little bit about your upbringing and any meaningful moments or vignettes or mentors or life experiences that led you to where you are today, which is doing probably one of the hardest things in the entire AI ecosystem that is imaginable, which is a gigantic GPU kicking Nvidia's ass.
So yeah. And tell us about you.
So I was born and raised on the Stanford campus. And so my parents were professors and I think one of the things that really teaches you is that the world isn't going to get a whole lot smarter than your first grade class, right? First grade we had only faculty children. We were in that class. And we grew up in an environment where ideas were the currency, not money. And where, you know, my dad had a tennis match every weekend. There were six of them who played tennis in rotation and four ended up with Nobel prizes and one with a Fields medal.
You learn what brilliance is and you learn that it's not arrogance. If you brought them a question in their field after tennis, they'd explain to you how something worked and they were willing to teach. And when you stopped by when you were eight doing a readathon when your school was raising money for books, they would sign up for you. And years later you read their definitive work in their field.
So this early experience bred a deep curiosity, deep intellectual curiosity. Is that what you're saying?
Yeah, it bred real curiosity and curiosity without sort of absurd arrogance, without sort of it being attached to wealth. Two doors down our neighbor had two Nobel prizes. Two. And what we knew growing up was his wife gave full-size Three Musketeer bars at Halloween. That's what we knew.
You know, we have another founder, this woman who's the founder of Vanta did a podcast with her, Christina. And she, you know, her parents were college professors. And I thought what was interesting about her journey was going from this deeply academic environment to this just this raw, you know, entrepreneurial company building capitalistic sort of environment. Tell us about that transition. So you go from this academic-heavy, brilliant Nobel Prize. How did you start thinking about, hey, I want to go out and do something in the business world because you've done like four or five startups now and you've been a company builder for decades.
I was always somebody who was an entrepreneur and in high school I had a t-shirt company and Nordstrom's was carrying our label. I've always enjoyed finding things that people want or need and trying to build them and sell them. That was a passion from the get-go, you know, my brothers are doctors. And I never had any interest in that. I was always interested in working with people who could build things and finding customers who could use them and getting sort of a high from watching your customers really enjoy something you built. And if you don't like that, then this is the wrong thing for you. You need to really love watching your customers enjoy a product or a service that you invented. And that I knew early on. I took great pride when I'd walk around campus and I'd see people with t-shirts that I'd designed or that we'd made. One of the companies I was a part of was one of the first to build hardware-based switches and routers. We took great pride in being part of a movement that made IP networking free. We didn't know anything about WhatsApp. That wasn't our thing. We didn't know about Viber or any of these different techniques. But what we knew was that if you could make networking approximately free, other people could build amazing things on your stuff and the world would move forward. And so that's always been a passion and I think that's what led me to infrastructure. To do infrastructure, Andy, you have to love building roads and you have to really take great pride in what other people drive to cool locations on the road you built.
Well, you know, it's funny because you and I grew up the same time as, you know, professionally speaking. And I remember graduating from school in the early '90s and being like, what's this internet thing all about? Like we had CompuServe, we had AOL and Prodigy and they were all closed. There was no Mosaic, there was no Netscape, there was no anything. There was no Yahoo. And I had to actually set up my first modem and that took me like three days to figure out how to program a modem.
And then so they like, we remember we called it the information superhighway. We were building the information superhighway.
Yeah. Now maybe it's building us, right?
I know. Exactly. It's incredible. So you started in this transformative world where you're building the on-ramps to the internet. Okay. And like that's where you and I met is when the equipment just wasn't there. We had the MAE-West, MAE-East. They had these concentrated, as you know, data center equivalents with lots of networking gear.
And it grew out of DARPA and became the backbone of the internet. And so you're building that key contributor to major companies that ended up going public and some of them just became, you know, big acquisitions, but effectively they were really difficult products to build. And as you recall, ASICs, application specific integrated circuits were getting more and more dense and kind of these were like optical transport networks and what have you and speed was always like the thing you were trying to push for, right? You just, it was the need for speed, density, IO speed. Do you think that that early experience 10 years ago, which is a freaking century ago in this world, led you to the aha moment that AI was going to be ultimately about speed? Like tell the audience about your aha moment where you were like, 'Hey, I'm going to do this wafer scale chip and all this crazy, you know, absurd Ferrari speed.' One of the things the internet taught you is that if you can make infrastructure fast, you can unlock new business models, right? When the internet was slow, Netflix was sending DVDs in the mail. And when the internet is fast, Netflix is a movie studio, right?
