Meet the CEO of Cerebras - The Quiet Company Making AI Possible

Welcome everyone. A bit of context. AI is one of the most significant technological advancements in recent years. It's a movement that's all about augmenting our human potential. Currently, large language models are increasingly, increasing exponentially in their power of an order of magnitude that we cannot kind of get our heads around. One of the stats I read is that AI is due to be a $1.4 trillion industry by 2029. Impacting every sector. However, it has been constrained by limitations in computing power, which is where Cerebras comes in. Cerebras is revolutionizing the frameworks for AI. They build supercomputers that give AI practitioners the computational muscle to do their job faster and at scale. So we're incredibly privileged to be hosting Andrew Feldman, Cerebras' co-founder and CEO. Andrew is an entrepreneur dedicated to pushing boundaries in the compute space. Andrew, I believe this is your 5th company. A few examples of his prior experience prior to Cerebras, he co-founded and was the CEO of C-Micro, a pioneer of energy-efficient high-bandwidth microservers, which was acquired by AMD for about $357 million. Before that, he was with Force10 Networks, which sold to Dell. And then prior to Force10, he was with Riverstone Networks from the company's inception all the way to IPO. This is a man who is pioneering fearless engineering, and we're really honored to be hosting you, Andrew. Joe, I will hand the reins over to you to kick this whole thing off.

Irene, I want you to do introductions at Thanksgiving with my parents. I think that was lovely. Thank you.

You're welcome.

We'll rent her out to you in exchange for one of your clusters that we can experiment with, like two. Andrew Feldman. So to the audience, you know, you've, you've heard me say in the recent months, right, that in my view, the investment in Cerebras is probably one of the most important and timely investments we've ever made on the LinkTo platform since we launched. And I don't use those words lightly because As Irene said, generative AI is going to be a foundational technology that will impact all sectors of commerce and interaction in the coming decades. But like all prior important foundational technology movements, they're all based at the end of the day on hardware infrastructure. You can't run generative AI and create generative AI without some really critical infrastructure. Our view is that Cerebras is that critical infrastructure that is going to be necessary to make generative AI proliferate and be truly productive, right? If you look in the past at chip companies, Intel and AMD made possible personal computing, right? We wouldn't have networking and cloud computing without the likes of Broadcom. We wouldn't have smartphones without the likes of Qualcomm and Arm. And everything we watch in terms of video graphics, the streaming of it, the processing of it, that's NVIDIA and companies of that ilk, right? So Cerebras is this new generation of chip company, and Andrew will tell you precisely what makes them different, that they're on the cusp of this new wave of technology that generative AI is going to be.

Right?

So you've all probably, like me, experimented with ChatGPT, and I, I think it's fun and also a learning experience. But at the end of the day, um, ChatGPT and any one of these large language models that, that generate all the texts that you read, right, they can't actually work until they're trained. So Let's start first maybe by having Andrew explain in layman's terms, what is the training of a GenAI model? What are the problems that arise in that training? And how do you guys at Cerebras solve that training problem?

Sure, Joe, thank you for having me.

We're honored and we always enjoy talking about about AI in the market.

And our wives are tired of hearing about it.

And so we're eager to share our views with others.

Prior to about 2014, 2015, there was a collection of obscure academics who were working in the domain of neural networks and of artificial intelligence.

