transcribe

Andrew Feldman on Building Cerebras and the Future of Chips | Ep. 57

Uncapped with Jack Altman · 1h 2m · transcribed Sep 2026
More from Uncapped with Jack Altman Business
𝕏 Share ▶ YouTube 📥 PDF 🤖 .md

Section Insights

# 0:00

Cerebrris' Journey and Initial Challenges

What were the initial challenges faced by Cerebrris?

Cerebrris faced significant challenges in its early days, particularly in board meetings where the recurring theme was the inability to meet expectations. Despite this, there was a strong belief in the potential demand for their product if they could overcome these hurdles.

  • Cerebrris struggled to meet expectations in early board meetings.
  • The team believed in the potential demand for their product.
  • Support from the board was limited in technical problem-solving.
# 12:32

Innovations in Chip Design

What are the key components of Cerebrris' chip design strategy?

Cerebrris focuses on three fundamental elements of computer architecture: calculations, memory storage, and data movement. The company aims to innovate in all three areas to enhance performance and efficiency in AI applications.

  • Cerebrris emphasizes calculations, memory, and data movement in chip design.
  • Innovation in all three areas is crucial for AI performance.
  • The company is committed to making chips faster and more efficient.
# 25:05

The Role of the Board in Hardware Companies

How can a board contribute to a hardware company's success?

A board can provide valuable insights and support in areas like financing and strategic direction, especially in long-term projects. Their ability to see across the industry can help guide the company, even if they lack technical expertise.

  • Boards can offer strategic insights and support in financing.
  • Their industry-wide perspective is valuable for long-term projects.
  • Effective product management is challenging in hardware due to customer feedback limitations.
# 37:38

Building Trust and Relationships in Teams

How important are relationships and trust among founders?

Trust and relationships among founders are critical for success. Working together previously can foster respect and understanding, allowing for effective collaboration and minimizing conflicts.

  • Strong relationships among founders enhance collaboration.
  • Trust can be established quickly through effective communication.
  • Respect for each other's skills is vital for team dynamics.
# 50:11

Future of Data Centers and AI Development

What are the future challenges for data centers in AI?

The demand for data centers is expected to continue growing, with challenges in predicting future needs. The landscape of AI is rapidly evolving, making it difficult to foresee the exact requirements for infrastructure and resources.

  • The demand for data centers will persist in the coming years.
  • Predicting future needs in AI infrastructure is challenging.
  • Government involvement in tech and AI is increasing but may lack informed decision-making.

Transcript

0:00 You got board meetings every six weeks and all you've got to say is still can't make it. Right? Is that that's the board meeting? I mean, what else are we going to talk about? >> Still can't make it again and again. >> And they're like, "Should I get in? Can I do anything to help?" >> That's right. They're like, "Can we help?" It's like, "Nope." But we always believed that if we could make it, there would be huge demand for it.

0:24 >> All right. Super excited to be doing this today. Andrew, thanks a bunch for for being here. So Andrew, you're the, you know, CEO and founder of Cerebrris, which is obviously one of the most important chips companies. And also really happy to have my partner Eric here, who's going to be kind of half host, half guest, who, you know, has been working with Andrew since the beginning. But thank you both for for making time. What I want to start with is it's 2026 now and obviously Cerebrris is very important in the sort of AI landscape but you started the company in 2016 and AI was not what it is you know now back then so I guess what was the kind of headsp space then what was the idea the insight like what were you what were you building when you started the company and what was the thinking >> well Jack thanks for for having me it's always fun to hang out with Eric and so appreciate you having me I think as as a computer architect when when you see new a new workload on the horizon you get excited right it it's very very difficult to win share in a mature market in compute and so when when something new emerges you get excited and you lean forward and you say well can I make it faster can I build a chip that that that is better at this work and then you ask the other question is is there enough of this work to justify building a chip that's better? So you have two questions. Can I and should I? A and what we saw with AI was an extremely computationally intensive workload.

1:56 Unlike for example the the rise of ARM processors for cell phones that they weren't computationally intensive, they were power intensive. Right? We we saw a problem that was going to be hard on compute and we we we saw that that what was being used was an architecture that was that was being re remodeled for it and that that we could build something better. And that was sort of the the insight and then it was a a long hard slo.

2:30 Before we go to the long hard slug, Eric, obviously for you, like as an investor, you know, you were not, you know, also in this AI era, but you saw something that made you invest and you were, I think, you know, previously not a chips investor and you done more like >> and probably never will be again. >> Probably never will be again. Like, would you do that every time again? Was there something that made it clear to you or was it like what what happened for you when you made the investment?

2:54 I I I think one of the scariest things about venture and I think about this all the time is the more you work on companies and you see the challenges and you see how hard it is you build up scar tissue and then the thing is like well you should do it again but like will you do it again and I ask myself that all the time like there you actually need some naivee like you need to have this like oh can do attitude and this like naive vet around it to attempt it like otherwise you just don't do it again. And I've definitely built up a bunch of scar tissue. and I don't know had I actually had any idea how hard it was going to be for them, like how much science and technology had to be built, how many innovations, how many times like the company was going to come to the point.

3:50 >> Don't tell your VCs how hard your shoe is. And the moral of the story is tell them it's no problem. Keep it inside. Push it down. Push it down. Just don't don't tell them hard problem. All's good, man. >> Yeah. So, what happened? So, you basically, you know, you you start the company, you get to the tape out, and then you're done, right? It's like that. >> No, never. >> So, how's it go? So, what what happened?

4:10 >> I think we we had you know, this isn't our first chip company. This this is my fifth. and the team had been together the founding team had been together at the last one and we we had some really clear sort of philosophies on on on what you should do and I our view is that if if you're going to attack Goliath if if there's a a giant standing in the market like Nvidia was even at that time that being a little bit better or a little bit cheaper is is not an available strategy right they had high margins so if you come in and and Even if you're two/ird the price, they can just cut cost. They can just charge less. They can bundle it. They can do a hundred other things. What that means is you have to go out with something way better, 10, 100, 500 times faster. You have to come up with a product that has a value proposition that that even if the other guys give it away, all right, they can't compete.