Right. Great example. There was no reason for that to happen except that the internet got so fast that delivery became dead simple for them such that they could bypass all the delivery mechanisms, right? All the other studios who had been in the way. And we believe that when you make communication blisteringly fast, all sorts of new business models emerge.
Absolutely. Well, wouldn't it then, can you tell me when the we're going to get 5G actually working? Just kidding. We'll talk about 5G some other time.
Go to Beijing. I mean, it's shocking. It's shocking. Our telecom infrastructure is disappointing. Our roads, bridges, trains, our physical infrastructure is really lagging.
It's embarrassing. Few weeks ago I was in Riyadh, Abu Dhabi and JFK and JFK looked like a third world airport compared to the others. And there's no reason for that. I mean this is New York, right?
Well, we know the reasons. It was just, it's uncomfortable to discuss them.
To discuss them. And there's no reason why our airports should be falling apart and our other nations who we think of as behind us in some sort of developmental hierarchy are building way better things than we are. And so that was disappointing. But I think to your point that the first thing is we know and we've seen it in compute, we've seen it in the internet, we knew going in that if you could make training faster, if you could make inference faster, the world would do all sorts of interesting things with it.
All right, but let's back up for a second. It's 2015, you know, and we're removed from the Lehman crazy, you know, financial collapse, right? Like we're removed from that, but it's not 2018. It's not exciting. It's not like a lot of capital flowing. One of the things you'll find if you read, if you see any of my other podcasts, which I don't, you know, recommend you do. Some of them are more interesting than others, but they're, I'll only watch the boring ones, Andy. Just send me the list of the really boring ones. I'll watch those.
Yeah. Well, I'll just try to make this one less boring. But here's the funny thing, man. You know what they all have in common is that the people I like to interview like to grind.
Because I look at this world today and I'm like, 'Okay, all these AI startups, they're so cool. Oh my god, that's so great.' You look at these blistering fast growth companies, you're like that thing could vaporize tomorrow. It's like watching an F1 car driving down a country highway with, you know, no pavement and like gravel and you're like, that's not going to, that might not end well. It might be really exciting for a bit, but it might not end well. And so I really appreciate people that grind and they really know what it takes to actually like dig out a success story. So I'm curious in 2015, nobody knows what the hell AI is doing. Everybody's doing this cloud application service providers are moving to cloud. Amazon's launching AWS around that time maybe a couple years later but like, you know, the whole world that we know now did not exist and so what were you and your founders thinking about when you decided to build Cerebras and go after a wafer scale chip when the world had forgotten about semi, all the fabs moved from Arizona and Santa Clara and Milpitas to, you know, Taipei. So like what was in your head and you have obviously a few founders that had to believe the same thing you believed. Tell us about that.
And the state of AI was pathetic. I mean guys like Fei-Fei Li who was finishing her doctorate at Stanford were identifying cats out of images, right? I mean this was the most rudimentary stuff. We were identifying numbers, right? Handwritten numbers was a big thing. So why bother? We came to believe that the trajectory they're on. When you bet on the future, you're really betting on a trajectory, right? You're betting on an accretion of small gains that quickly add up to meaningful gains. And we saw this on the horizon. And what we saw was that this type of compute could produce profoundly different results. That what it was trying to do was solve problems that CPUs and GPUs were really bad at. That our traditional compute infrastructure was terrible at. Traditionally, our compute infrastructure was bad at understanding images. It was terrible. We could store them. We could retrieve them. We could replicate them, but we couldn't understand them.
That's right. We had no context. That we weren't very good with language. Again, we could store, we could type, but what we couldn't do is understand the meaning of what was being said. And there were a few sort of madmen out there. I mean, we met in 2015 or early 2016 with Sam Altman and Ilya Sutskever and Greg Brockman and Scott Gray and, you know, they're doing this, telling me how, you know, we're just a few years away from AGI and how they have a plan and that, you know, we're going to build super intelligence and you're thinking to yourself, 'Holy, can't get my head around that.'
You might have thought they were like micro-dosing or something.