And they were laboring in this obscurity because their models never did anything. And that's very interesting. These are guys that we think of now as pioneers in the field, Geoffrey Hinton, Bengio, LeCun, et cetera. Starting in about mid-2014, early 2015, they had an idea, and that was to change the compute platform that they were working on from CPUs to GPUs. And in so doing, they brought enough compute power to the problem that suddenly these algorithms that had in the past produced absolutely nothing began producing very interesting results. And that's sort of the, the kickoff for this modern AI push of the last 8 or so years. And what we saw first was it in the domain of vision. And suddenly these algorithms were able to outperform humans in a whole set of tasks involved in vision. But we saw this, and in late 2015, my co-founders and I had decided that we wanted to work together again. We wanted to put our shoulders to an industry. We want to try and move an industry with, uh, the next 5 or 10 years of our lives. And the industry we love is, is the compute industry. And we saw this on the horizon, and we, we saw that this particular problem had basically infinite compute demand. That even in its earliest stages, what we could see was that as you applied more compute to the problem, the results got better and better. And there are very few problems like that. As you apply more compute to database problems, the results don't get better. As you apply compute to almost all of the problems you encounter in the computational world, they don't get better results. But for this work, as you applied more compute, the results got better and better and more accurate and more accurate. And so we, we saw this as an opportunity. We, we saw this as a, uh, a tectonic shift. And historically, tectonic shifts in, in applications have produced extraordinary companies. Underneath in the hardware realm, and they've been tremendous challenges to the existing status quo, right? For those of you who, who remember the, the rise of Ethernet and IP networking in the mid to late '90s, didn't help Nortel. Nortel had tremendous share. It didn't help Ericsson. It didn't help the installed players. All of them got demolished, and, uh, the results were new companies. Cisco, Broadcom, uh, and later Arista and Juniper. Um, in the same way, the companies that had approximately all the share in the PC space, uh, with the rise of a new type of compute, uh, a handheld compute that was limited by, by battery power, um, they won approximately zero share, right? There are, are no cell phones running x86 machines, and a whole set of new companies stepped in. We, we saw that sort of opportunity in 2015. We saw an opportunity to build a company on a new technology, on a technology that was not repurposed, right? The graphics processing unit was so named because for 20 years it was targeted at one particular problem, and that was pushing pixels to a monitor, right? It was a graphics engine, and at graphics it is a perfect machine. It has, like remodeling your house or other efforts to remodel, it is an imperfect machine for the AI workload. And this is what we saw in 2015. And it was catalyzed by a comment our CTO made when he said, wouldn't it be serendipitous if 20 years of tuning for one workload produce the machine well suited for another? And wouldn't it be odd if a machine designed for pushing pixels to a monitor was good at AI? And with that, we, uh, we set off on this journey that is Cerebras. Uh, we saw an opportunity, uh, to build a different type of chip, a chip that was in every way optimized for one thing, and that was doing AI work. It looks different, it behaves differently in every aspect. The chip is different than a graphics processing unit. It's different than a traditional CPU. Just by give you some idea, it's 56 times larger. Our current chip has 2.6 trillion transistors, whereas the largest chip today has about 80 billion. It has thousands of times more memory. It has tens of thousands of times more memory bandwidth. It, it was a profoundly different machine we chose to build. And, uh, you know, one of the few times in my life that I underestimated the size of the market I was targeting. It turned out that we were wrong, that AI wasn't going to be big, it was going to be overwhelming and transformative, not just for our portion of the tech market but for whole economies. That's what we've seen in the last 6 or 8 months. That was a bit of backswing. I hope that wasn't too long. Tell you a little bit about how we got to where we are.

Great, Andrew. That's super helpful for you to set the context. Let's talk about the whole training requirement for any type of a large language model. What is that training? If you could explain in plain English, What's involved from a compute standpoint in doing that training? And why is your particular chip and system design more advantageous to somebody doing that training than the other devices that are on the market?

Sure.

First, AI is divided up into two stages. There's a training stage and there's an inference stage, or sometimes the latter stage is called serving the model. What is different about AI than other forms of traditional computation is that— and we'll go by, by example— in the past, when you wanted to teach a machine to identify a cat, you wrote a description of whiskers, you wrote a description of paws, you wrote a description of fur, and you, in your algorithm, you described the elements of catness, and this was very specific and prescriptive. Um, when an AI identifies a cat, the algorithm doesn't— begins with none of that. The algorithm is instead designed to be shown data. And after thousands, tens of thousands, hundreds of thousands of images in which, uh, it was sort of— it got it right or it got it wrong— it improved step by step. It got better and better and better at identifying the cat, to the point where even if the cat's in a shadow or you can't see its whiskers or etc., the there is a, the model has learned from huge amounts of data. This is how we began. We moved into the natural language processing domain where we showed the model, we showed the algorithm huge bodies of text. And over time, it began to understand English. And we didn't provide it a dictionary. To do a translation. We didn't, uh, provide it a correlation between a picture of a dog and the word dog. What we did is we provided vast amounts of data, and it sifted through this, found patterns in language, um, and after this sort of inordinate computational task of sifting through this tremendous data, it came to understand the patterns in the English language or other languages. And to understand in some sense the meaning of words. And this is a brute force compute approach. You show in these models, you show a model approximately everything ever written in English. And to do this, we scrape the internet, we use all digitized books, we use, I mean, absurd amounts of data. And we, grind through this with some of the largest compute clusters ever built. And the result is a model that has accumulated insight by being shown data. And when you then stop the training and you ask it a question, it can retrieve this insight, and it can generate a response. And that's what ChatGPT was. And so this was, uh, it is at its core, uh, a profoundly computationally intensive task, um, because the model has to see vast amounts of data, and that data has to be computed in a way so the model can learn. Now, we saw this on the horizon. This part we, we managed to get right, um, and we, we saw that, that the problem was going to be too big for any given chip. In fact, the problem was going to be too big for tens of thousands of chips at a time, and the, uh, the cost in computing of tying together lots of little chips is high. Uh, and the performance of 100 little chips isn't 100 times one chip, right? It's vastly lower and costs more because you have to tie them together, and it uses more power. And so we had an idea that we could, by building a vastly bigger chip, we could reduce the communication overheads we could reduce the amount of, of times you had to move information, and in so doing, we could vastly accelerate the work. And that's what we did.