5:13 And so that that's sort of one direction of our thinking. And the other building off that is is we want to do hard things that produce that advantage, right? That we want all that hard stuff within the building because that's under our control. All the other stuff's not under our control. But if we can build something that is so fast nobody else can do it, you can't give away your parts to achieve what we can achieve.

5:42 then you're on to something. And the only way to do that in our opinion is radical innovation. >> You you can't incremental your way to vastly better. >> So what has to be in your control? >> You you have to have all aspects of the design, the implementation, you have to understand the manufacturing, you have to understand every part. And for us that meant chip, board, system, software, all the way up to the API.

6:13 And that was not easy. That makes it more expensive. It makes it take longer. It means there are more more ways to fail there. When you do radical innovation, there are no vendors waiting for you, right? When you build a a chip the size of a dinner plate, you can't go to a catalog and find a heat sink, right? Because nobody' ever built one bigger than than a postage stamp. So nobody has stuff ready for you. And so you you end up investing an enormous amount of time in in building things that support your thing, right? The the surrounding components. But the result of that is is you develop this extraordinary expertise. We didn't start world leaders in packaging. We're right now the best in the world of packaging.

7:00 we earned it failure after failure year after year until we got it. And so what you want to do is you you want to think about sort of h how you can innovate across the the the various elements of of a full solution and push as hard as you can. >> I I think like Andrew taught me this in in in semis and everything else just to like make this like what you need to be shooting for concrete.

7:30 So if you think of whoever the incumbent is in whatever market you know they're getting say twice as good every year in a market like this is very dynamic for a new company to get to scale to start to get to scale it's going to take five years at least let's say with everyone moving as fast as it can AI tools you nail everything first tape out works first bring up works you know everything goes right five years okay to to like to get to scale and so you're basically at two to the fifth. So that puts you at 32x and then you need at least a multiple advantage of that. So you say 3x. So you're basically at 100x.

8:10 So one of the things that happens when you hear a lot about these new ideas is they're like, "Hey, we're going to be 50% better. We're going to be two times as better." But what they are, they're like, "Once we have this design, it's going to be even 10 times better than what exists today." But what exists today isn't the target. Because if you're looking at the companies that are out there right now, these big companies with a lot of resources are actually doing they're they're doing great work. Everyone's Nvidia is advancing the ball materially every year like very significantly.

8:42 >> Nvidia and Google and Tranium, they're moving the ball. And so you got to aim at at 100 500,000 times better if you're going to arrive. >> Yeah. if you're going to intersect the market ahead of them. >> And so that that just like when you realize that you're like, "Oh, wait. It can't be small changes in something. There has to be a very big underlying like architectural change, a big thought like there has to be something materially different whether it's wafer scale, es whatever it is like there those things matter. You >> you can't get there with a collection of modest improvements."

9:22 >> Mhm. >> Right. your your your biggest competitor buys silicon for less than you. They buy manufacturing capacity for less than you, right? They they have they probably pay less for their EDA tools. And so you can't go at them, you know, you can't run it Goliath straight on. I mean, you got to think about how you can deliver something profoundly different. >> What were the like along the way after some of these like early moments? I know that the company had like, you know, some near-death experiences or whatever.

9:52 Like what were the what were sort of the hard hurdles? >> I had some near-death experiences pain in my chest when when you can't build the thing you're supposed to build. Yeah. >> Like what are those per like how would you frame what those periods are where you're like we just need I don't know is it a certain amount of time and a certain amount of capital and it's just going to be a valley of death? Is it that you just don't know whether you're going to make a technical breakthrough?

10:15 Like what are the things that lead to those? I I think in in our space we were always confident if we could make it we'd sell it right that that there are sort of two axes in our life there's can you make it and the other axis is can you sell it right and we always believed that if we could make it there would be huge demand for it I think nobody had ever done way for scale there was a a whole sort of collection of people who were saying it could never work and we were unsure too. I mean we believed we could do it but there was no evidence and there was a period of time about 18 months and we couldn't make it and we're spending about 8 million a month. You got board meetings every six weeks and all you've got to say is still can't make it right that that's the board meeting. I mean what else we going to talk about? I mean, still can't make it again and again and again and and I I give a hu huge Right. Right.

11:22 >> And they're like, "Should I get in? Can I do anything to help?" >> That's right. They're like, "Can we help?" Like, "No, >> there's nothing nothing that can be done. Nothing to be done." But >> we believed because we we had ideas still. We hadn't run out of ideas. And each time we we built it and it failed, we'd go through sort of good engineering practice. We'd do a full failure analysis. We'd understand it. We would and we wouldn't fail that way again.

11:54 >> All new mistakes. >> All new mistakes. So we we sort of had a that that was sort of our mantra. Only new mistakes, right? Only new failures. and over time we we could see progress. I mean, in the beginning, we were shattering wafers, right, in seconds, and then it took minutes, and then we had one run for for an hour, and then we shattered some in minutes again, and then we we we built back up to the point. We had a day in sort of July of of 2019 where we were running and temperature was was flat and we just stood there in a tiny little office that had been converted into a lab and we drilled a hole in the wall to suck the air out and stared at a server which is about as exciting as looking at at Paint Drive.

12:43 Said, "Holy crap, we've solved this problem that nobody in 75 years of compute had ever solved." And that was one of the great minutes of my life. >> That's cool. When you look forward like if you so you know there's the wafer scale and then everything around that the packaging and and cooling and powering it all these things. But you know when you look at the R&D envelope that you have going forward over the next five years, >> right? I I know there's like things you're excited about, but like where do you take like where do you take this like giant thing?

13:23 >> Like what could be the next wafer scale innovation? >> I I think if you simplify what we build down to its most fundamental elements, a computer is built of of of three things. we do calculations, we store the results, and then we move the results to where they're useful. So we we build a core that does the calculation. we use memory and we move data to and from memory where we store the results and then we have IO and that that that's how we ship the results to somewhere where it's useful.

14:05 I I think if if you're in the computer business like we are for AI, you better be working on all three, you be thinking about how to make your course faster, how to make them tuned for AI, but general enough to to withstand the innovation happening in AI. You better be thinking about memory and both capacity, how much you can store and how fast you can get data on and off it. And then you got to be thinking about how you get the results off your chip and somewhere where they're useful and that's your IO. And so we're working on all three of those.