Well, you're right. I feel much better. The problem with vision is that real visionaries are very close to madmen. And you look at, you know, the great science fiction, 30,000 Leagues Under the Sea, Around the World in 80 Days, these great sort of acts of science fiction, of creativity turned out to be underestimates of where the world ended up. And clearly Sam and Ilya and Greg, they were profoundly visionary. And what we took away from this was that this work was coming. I didn't know how long it would take. I didn't know how fast it would get here, but that it would require a very different type of compute and that CPUs weren't the answer. And that we thought that the way they were building GPUs wasn't the answer. And that in particular, there were two things that would be very different about this type of compute. Number one, you'd have to move a lot of information. And traditionally when you go to a processor, you don't move information. You bring it in, you work on it, it leaves. That's not what happens with AI. In AI, you bring in information, you work on it, you get to an intermediate result, you move it somewhere else, more work is done on it, you move it somewhere else. This movement of information is problematic to traditional architectures. And we thought to ourselves, we need a solution for that. The second thing we saw was that this work, particularly in inference, required a different pattern of going to and from memory. That to generate inference tokens you had to bang on memory and it was unlike any other problem we'd seen and that the type of memory that was being used by GPUs, HBM, would be insufficient and would end up being a massive constraint.
Yeah. What happened very interestingly is we were right on both fronts on those. I mean we were wrong on a lot of other things but on the big ones we were right. One, Nvidia also came to realize that you were going to have to move information a lot. So they bought Mellanox. They bought the thing to tie together little chips so that you can move information.
All right. But wasn't Mellanox all about IO and IO speed and interconnects?
An interconnect switch. That's exactly right. So they could tie together lots of little chips. All right. So they could move information to build clusters. That's right. Now, when you build clusters, you use a great deal of power and you move information very slowly because you have to go off the chip onto a motherboard, off a NIC into a switch, maybe through another switch, back onto a motherboard. You use a ton of power because those are lasers. Extremely power hungry, very slow. We thought to ourselves, if we built a bigger chip, we could move information less often. We could store more information on the chip. We could keep it in this super fast, extremely low power domain on a chip and we would be faster and use less power and that if we stuffed this giant chip to the gills with SRAM, all right, we would be able to overcome SRAM's traditional limitations. There are two types of memory. There's DRAM and HBM, which GPUs use, is a type of DRAM and it is extremely high capacity, it can hold a lot of stuff but it's dead-ass slow. Moving information onto and off of it is unbelievably painful. SRAM is the exact opposite. Can't hold very much, but you can move information onto and off of it unbelievably quickly. And so what we thought was that if we built a giant chip and stuffed it to the gills with SRAM, we could use enough SRAM on that chip, 44 GB of SRAM, such that we overcame SRAM's limitation of capacity and we're left with this blazing speed.
And as a result, we're 40, 50 times faster than Nvidia in inference today. We're the fastest in the world on every inference workload.
Let's talk about this for a second. I actually started my career working in, believe it or not, semiconductors for guys like Wilf Corrigan and TJ Rodgers, all these legendary...
Easy dudes to work for is what I hear.
Oh my god, they were brutal. But honestly, I learned so much from them and I feel like we all stand on their shoulders and that they're forgotten heroes of tech. I mean the people that built the semiconductor industry in the 60s and 70s like Fairchild and all the Hogan's heroes and all these people maybe 1% of technology people even know who these people are and so as you know it is extraordinarily difficult to tape out any chip that has any kind of complexity to it. I remember when we had, you were struggling with the taping out of the ASIC at Force10 for example because they were like the biggest ASIC that anybody had ever seen in the networking industry and that was sort of the basis of the company but you know if you think about wafer scale that is like a gigantic atom-splitting, likely to fail thing. So that's 2015 you guys go, okay, we're going to do wafer scale, we're going to jam this SRAM, I'm sorry, DRAM, right?
SRAM. We're going to build a bigger chip. We're going to build a chip 56 times larger than anybody else in history.
How the hell did you get anybody to back this idea? Like, you know how, what high stakes this is. We're talking about tens upon tens of millions per attempt. I mean, this is why the semiconductor industry from a venture capital perspective died in like 1999. I mean, I don't think anybody funded anything past like Brio or something, but selling it to the VC community like who was your first VC? Was it Vaizra?
We took it to Benchmark and we took it to Foundation Capital. We took it to Eclipse. We took it to six or eight others and we did eight meetings and got eight term sheets.
Those three are still super active and super helpful, are they not?
They are super active and super helpful and I couldn't ask for better investors.
The best in the world, but it's not like those guys are huge funds. They're like, they're not an Insight where they can go, I write another billion dollar check to Cerebras. Like how did that work? And this is risky as hell for their partnership.
It was. I think we also had, we sort of ran the table on the absolutely most knowledgeable people in chips. Andy Bechtolsheim, Vinod Khosla, a tech titan, Bechtolsheim. I mean, every single person was a legend.
Oster is as good as it gets.