And just to give you an idea, I, I, I have my props.

The good thing about being in hardware is we get good props. This sad little chip is an NVIDIA A100, and you should lean forward because it's hard to see, it's so small.

Um, and, uh, it's about 800 square millimeters, or for those of you who don't know metric, it's about the size of a postage stamp. This is our chip, all right, and it's 46,000 square millimeters. That allowed us to have in one place vast communication resources, vast compute resources, and vast memory.

And that allowed us to do work faster and produce results in less time. And it allowed us to, to build clusters of them. Even one of these is insufficient for some of these large networks to build clusters very differently than with lots of little chips. Now, we weren't alone in, in the observation that, that the communication overhead in tying together lots of little chips was going to be a problem in this, in this domain.

And in 2019, uh, NVIDIA bought, uh, Mellanox for this exact reason.

They paid $7 billion to acquire a company whose sole reason for existence is to tie little chips together. And that's a different way of doing it, um, but our approach has produced results that are vastly faster, use less power, and produce answers in less time.

And, and that was our goal, uh, and we've been able to deliver this for enterprise customers, uh, for government, for intelligence communities, uh, both domestically and internationally. And, uh, that's sort of where we are.

That's an amazing, uh, from a design chip design standpoint as well as a manufacturing standpoint, quite a a feat of engineering, Andrew Feldman.

So it was, you know, we, we built a chip bigger than anybody had ever built in the history of compute. There had been other efforts to build chips this large, including by some of the biggest names in compute, including AMD. But we, we thought the time was right. We thought the workload was right that made it worth it. and, uh, we, we were able to solve a problem that had been unsolved in 80 years and, uh, began shipping product in, uh, late 2020. Uh, we have since, uh, shrunk our, our chip from 16 nanometer.

We're now at the 7 nanometer node. Uh, to give you an idea, you know, the, the chip I showed you has 850,000 cores.

Right.

How many transistors in there?

2.6 trillion.

That's an astronomical—

These are astronomical numbers.

Hard to even imagine. By the way, for those of you in the audience, when Andrew Feldman said that their latest generation chip has transistors of 7 nanometers, 7 nanometers is about 1 10,000th, I think, of the width of a human hair.

And that's the distance between transistors.

That gives you approximately how many transistors you can pack in a unit of space. And transistors are built up into your compute horsepower.

Yeah, yeah. And the NVIDIA chip, the GPU, has roughly how many cores, Andrew, in comparison to—

About 15,000. About 15,000. Today, NVIDIA has the lion's share of the market. Tremendous credit to them for making hay when the sun shined. They were a better part than the CPU. They weren't a great part for this work. They've tried to remodel and remodel and remodel and make it better and better. But the market has absolutely exploded. And you have individual customers. I don't know if you guys saw this, but the other day Google announced that it had, in addition to developing its own processor, it had procured a cluster of 25,000 GPUs.

That's a billion-dollar buy with a B.

And these are being stood up all over the world. And it is— the market is so big, so dynamic.

It's an extraordinary place to be right now.

Yeah, yeah. The ability to bring such a huge scale, right, of compute as well as associated memory on a single chip is important because these large language models are already large in terms of the datasets that have to be processed, but they're only getting larger, right? So I think GPT-3 had, call it 175 billion parameters, right? But the latest sort of Microsoft collaboration with OpenAI is gonna approach 1 trillion, I'm told. Parameters, right? So you can just, for the audience's perspective, think about what you would need to do to process a dataset that huge if you have to go through the division of that dataset to all those little GPUs that Andrew Feldman was talking about, compared to being able to squeeze big chunks of that into the Cerebras processor, right? So you're, you're ending up with a training process that that is significantly less complicated and therefore less costly than it would otherwise be if you had to deal with all the little GPUs that you now have to— you basically have to allocate or sequence work to.