14:41 And we have programs with the with the US government already, big programs where we're thinking about how to stack memory, right? How how to put HBM onto an SRAM based wafer. And what this does is it it enables HBM to behave like SRAMM which is exactly what everybody wants. You get the the capacity of HBM and the speed of SRAMM, right? We're we're working on on optical wafer stacking. So you put an optical switch onto a wafer and this would change the world. you know Jack Dungara said and he he's sort of one of the the pioneers in in big computing. He said, "We've we've been better at making flops and moving flops, right?" And that that that's that's the IO part. And so thinking about how to how to solve that problem by bringing optical switching smack up against compute is something that that we also have large government contracts for and we're enormously excited about. So, you know, you got to get faster. You got to find ways to store more and get to memory faster. And you got to find ways to move that move those results at 10, 100, a thousandx faster to where they're useful.

15:59 >> I want to go to a bit broader in the supply chain. And you know, I I think it's kind of like well understood right now that AI is very supply constrained, but I like probably many other people, I wouldn't say I have like a perfect understanding of the supply chain. And one of my favorite things doing on this podcast is you get a, you know, successful person like you and because there's cameras on, I get to ask you a really simple question and I get to, you know, be humored with it. So could you explain sort of in a somewhat simple way how do we go from like sand to a chatbt answer?

16:34 >> Dude, it's just TSMC. That's it. There's >> there's a black box. You call that TSMC by another black box called SML. And out the other end comes comes great. >> We're supply constrained. But I don't know what the Can you teach me what the supply how it works? >> Yeah. I I I think the first thing to to to think about is that that that a fab and especially a fab that that builds at cutting edge geometries is a is a modern pyramid. It's one of the greatest things humans make.

17:05 >> What is it? It is a collection of machines that that take a chunk of silicon and use a photoiththographic process in which they etch transistors into that piece of silicon such that when you deliver power to it they do calculations and it it is a collection of I mean this is a factory it's just a reasonable way a real factory cost 40 or 50 billion to make it has a four or five year lifetime.

17:36 >> How big is it? >> Football fields. >> Okay. >> When they when they when Samsung was building a factory in Texas, you they began building a power plant. The power plant was used to make concrete. They ran concrete trucks, hundreds of concrete trucks, 7 by 24 for years to pour enough concrete to build the foundation on which to put the >> the FAP, right? These are unbelievably complicated things. >> So, ASML makes the machines. ASML makes makes the the machine that does the photo lithography.

18:11 >> what's that machine like? >> It's the size of each machine is what is it 50 or 60 feet long and 20 ft high >> costs what half a billion? >> Yeah, they're they're expensive. what's interesting is that they sell the same machines to different fabs and fabs use them in different ways and are able to do different things with TSMC able to to to to achieve things that others can't. >> And can anyone else make these mach is it really just ASML that can make these machines?

18:42 >> Right now, it's only ASML >> in the world. >> In the world. >> That's weird, isn't it? >> Yeah. >> Yeah. I mean there are very few things that that this is a true monopoly not born of you know what debeers did which was try and control supply. They have technology that others haven't been able to to replicate. >> Are we supply constrained on lithography machines? >> probably not. I I think the the challenge in our business is that that demand moves extremely quickly. And if you've looked at sort of a great company like Nvidia's demand, it's it's exponential. But you you can't build fabs that fast. If a fab takes 5 years to build, >> right, and you see demand increasing, there's no way anybody would have expected demand to to shoot up the way it has for chips, they're always behind.

19:34 >> Yeah. and they're investing huge capital blocks in in in the the construction of these facilities and demand is is racing past the ability to make these huge 5-year bets and we have the same problem right now with data centers, right? AI is moving at the speed of software and data centers are moving at the speed of real estate and that that's why we're behind. That's just it. >> Good articulation is good. Okay, so TSMC buys the lithography machine, >> right? So TSMC buys lithography machines. They have decades of experience. They they wrap it together.

20:09 They take an ingot which is a a tube of silicon. It's sliced into discs we call wafers. The wafers are sent through the process. The process etches through a lithographic process transistors into the silicon. It moves through the process. They're cut. we call it diced or in our case they're not cut. Out the other end comes chips. And what's amazing about this is you take sort of some of the smartest people you've ever met and you put them on a project for two years and you send the the the the results to TSMC and they run it through a factory that cost 40 or 50 billion to make and took years to make and out the other end comes a $22 chip.

21:04 It blows your mind when you think about it. The each one of these took sort of the smartest people in their field. >> It's the most impressive thing humans have ever done, right? It's among the most impressive things. >> And a burrito is $23. >> That's right. Right. Exactly. Right. And you go down the street here in the Mission District and you pay $17 for a burrito. That's right. Right. And so, >> and once you have the chip, you're done.

21:26 Right. Then you're good. >> No. Once you have the chip, you've you've solved about a a fourth your problem. >> Okay. But still stay TSMC. >> Yes, >> they operate the they they receive a design, they operate the machine, they give you the chip. >> Yes. >> But that itself seems like not replicable. Like we don't have fabs like this in the US. >> We have them run by TSMC and by Samsung. >> So what >> and Global Foundaries have got one too.

21:50 And >> how much of the bottleneck is at the TSMC level? >> Huge amount. >> Okay. Why huge amount is that? Because of the speed of building fabs. >> It's because the demand has outpaced the the ability to build fabs. It's not like it's it's not like a an apartment building where there are hundreds of builders you could choose to build your apartment building. The actual making of the fab is skills that only TSMC has. And so and even there they can't do 12 at once, right? that the the skills to make these things are so rare and in so few hands that that you can't just knock them out. It's not a cookie cutter and and that's why we we don't have enough of them right now. And the fact in the US sort of three decades of bad policy pushed the fabs away. When the fabs left, the tool vendors, the the collection of of vendors who who provided services to them, they all left. The next step in the process called packaging, those are companies like Amcor and ASC. They left >> and we just punted a strategic industry and we we got to do better. That that was that's not smart.

23:06 >> Yeah. So, what do you think should happen there? like like >> well I think I think we should sit down with Global Foundaries and TSMC and Samsung and have a 20-year period where we wave all local ordinances to allow them to build fabs. I I I think this is about building US domestic fab capacity. >> Yeah. I I think I mean look when when a a big ship right tried to parallel park in the US canal right we were delayed in chips and we couldn't buy washing machines.