That's right. That's the best in the world. They knew the power of the idea. Now, they weren't sure we could do it. It was hard. And it involved building a partnership with TSMC, convincing TSMC to modify some of the steps in the lithographic process. It involved inventing packaging techniques. It involved, from the get-go, we knew we were going to be a system company. We weren't going to do a PCIe board. We were going to build a whole system. 7 years later, AMD says, you know, we should probably be a system company. And they buy ZT for $5 billion. We knew this going in. We knew this was not going to be a low-cost project and our investors have been along with us for the ride. After a while, Altimeter, we have had extraordinary investors. They've been wise. When you do fundamental work there are months and months where your stuff does not work where you cannot figure out how to solve the problem and you're spending millions of dollars a month and you go into your board meeting, it's not working yet, it's not working yet. And, you know, to their great credit, these were really seasoned investors and they asked about the methodology we're using and we were using good engineering methodology and we were solving a problem that nobody had ever solved. Gene Amdahl had failed at it. IBM had failed at it. TI had failed at it. The NSA had tried it and failed at it. Elon Musk recently gave up on it. After trying to buy us in 2018, he started Dojo. Just gave up on it.
Okay. So, you're grinding, grinding, grinding, grinding. By the way, I ran, you and I have a mutual friend who we won't name here. And I was at a dinner party with him because his mother-in-law is one of our dear friends and was our neighbor and blah blah blah. And he was telling me about you and I had lost track of you and he was like, 'God, man, Feldman's trying to do this like really crazy AI thing.' And I was like, 'Well, I kind of see this company in the press, but like whatever.' This is, you know, years, 5 years ago or something. And he kept telling me, he's like, 'I think Andrew is going to go crazy.' He was worried about your sanity because he was like, every time I talk to him, he thinks he's under existential threat and he's going to be out of business like tomorrow. So, you're going grinding through the desert. Okay. You got no water. You got no compass. You know, you got nothing but continuous failure and yet you guys stuck together. Did you just know you'd get there?
What you do have is colleagues who care about you and who you have profound trust in. And what you do have is a methodology. You have a belief that an engineering methodology where each time you fail, you do a failure analysis. You figure out what you got better at from the previous failure, why this one failed, and you move, you do it again.
And that's why your investors clued in on that as their measuring stick.
That's right. I think it's no different than pitchers, right? You know, you got to work on the process. Sometimes they're going to hit a good curveball, but more likely they're not. And so you dial into the process. You dial into extraordinary colleagues. You work closely with them. These are people, some of them on the hardware side who were with me since '96. I mean, we have people, in fact, our hardware team was with me at Yago all the way through Riverstone and have been with us for decades. And I have such profound faith in them. There was never a minute I doubted them. I doubted myself that I was leading them on the right path, but I never doubted their ability to execute.
So, okay. When did you GA?
Right at the end of 2019, beginning 2020.
Wow. Okay. Okay. From that day to current day.
Andy, love. We had failed. We had failed for five years in making it work. I mean, it wasn't like the market said no. The market's like, 'When do we have our part?' And we're like, 'Can't get it to work.' And one day, you know, have failure after failure, it worked. And the founders stood there in the lab. And we were in a tiny little building in Los Altos, not meant for hardware. We drilled a hole in the wall to put in a chiller. We had the windows open. You know, there was nothing about this building that was right for hardware. And the system began running. The founders stood there and we couldn't speak for half an hour. I mean, every other team that had ever tried this had failed and we were watching it succeed and it was one of the greatest minutes of my life.
Describe that feeling for a second.
You know, obviously having kids and there are lots of things that you love in your life with every ounce of your soul. This was our idea working and the fruit of our thinking. And in 70 years of history starting with John von Neumann all the way up to the greats in the industry. This had been tried again and again and we were the only ones who ever made it work. Nobody could speak. We just stood there and, I mean, watching a computer run is about as interesting as watching paint dry on a wall. We're just watching it and the monitor's ripping. So, we're watching to make sure that the calculations... It was one of the greatest minutes of my life. I mean, I can't even imagine.
Okay, so it's, I mean, just real quick, this is kind of funny. In 1981, my dad brought home an Apple II Plus. My brother and I were like, I was maybe 10, he was 12. We plugged into the TV and I never forget. I was like, 'This is the most crazy thing that I've ever contemplated.' I was young, so I didn't contemplate a whole lot.
Plus, it's a cool machine. That was my first machine.
Great. We were writing video games. That was terrible.