That's exactly right, Joe. And so an easy way to think about it is a parameter is equivalent to an independent variable. And so these models are looking at hundreds of billions of independent variables, independent explanatory units for understanding language, or now heading towards trillions. And then you show this enormous set of independent variables, you show it a trillion tokens. And a token is, you can think about it as a word or a letter or two. And so you're building matrices for these calculations that are among the largest matrices ever computed, right? You might imagine two matrices, one that's 40,000 rows by 10,000 columns, right? And you're multiplying two together.

These are so big they don't fit on traditional machines.

Once you have to divide up this work, you have this extraordinary challenge of allocating chunks of work to different little machines, to organizing how they're back, how they're reassembled. Just to give you an idea, at the back of the paper describing GPT-4, they talk about the credit to people who were involved just in the breaking up of the work. and there are 100 people cited. There were 100 people involved in allocating the network, the neural network, over little compute units. It was an absurd amount of work, and, uh, uh, it cost billions of dollars to train, with a B. Uh, you know, we, we attacked those problems. Uh, it we put into the open source community 7 GPT models. We did it in a month, maybe cost a million or two to train, and we did it with one person. And so now their models were bigger and newer, but the order of magnitude of different allocation of resources to achieve results is useful there.

That's really, that's really interesting. And I've also heard from some engineering friends that, you know, other than the chip itself, you've devised some very, very innovative software, I guess, for lack of a better term, that, you know, that the way that these chips run that deal with some of the issues that you just talked about in, in, in processing these very large datasets, right? So you talked about matrices. I just recalling my college matrix algebra course, the— in order for the process to work properly, the, the, the matrices have to typically be homogeneous. In, in, in size. So think about a rectangle, and it's got a uniform width and, and, and length, and all the rectangles have to kind of look that way in order for typical processes to run. But you guys have created a way of having your chips process heterogeneous matrices that are not necessarily the same size and yet do so efficiently.

So what do you do?

I think that's exactly right.

What do you do when the data doesn't come in a nice rectangular matrix?

What you do is you make it rectangular by stuffing the entries filled with zeros.

By padding.

Right, by padding. And then you present it to your compute engine. And if your compute engine is a standard GPU, it just multiplies those zeros. And in the world of compute, there are few things less intelligent than multiplying by zero. You use power and it takes time and you generate no new information. You knew before the calculation that if you multiply by 0, the result is 0. After the calculation, the result is 0, and all you've done is take time and power. We don't need to do that. In fact, we can decompose the matrix into vectors and multiply them by scalar.

We avoid this problem.

We avoid the problem of ever multiplying by 0. We do a collection of things that we're able to do by virtue of the fact that we specialize in this one type of work.

We don't have to carry the burden or the tax of also doing graphics, of also doing 64-bit double precision work for supercomputer workloads.

We are a specialist machine and we're designed for AI. And for workloads that, that, that have the same fundamental underpinnings of AI. And much of the advantage comes from the fact that when you design for and are focused on one particular type of work, you can do away with a lot of other stuff.

Yeah. Great. Let's talk about some of the commercial aspects of your business now, Andrew Feldman. Perhaps you could highlight who some of your current customers are, how they're using your gear, and then as importantly, how would you expect that customer base to change and expand and evolve as GenAI usage proliferates and becomes more mainstream?

Sure, we go to market, thank you for asking about us. We go to market in several different ways. You can purchase our equipment and put it on your premise. You can subscribe to our equipment and it can be on your premise, or you can use our equipment in our cloud. And there you can gain access to it by the sip, so to say, right? By the day or by the model or by the week or by the month. We have big clusters available to do some of the largest language models on Earth. We have small clusters available for smaller problems. And so customers, if they have equipment on, if they have a data center like our customer GSK, the giant pharmaceutical, they have equipment on-premise and we've been able to do all sorts of interesting work. They've in fact published work where they use natural language processing models on epigenomic data. And that's a very, very interesting application of language models because DNA is a language. And so both genomic and epigenomic data work really well on when analyzed with large language models. We have customers in the energy space. We have many customers in pharma.

We have customers in the energy space like the French supermajor Total.

There they published work in which we were— a single one of our systems was more than 228 times faster on their work than the top-of-the-line GPU. We have government customers, both throughout the Department of Energy and the intelligence communities. At the Department of Energy, we, we sort of swept the landscape. Our customers include Lawrence Livermore, Argonne National Labs, National Energy Technology Center. We have customers in, in Japan like Tokyo Electron Devices, in Europe. So there's a collection of, of different customers of various sizes and various types. On our cloud, we have both large and small customers like Jasper. Many of you might use Jasper's marketing tools for you to generate marketing content. So big and small. Jasper is 2 years old, GSK is 300 years old. We have companies that just raised their Series A and companies that have hundreds of billions in revenue like Total.