23:41 >> Yeah. >> Right. I mean if we lost our chip capacity it would be catastrophic for our for our industry. Not just for our industry but for for the country for the economy. Eric, how did you as a board member when you were like learning this stuff as you went, you still obviously, you know, I don't think Andrew's been lying to me. I think you've managed to be very helpful. Like what? >> Eric was extremely helpful.

24:04 >> So like how did you approach this? >> I have a very small circle of competence. It is not any of the stuff that he talked about. If he if he if you asked me to explain what packaging is right now, I could not do it. I certainly could not do it in an adequate way. And Andrew explains things to me all the time. and you know so I think it's important just to be like hey this is where you what you can do and this is what you can't do you know and there are periods of time in any company particularly if you're doing technological innovation where there's just like there's just you just have to let the engineers engineer and like and the scientists do their thing and and and stay out of the way and keep them financed maybe keeping him financed was very important I guess for >> I think there are a couple things that that made Eric a good board member And I think that are foundational in being a good member, right? You're you don't know about everything and and share and make us better in those domains where you're a real expert and don't talk about those other domains, right? That sometimes there a lot of words and and we were lucky. We had a really good board and everybody knew what they were good at >> and they they helped us in those domains in which they had real expertise, right?

25:17 Right. I mean, one of the the advantages of of of the venture world is you you can see across an industry, right? When when we're deep in it, we're going deep. They can see wide, right? But they weren't efforts by the board to try and solve technical problems. They they didn't have that that wasn't their their expertise. talking about how we might finance the company, thinking about all sorts of other things. They were enormously helpful and they were patient and I I think they asked thoughtful questions and our board members our our board meetings made us better.

25:52 >> It's like what what what what should a board do in a hardware company in general when you know you expect that there's going to be much less to contribute on the product? Is it financing and recruiting? And >> I I I think the the question when you've got a long hard project, right? And this is the opposite of SAS, right? you you raise money and then you spend two or three years to build one before you have any real idea what the customer is going to say. You go and you talk to customers and they say, "Yeah, yeah, that sounds great." Because who's not going to say, "Yeah, yeah, that sounds great." Right?

26:26 And so the product management is unbelievably difficult. You survey customers and because it's no cost and it's easy to say, "Yeah, yeah, this is great." Nobody wants to to to sort of put their foot on the throat of of somebody else's idea, right? they say, "Yeah, yeah, that's great." It takes you three years to to to two and a half years to get to the point where you can bring it to a customer. I think what you can do as investors is understand that trajectory. Understand that your first chip is very rarely a good one and it's a second or third one. I mean, even really strong teams like Google's TPU team, the fourth one was good.

27:00 >> The first two were the fourth or fifth were really good parts. It takes years. You have to to know that going in. You have to know it's a long game, right? ask questions related to the long game, right? Are you hiring the right people? Right? Are you thinking about this in the is it have you made the right decision between specialization and flexibility, right? These are questions that that can really help sharpen our thinking. But you know w whether to use this one technique or this other technique in in in design whether you use one tool vendor or the other you got to let that you got to let the team >> team pick that.

27:48 >> Did it feel really different for you versus like an infrastructure >> totally role? >> Yeah totally. It's just really different. I mean your time to revenue is so much longer. The revenue comes in like giant chunks. It it the initial customers you know we had US government customers then we had you know huge sovereign cloud and G42 you know then you have now today obviously open AAI and so you have these like things so your customer concentration is more >> just like every dynamic is different the investment in go to market so much less so I think it it feels really really different the things that are the same by the way are You need really good people from a whole bunch of domains coming together who stay motivated over a really long period of time. That part's actually the same and and then the market is very dynamic like I mean think about it pre-transformer you started the company pre-transformer >> we did >> right at that time tensorflow was dominant so the first >> flow was dominant resnet was was still a thing >> right >> the these were very small very simple networks compared to where we are today yeah the the actual ML the AI looked nothing like it does today.

29:07 >> Well, to actually I think it's really, you know, for entrepreneurs who are who who listen to this like >> it was started as a training system. >> Like part of the vision that you pitched and part of what you explained was like training is a harder problem because of back proper you thought was going to be on devices and computers and really distributed which may still end up happening. And like fast forward to to 2022, 2023, maybe 24, we you pivoted the company >> or I don't know if pivot is the right word, but like to to inference. So where h how and why and where did we get lucky and where did where did you have real foresight on that? I >> I think there's sort of there two ways to to to make decisions and to to think about it. one I think about like an old telephone circuit, right? You you you they they set up a a dedicated view all the way to your brother New York and that that's one way to do vision, right? That that that's one way that that you've got this idea of the way the future's going to look way out there. And that's a circuit we we think of, >> right? the other way to do it is the way routers work with the internet works where you go hop to hop. You go to Cleveland and and then there's a decision whether it's best to take the next hop or and that's sort of the way we worked where we we knew there was a pot of gold out there, but the path to it we knew was was unknowable.

30:43 And so each time you get over a new mountain, you look around and you redecide. So it's it's so this sort of hop to hop thinking. So can we can we build it? Yes. Can we make it work in routing? Yes. Is routing now where everybody is focusing. That's where they're focusing. Can we be a fast router? Yes. But now we're we're seeing the lay of the land. We're seeing the sort of unfolding of AI, the rapid growth of intelligence.

31:12 Well, who's going to use it? We use AI through inference. So we we were in a position we're engaged in conversations. We had product being used where it allowed us to see something new and that new thing was holy cow the trajectory of AI will make us smart enough that everybody will want to use it. If everybody wants to use it inference is going to crush >> right the compute infrastructure we better be there. And so each time we achieve something it it sort of moved us up a a mountain or a hill. It gave us a new view of the landscape. we could make some new decisions and we were always sort of moving in the same direction but we didn't know how to how to get there.

31:54 And so decision hop you you get to a a point of view think carefully earn the next viewpoint. >> Yeah. >> Earn the next >> Well, it's interesting because you know with software you obviously can both like think nimly and change everything the next day. >> Right. >> Hardware you can't do it like that. That that's right. O ours we we have bigger discrete decisions, right? And that's why you the choice in your chip of specialization versus flexibility is so important. We we made a couple really good decisions in in our first architecture where we decided not to sort of embed technology that that would accelerate convolutional networks.