You probably remember playing Pong on Atari. It's incredible. If that's where you grew up with, I mean, there are a lot of games that are just fine now.
But so, so you're watching this thing spin and you're saying, 'Oh my god, our vision was right.' But it's still 2019, OpenAI's open source and there's no Anthropic. There's like, the business was nowhere.
But the technology solved. We had come up with a recipe that allowed us to solve a problem that nobody else has solved. And that at the time very few people cared about.
I think very few people did care about. But I think when you do deep research, when you do deep innovation, you are by definition ahead of where most people care, right? That's what it means. You're ahead of your vendors, right? When you do pioneering work, it's not like you can go to a catalog and get a heat sink for a part that had never been invented before. And so what we had to do is we had to meet with vendors and say, 'Look, this is what we've achieved.' We ended up inspiring many of our vendors who were like, 'Holy crap, we want to be a part of this.' And that's something that I underestimated profoundly was that you can inspire your vendors and your partners with the power of your vision as well. They want to be part of the innovation.
They want to be a part of the innovation too. And one of the field engineers from one of our vendors said, 'You know, working with you is why I became an engineer.' I mean, this is what I came to do. I didn't become an engineer to do a little bit better than the last time. It was projects like this. And so when you can inspire people, we had many vendors who did heroic things for us. And now, you know, six or seven years later, we've got three factories working. We've got yields higher than Nvidia on our wafer scale part. We are ripping. Our quality, delivered quality is better than the DGX. Fastest inference in history of humankind by far.
Fastest inference in the history of humankind by a lot. Okay, so but here's one for you. Okay, here this is fun. So now you're starting to roll and you're starting to get some early doctor customers. These are the kind of wild scientific lab type peoples, right? And then you all of a sudden the B2B world starts going, maybe one server versus 40 is probably a good idea.
Might be. Maybe it's a good idea. Maybe I don't want to have racks that are 40 feet high full of a bunch of Minix and cables and crap like that. That's kind of a 1999 thing.
So you do something crazy and you go to the Middle East. I was watching at that point because Vaizra and I do a lot of company building...
Together. I mean he does all the hard work. I just sort of try to be useful here and there. And you were one of the companies that would come up every once in a while and I'd be like, 'God, you know, Feldman, man, he's just he's seems like he's got it.' And then you do something crazy. You go to the Middle East and people at the time were always like, 'Oh, Middle East, it's just about Saudi Arabia and it's maybe UAE or whatever.' And all of a sudden today, I was told, I don't know if this is current data, that it's still by far the number two market in the world after the United States. Is that correct?
Okay. So, what made you do something that's like I heard you were on a plane like once a week back to the Middle East. I don't envy that at all. But what was driving that decision? Because that's another completely whacked nonlinear non-traditional file process.
Well, first, Emirates flies an A300 out of San Francisco. It's 16 hours. It's a beautiful plane.
Beautiful plane. If you can fly on that, I highly recommend that.
I did it once a month for more than a year.
And look, I met the CEO of G42. His name is Peng.
Yeah. But why did you go in there the first place? I guess that's what I'm looking for. Like what were you thinking?
We were looking around the world for people who were working on interesting AI problems and we knew that there was a project in the UAE to train in Arabic. Arabic, like many other languages, we call under-resourced, which means that if you scrape the internet, which is what the big US players did, you don't come up with a lot of Arabic. And that means that you don't have a lot of tokens, you don't have a lot of data on which to train your models. And we heard they had a program, we reached out, I met the CEO Peng, and I flew out to meet him. And I found a kindred spirit. I found somebody who had a monstrous vision. I was honored to meet his chairman, Sheikh Tahnoon, and I found Inception. I found the largest and most compelling and sort of coherent AI vision that I'd ever seen. And we began working with them and showing them how our products could solve problems that were of interest to them. They had a problem that they'd been working on for quite some time. We solved it in a few days and we started a partnership that is now sort of fundamental and strategic, a cornerstone of our work. And through this partnership we have delivered hundreds and hundreds of millions of dollars worth of equipment. We've trained the leading models in Arabic. We've trained genetic models for them. We have collaborated with them on a range of different projects. It has been one of the most productive and rewarding partnerships that one could have imagined. And you know, I didn't know anything about the UAE and I love Europe, obviously we're of European descent, many of us in the United States, and who doesn't love to go to Europe. But if you think about the Middle East leapfrogging Europe in a lot of areas of innovation using products like this, by far you could see a sea change in industries because as you know they need to diversify from energy-centric to resource-centric yesterday.