32:45 Instead, we said we don't know how long those will last if we work underneath that and accelerate the the algebra that underpins all AI we knew about. That was a really good decision because when Transformers came out, we were the fastest at those two, even though we'd never seen them and never heard of them and they hadn't been invented when we set the architecture. And so it it helps to make a few good decisions. >> Yeah, it's interesting like on the you know on the backs of you know companies like SpaceX, Andreal, Cerebras, you know, Palunteer, maybe others, but obviously like hardware is now hotter than hot and everyone wants to fund hardware. But it's it does just talking us through it is crazy hard.

33:28 >> It's really hard. >> Yeah, >> it it's hard and it it requires enormous internal fortitude and it requires success has historically been predicted by some experience in the field. >> Yeah. >> >> different than like you know AI where you see the advantage. I think in look in in AI and social networking a lot of the the founders and the leaders were building tools for themselves and their classmates and their friends and there they had unique insight. I mean if if you look at cognition if you look at cursor these are some of the best engineers software engineers in the world. building tools for themselves, right? That's true. And that's different from >> I also wonder if this last topic probably plays a big role in it, which is >> when you can change your opinion the next day and it's fine that you were wrong yesterday. Speed and decision-m speed trumps versus if you just, you know, you do something and then you have to live with that for a year, >> years. Years.

34:29 >> You got to you you can't just be wrong and fix it tomorrow. I I think in in chips we we we measure three times and cut once, >> right? We it is really unforgiving. >> Yeah. >> To big mistakes. And so you have to be very very sure that that you you get the big items right. >> Yeah. One of the things that I think is interesting about how you built the company, which is, you know, you and you mentioned a lot of the early team had have done multiple companies together.

34:59 Like a lot of you guys had had worked together for a long time. >> you're older and and more experienced. >> I think you're not that old, >> but somebody called me a boomer. I'm just like, my dad's a boomer. A boomer. >> You know, at some point you're looking up and that's just it's all the same. when you're that low, you're looking up. It's just like like when you were eight, you really old, like 27.

35:24 >> Well, but but you know what's interesting actually is if I look at the product leaders in your organization. >> If I look at some of the go to market leadership, >> yeah, >> these are they're actually very young. >> They are. >> And and I'm just interested in like how was there intentionality around that? Not necessarily agist obviously, but just like how you combined the perspective of you know young at the cutting edge product leaders with seasoned hardware engineers. I >> I think we tried to think really hard about where experience mattered where blistering intelligence matters. you know, if you look at at our product organization, unbelievably smart. I mean, some of the best product people I've I've ever seen. all young, all promoted from within. you know, we we we don't have big company rules. You got to be in a job for this amount of time before you can get promoted, right? I mean, if you're extraordinary, we're going to give you more and more responsibility and more and more and if you do a great job with it, there's no end to where we will we will take you. you know, my co-founder Sean, you know, one of the five of us, I mean, in my last company, we we hired him as an individual contributor and when we were acquired four years later, the guy was 25 or 26 when we when we hired him in 2007, right? when we were acquired by AMD, we made him a a corporate fellow, right? That there was just no end. I mean, and now then he is a founder and he's co CTO of a of a public company.

37:02 >> and we we do that sort of ruthlessly. and there are some areas where to be exceptional requires a tremendous amount of experience. There are some areas where it doesn't require any experience. we we we don't have a long history of of understanding what customers want in AI. So their methodology, smarts, insight trump experience. In other areas in the making of chips, it's been my experience that that that some odd years of previously building chips is the best predictor of whether you're going to deliver exceptional chips. And that's also true on the mechanical side, on the system side.

37:44 >> There's just not a lot of chance in college or in graduate school to actually build silicon, to actually build a machine. They're so expensive. They're they're so hard to build. they take so long that even in a doctoral program, you don't get a chance to tape out a chip, to deliver it, to bring it up, to put it on a board, to power it, to write the software for it. And so it takes some time. Are are the relationships something that are critical or can those be earned quickly by young people?

38:12 >> I I think the following. First, five is too many founders without question except that we'd worked together before >> and so everybody knew what the other other folks were good at, >> right? And and so there there wasn't a lot of headbutting at all. I mean, we'd all worked together previously. We all had tremendous respect for what the others could do and and some humility about what we couldn't do. And so we we were able to to to do that. your specific question is does it take a lot of time to build trust? I I think within the six or eight weeks of working with someone, you can tell if they're extraordinary.

38:48 >> Mhm. >> Right. I mean extraordinary people in the first email they send you, you go, "Whoa, that's exactly what I needed." >> Yeah. >> Right. Every list is in descending order of importance. There's not a lot of fluff. There's high signal. There's You go, "Whoa." And then you see that again. And then you watch the way they run a meeting and you go, "Whoa." And then you watch them deliver something. They motivate a bunch of people around them.

39:10 They can work across the organization. You say, >> "Well, that's a person I need on the next important project." Yeah. >> And then they they just they chew through work. >> I always thought with recruiting, it's like if I came away like learning real things and I wanted to have another interview or meeting just cuz I was like, I'm going to learn more stuff. I was like, that's my best prediction. >> The best. >> It's so good.

39:30 >> The best. >> What about the external relationships? It's like, you know, you got to work with a lot of people outside your company. You know, you got to be in you got to be in Taiwan. You got to work with, you know, >> you know, it helps to bring those with you a little bit, some of them, right? So, you know, we'd been been building chips with with TSMC for decades. We'd been working with our contract manufacturers for decades. And so, especially in a time of contention that that that that those relationships had had been in place for years, that you'd been good to your word, not once, not twice, not just in good times, but you you'd been good to your word over good and bad times. That was really, really important. I I think you earn relationships with new with new partners in exactly the same way. You you're good to your word. you get them information early. You write I mean it's, you know, write a thank you note.

40:26 No, really. I mean, be do do what your mother said, right? Be a good person. Write a thank you note. Do what you say you're going to do. >> this this to me was one of the big learnings like particularly during the COVID era and everything. If I think of the software companies, even the software infrastructure companies we work on, it's really like their vendor that matter is AWS and like maybe now their vendor that matters is AWS and you know a foundational model company or something like that. There's like two vendors that matter. This is like I don't know what the vendor list is but it it's >> it's insane. It's dozens, right? It's it is the all of these components that go into the system. All of these specialty manufacturers that build the cooling plate and part of the water system and like and so you have all of these things that have to kind of come together and and we just aren't used to that like that those me that many dependencies in order to deliver our product or at least >> I wasn't that was like a big >> thing and you have to kind of work with them on >> software guy coming to grips with a supply chain.

41:32 >> Yeah. Yeah. Yeah. >> Yeah. Exactly. >> They call it a chain for a reason. >> And and and it really is. And you you're like, "Wait a minute. Holy cow. This all of this stuff." And when you're growing exponentially, like then everything becomes even more complicated, right? Cuz it's like you're putting in I had never thought about this. You know, it's like AWS, we want more capacity. You know, you go online and you add capacity. like you want more capacity on in terms of like number of wafers like that's a 15month lead time kind of decision >> and a huge amount of capital and your vendor is also making a huge allocation decision and so they have to buy into it too and so I I that that was a big learning for me in just terms of like how much complexity there is in that >> one of the things we've been talking about a lot internally is that it feels like no matter how big you think it is, it it is hard to wrap your head around the size of the AI buildout that's happening.

42:34 >> Yeah. >> And the capex going in, you know, I actually saw a chart this morning that was like 1% of GDP >> was spent per year on >> highway and telecom, >> right? >> And then it was 2% for railroads >> and then this AI buildout is like 3 and a half% of GDP. >> It's like >> only one person got it right, >> which was >> your brother. >> No, that's true. >> Yeah. I mean, the only person and everybody thought he was out of his mind. I mean, what what what what Sam's really good at and and I think is so hard is he saw an exponential and wasn't afraid, >> right? They you take that exponential out two or three or four or five years, you holy crap, >> right? He wasn't afraid. Everybody else was afraid. Every now it's going to slow down. It's going to He was like, no.

43:21 >> Remember the Stargate like 7 trillion or whatever, >> right? I mean now now the initial Stargate is is is sadly small, >> right? And it was mindboggling. They were laughing. >> And so all of us got it wrong except him. >> I think we got it wrong. I think TSMC got it wrong. I think Nvidia got it wrong. Everybody, the memory guys got it wrong. We all got it wrong except Sam. >> Yeah. So, you know, you've got this insane buildout going. And, you know, the the demand for the demand for inference seems like it's going even faster than no matter what we're going to be able to build, you know, and so then there's a lot of questions about, you know, what will that mean for the price of compute and what will it mean in terms of how much will be available and token spend and all of this, but I'm curious like day-to-day what you're seeing on the data center side and your experience, you know, now what what is that like for you?

44:15 The data center is one of the the links in the supply chain to deliver compute by the cloud. And the truth is it's whether you deliver it by the cloud to via a cloud or you deliver it what we call on prem it impacts both case. In one case you rent the capacity and you put your your equipment in it. In the other case your customer rents the data center. And so it's that it is a a limitation in the market right now.

44:44 it exposes a whole bunch of weaknesses in the US. our grid is pathetic and sort of built on 1940s or 50s technology. we we stopped doing work in in very interesting technologies that turn out to be extremely clean like nuclear. I mean wouldn't it be ironic if it took AI to get bring us back to doing nuclear, right? I mean that's right. Pretty right. it it showed that you know in these data centers we use generators as backups and these are either diesel or or LNG or one form of gas right and these come from G Vernova or they come from Caterpillar there've been no innovation in in generator and diesel gen sets for for decades now suddenly there's innovation the guys at boom want to use what they designed for jets to to to power data centers. We're seeing all sorts of interesting things in in battery backup and like from Bloom Energy, you're seeing interesting fuel cell. I mean, there's just this enormous innovation be because there's necessity driving it. We don't have enough data centers. They're coming on too slowly. the grid is old. and so we it's an area of tremendous innovation. I I think the industry did itself no no favors by doing some dumb stuff at the beginning.

46:13 Tried to pawn off some costs on local communities. It tried to take advantage of local municipalities. There there is no reason a data center shouldn't pay its way. There's no reason why it should use very much water. We use closed loop systems, right? All the data centers in the US use less than California almond growers. Not by 1x or 2x or 4x, but between four and seven times, the almond growers use more. So, I mean, >> the water thing is just not a thing.

46:39 >> The water thing's not just a thing. Yeah. >> But as a community, we didn't do a good job of communicating with local communities, getting their buy off, showing them that we're going to bring thousands and in some cases tous thousands of high-paying construction jobs. We're going to pay ongoing jobs, and their tax base ought to go down over time. And we didn't do a good job of that. And now we're paying the price. >> I mean, it kind of goes with the whole theme of tech doing a terrible job communicating about AI in general.

47:06 >> Horrible job. We're doing a horrible job. >> Yeah. >> I think the data center thing, the other part of the data center thing is interesting to me is like if you if I asked you, I don't know, 2018, 2020, whatever. What's the probability that data centers were going to be a critical factor in our >> Another thing I'd have gotten wrong, >> right? Yeah. >> We'd all got it wrong. >> Got it wrong. >> And it just And that was just like a it's a new thing. It's like, okay, now you have to build it all the way through and deliver it. But >> how long does it take to go from shovel?

47:33 I know you you joked about shovel. I actually want to ask you about that, but how long does it take to go from start to finish on a data center? >> So shovel ready is an expression you hear in the data center world. It means we haven't done >> but we're ready, right? >> We want to sell you a pile of dirt, ready for your shovels to start doing something. It >> does sound better. It does sound better than I've got raw land or I've got dirt in Oklahoma, right?

47:59 >> Doesn't it mean you have permits and other things that are important? >> Sometimes, sometimes not. >> we we would never have expected to get where we to be where we are today. >> Yeah. from if you have a good builder it takes probably from raw land and permits to stand up 50 megawatts which is a reasonable block and often even the big sites unfold in 50 megawatt blocks >> 50 megawatt blocks is 18 months now if there's an existing building there and there's grid power already we call those brownfield sites so a lot of what's going on right now is people are going to the rust belt and they're buying old factories, paper mills, because they had >> a lot of power, >> a lot of power, right? That they had the permits for power already >> and retrofitting them.

48:51 >> So, how much of their kind of supply chain bottleneck is energy versus chips versus the construction versus permitting versus whatever else? >> Each chunk of the supply chain has its own supply chain. So, data centers need there's plenty of concrete. in most places there's labor. Although in Wyoming where there was a battle for all these data center sites, they were bringing in electricians from as far away as Denver. generators and electrical transmission switches are long lead time items right now and those are hard to come by. And so those are the long poles usually.

49:33 Usually you can get up what's called a cold shell. You can build a concrete tiltup building or a metal building fairly quickly. but then you have to fit it out and turn it into a a high-powered hotel for for compute. >> Yeah. >> Right. And that takes electricians, takes cooling, chillers, all these other things. Those all are long lead time right now. >> I was really I happened to fly over leaving Memphis last night. I happened to fly over macro hard and macro harder.

50:06 I it it was astonishing like the size. >> Nobody can build like Elon. It's unbelievable. >> The the size of buildings and you know the construction nearby and everything. I was like whoa that is tremendous. Which leads to an interesting question now which is like I could imagine a world where either it'll just never be enough data centers you know and for 10 more years it's always going to not be enough or because of the dynamics of the lead time and what's going on right now you could imagine like an overbuild and betting on which of those worlds you think is going to come out is there's probably some alpha there for whoever we >> don't have enough data centers >> we won't certainly I mean I seeing out 10 years in our space right now is is really hard Right. 10 years ago, transformers weren't running.

50:53 People were running tiny little models and trying to identify faces or cats and chairs or whatever. I mean, it it was 2016 was not a big AI year, right? but five years we will still be chasing data centers and we'll still be chasing chips. And for those who use HBM, they'll still be chasing HBM. >> Yeah. It'll be interesting also how politics plays into all this because like now you got probably, you know, more government interest and involvement in tech and AI specifically than we've ever had obviously.

51:24 >> No, what we need is more people who don't understand making decisions. That's clear clearly what we need. I mean I think more people with absolutely no clue making making important decisions. You know, I I I think this is an area where they're really thoughtful discussions to be had about what's good for the US. >> Yeah. and good for different municipalities, different rural areas. and those aren't being had by the politicians. >> Yeah. >> Right. They're knee-jerk. They're sort of >> Let me ask you a question on this moving away from politics for a second. Just let let's say we do agree on this like compute thing that compute is constrained in some form or another whether it's memory or chips or data center space, whatever. There's some constraint there and the demand is is overwhelming.

52:14 Doesn't that argue then the only thing that matters is like tokens per watt? Like basically we're going to have a limited number of denominator and so however much whe whether there's tokens per watt or tokens per facility or tokens per chip or I don't know something like that but it feels like we need the most of the numerator for the limited denominator task per watt maybe >> and and so like if that's the case then you have cerebrus has a speed advantage a huge speed advantage >> but like how do you think about that dynamic if we fast forward in terms of what the conr actually is.

52:52 >> I I think there are a couple things. First, speed, which is what we've chosen to focus on, has historically created markets. >> Mhm. >> Right. and if you think about it, right, there is no market for slow search, right? There's no market for dialup. your kids are are are what 14 and 12. If you want to punish them, right, don't don't take away their phone. Ratchet it back to dialup speed.

53:22 >> Right. Right. This is a real punishment. Let him use it for a week at dialup speed slowly. Right. That why is we laugh and then we say it's okay for for AI to be slow. I mean think about it. I mean when when the internet was slow Netflix delivered DVDs and envelopes. >> Yeah. >> And when the internet got fast they became a movie studio. Right. That's not they didn't get better at their other thing. they became something entirely new. And I I think what we're seeing with the launch of of GPT56 Soul, right? It's put out in limited limited availability last week that people are thinking of whole new applications.

54:01 You've got frontier intelligence instantly and that that opens up all sorts of new opportunities. So that's one thing you do with speed. The other thing you do is you try and think about how you can drive up throughput, how you can make more tokens. And one of the ways we're doing that is by partnering in something called disagregation and we're doing it with AMD we're doing it with AWS where you think about in the work of inference is there a part that can be done by somebody else and is there a part that can be done by you such that the the result is higher throughput and you know with with AMD we're seeing 5x additional throughput I mean five times times as much throughput while keeping the speed the same. And we're seeing similar numbers with with with AWS. We have an opportunity because of our architecture to do that with with the entire GPU landscape. We we we could do it across the board. And so you know there there four major chip makers right now in in our category.

55:08 Obviously Nvidia, we'd love to partner with them. There's AMD, there's the Google TPU, and there's AWS with their training parts, and we're already working with two. So, that's something we've thought a great deal about and and is a vector we are extremely interested in in chasing down. >> What has it been about Nvidia that has made them have the just crazy run that they've had, you know, like I've >> I think most people are are wrong about in in what makes Nvidia great. first and Nvidia's, you know, in the first quarter of this century, they're they're the great company without any question, right? and I I think people look to a bunch of things.

55:45 They look to CUDA. They I don't think it's CUDA. They they look to their chip architecture. I don't think it's a chip architecture. it's this unbelievable grit and intensity that was born of a decade of not having success as a public company. I think if you look at their stock chart between about 200 two what 2003 and 2013 2004 2014 for a decade right they traded horribly and you're a public company and you're fighting tooth and nail and no one's listening to you and you can't sell very much and the relentlessness and the grit that that takes is awesome and to come out of that as the most valuable company a decade later right the most valuable company in the world.

56:34 That sort of intensity and grit for a company of their size in my view makes them one of the great companies in history. It's it's not this these other things. These are just things people say. But that that sort of a decade of being a public company and fighting and fighting and fighting that gets in your DNA and that that's awesome. >> Yeah. I mean like you see Jensen like still at events >> fighting tooth and nail.

57:00 >> Yeah. >> Right. Right. I mean where could you have imagined >> trillion dollar companies? >> Could you imagine old tech leaders before Jensen doing that? >> No. >> Right. Right. >> Have that kind of fight. Yeah. >> I mean that in my view is what's awesome. >> Yeah. >> that level of fight I think the other stuff cool good but but when when I look at what what I can do better when I look at sort of what I ought to be thinking about as being a CEO those are the things I look to. Do you think did your near-death experiences give your company some of that DNA?

57:33 >> You know, we we we are and I I think this is what's interesting about Jensen, too, is he sees himself still as the underdog. Now he's the big dog, but you know, this is my my fifth startup. You know, we sold three and took one public previously and we've now taken this one public. I'm a I'm a professional David in the battle with Goliath and I I wake up every day with that mentality and that we're now bigger we we have sort of bigger competitors, right? We have bigger challenges. We have more people throwing stones at us. and so I I I I hope we take that. I personally wake up every day with that passion and that drive. I mean, every day I think to myself, when we started they said it would never work. you can't do way for scale and we made wave for scale and then they said all right you did way for scale but you can't yield it in volume and then we yield it in volume they said okay you can yield it in volume but you can't package it in volume production and then we packaged it in volume production they said okay now you've packaged it in volume production you only have a government customer and we said okay then we won a sovereign cloud and then they said you don't have a a hyper you don't have a a a frontier lab and then we won open AI and And they said, "Okay, you've got government, you you've got sovereign clouds, you've got a hyp you've got a a a frontier lab, you don't have a hypers scale. Then we want a hyperscaler. Right? Each time they said then they said you you've got all these cool customers, but you couldn't you couldn't do big models. Now we're serving GPT.

59:11 >> What are they saying? What's that now? Like what are they saying? >> CEO is a boomer. That's good. I mean I mean but each time >> I became a millennial and I shut down and you know right and so I I think when you're sort of a professional David when when you are an entrepreneur at heart right each one of those fires you up right we're only interested in solving problems that other people can't solve that we're only interested in doing things that that other people can't do that's why we get up every morning it's not the money it's not it's I mean, notoriety. It's because we love building cool things, and we really like building cool things that are so hard that other people can't build them.

59:56 >> Eric told me you had a pretty cool and unique childhood. I don't know if that fed into this at all, but like how'd you grow up? >> I grew up on the Stanford campus. my parents were faculty and there's a a little neighborhood where all your all your neighbors are are other professors. my neighbor was William Shockley, the inventor that's crazy transistor. >> So crazy. >> Right. So all we knew about him, you know, we're 10 or eight is that his wife gave out fulls size candy bars at Halloween, right? The inventor was at Bell Labs. He invent he brought Silicon Valley, right? He he by the movement of Shockley to the West Coast, created the foundation for Silicon Valley. And we're thinking he gives big three musketeer bars, right? but I I think there were a couple things that were glorious.

60:45 First, the only currency was intellectual horsepower, right? That that was nobody cared if you're rich. Nobody cared if you'd started a company. That this was the 70s. Nobody car what they cared about was, well, that dude's really smart. He does good work. My other neighbor was almost for his work with David Conorman, they got Nobel Prize in economics. my dad's tennis match there were six or eight guys in rotation. They played doubles on Saturday and Sunday.

61:19 somewhere in my mid20s I realized three had Nobel prizes and one had a Fields medal, >> right? and >> he sucks at doubles. But >> that's right. No, let me tell you, it was some old man tennis. I mean, their serves were grim. >> Their physics was good. >> It's a crazy way to grow up. >> It's a crazy way to grow up. And you know, we the neighborhood was safe. we'd get on our bikes and we'd just go all summer and come back at when it was dark.

61:45 >> and so did you think you'd be an academic? >> I did. I actually I was working on a PhD. I got a little bored. I went to the business school of Stanford while I completed my qualifying exams for my PhD. And I sort of got got sucked into this by mistake. And you know my dad still asks me he say you're going to finish your PhD Andrew like dad all my professors are dead. Nobody left. so >> yeah, amazing. All right. Well, Andrew, this was a blast. Thanks a ton for doing this with us and obviously you know, you've built something extremely special and it's been cool to just watch and learn vicariously through Eric. So, thanks for everything.

62:23 >> It's a pleasure to to be here and chat with you guys. And you know, Benchmark was an extraordinary partner. I mean, I I think if you do hardware, you're going to be in bed with your your your backers for a decade and pick good ones and and I'm proud we did. Thank you.

Summary

Andrew Feldman, CEO of Cerebras, discusses the evolution of AI and the critical role of hardware in its advancement. He reflects on the challenges faced in building Cerebras, including the need for radical innovation to compete with giants like Nvidia, and emphasizes the importance of speed and throughput in AI applications. The conversation also touches on the complexities of the semiconductor supply chain and the future of data centers.

- Cerebras was founded in 2016 to address the computational demands of AI, which were not well served by existing architectures.
- The company focuses on radical innovation, believing that incremental improvements are insufficient to compete against established players like Nvidia.
- Andrew highlights the importance of having control over all aspects of chip design and manufacturing to achieve significant performance advantages.
- The semiconductor supply chain is complex, with significant lead times and dependencies on various components and technologies.
- Data centers are currently constrained in capacity, impacting the ability to meet the growing demand for AI compute power.
- Speed and throughput are critical metrics for AI applications, with Cerebras aiming to deliver substantial performance improvements.
- The conversation emphasizes the need for strong relationships within the supply chain and the importance of trust in partnerships.
- Andrew's background growing up around influential figures in academia shaped his approach to innovation and problem-solving in technology.

Questions Answered

What were the initial challenges faced by Cerebrris?

Cerebrris faced significant challenges in its early days, particularly in board meetings where the recurring theme was the inability to meet expectations. Despite this, there was a strong belief in the potential demand for their product if they could overcome these hurdles.

What are the key components of Cerebrris' chip design strategy?

Cerebrris focuses on three fundamental elements of computer architecture: calculations, memory storage, and data movement. The company aims to innovate in all three areas to enhance performance and efficiency in AI applications.

How can a board contribute to a hardware company's success?

A board can provide valuable insights and support in areas like financing and strategic direction, especially in long-term projects. Their ability to see across the industry can help guide the company, even if they lack technical expertise.

How important are relationships and trust among founders?

Trust and relationships among founders are critical for success. Working together previously can foster respect and understanding, allowing for effective collaboration and minimizing conflicts.

What are the future challenges for data centers in AI?

The demand for data centers is expected to continue growing, with challenges in predicting future needs. The landscape of AI is rapidly evolving, making it difficult to foresee the exact requirements for infrastructure and resources.

© transcribe · For agents Built with care and craft by Gokul Rajaram