Have you had— just interestingly, let me ask, have you had any financial institutions approach you in order to use this for, you know, training very domain-specific types of models that they want to develop? So for example, a hedge fund that wants to, you know, create an AI-based predictive tool for trading decisions.

We have trained models for financial institutions. And we've done that on domain-specific financial data. We've done it on both proprietary data and on open-source data. There's a Thomson Reuters dataset that they made public a few years ago. We have done fine-tuning, which is where you begin with a general model and you, you then show it very, very specific data with the idea that you're then going to have it do work in a very specific domain. And I think maybe the way to think about it for the audience is that general training of these GPT models, and the G stands for general, right? It's a general pre-trained model. You might think of as maybe you're taking your 4-year-old and having them do cardio and weights. You do that for 2 years or 3 years. The idea is whether you then present your child a task in wrestling or javelin or karate, they are pretty well situated.

They're fit, they're strong, they can do lots of different things. The idea with pre-training is, is a foundation on, on which to build.

The idea with fine-tuning is if you then say, why don't we do 50 karate classes on the foundation of fitness, you add this very specific expertise, and then you say, how are you going to do on karate tasks, right? That, that's the inference. Here's a karate task. And what we've shown over time across the industry is that works really well. Uh, you train on a general dataset and then you show very, very specific tasks for your domain-specific activities. And so that's exactly what you described, Joe, and that's, uh, we do that all day, every day for, for customers.

Great. I'm gonna ask a final question and then turn you over to our audience for Q&A, Andrew. Sure. So in our business at Linqto, one of the private market databases we use consistently is provided by an outfit called PitchBook. And they recently built an interesting analytical tool that, you know, takes in inputs regarding a private company and its markets. and, and they basically, uh, use that data to make a prediction as to how successful that company will be, uh, in doing an IPO, right? Um, the— that, that predictor assigned a rating of 99% to Cerebras. Congratulations.

Well, clearly it's right. Um, and I— while I was going to, to to attack that predictor. I now realize the error of my ways and how obviously correct that model is. I think with—

So since it's obviously correct, here's my question for you. It's a little bit of a mind experiment. So fast forward to, you know, whenever, 2024 or some date in the near future, you are sitting in New York in a large conference room with all the buy-side analysts from the top institutional investors in the country, your Fidelities, your Vanguards, your BlackRocks, Franklin Templetons, what would be the top 3 reasons you would give them for buying into Cerebras' IPO?

You know, I think in my experience, large and growing markets are where you want to be. And I think in my analysis of opportunities, and I evaluate fewer opportunities than you do and the people on this call, because I mostly invest years of my life behind a single bet. And The structure of the market and the size of the opportunity is foundational in my thinking. And I, I think, uh, it's hard— you're hard-pressed to find people who, when they look carefully and thoughtfully at AI, don't come away saying, uh, this is an economic dislocation of, of tremendous, uh, reach, you Bill Gates talked about it as bigger than the internet. And so part one is what is happening in AI right now is transformative, not to a little corner of the compute world, but how we live, work, and play. And I tell people that my 9-year-old granddaughter uses AI unbeknownst to her when Netflix recommends a cartoon for her. My 92-year-old mother-in-law, she uses AI when she asks Alexa to play a Frank Sinatra playlist, and Alexa puts together a playlist that includes songs she forgot she liked. I think AI is creeping into the way we do things, right? It's in your phone right now in selecting your route home. It is sort of step by step in the background creeping in. And that's before the big impact, right? The big impacts of a technological shift are not when they replace things we already do, right? The big impact of compute was not when they replaced typewriters or they replaced, general ledgers, right? The general ledger accountants in the old world were extremely efficient. Um, the, the big, uh, economic change came when technology allows you to reorganize resources fundamentally. When we went to the cloud, there was a huge change that allowed software to be allocated differently. It opened the world for SaaS. It created vast sort of different economic opportunities. And we still even haven't seen that with AI. Right now, AI is doing copywriting better. It's producing a better type of search. It is, uh, gener— it is helping us generate better code. Um, it is, uh, removing the laborious tasks of, uh, of visiting a building before you insure it by driving by it instead of taking photos from satellites or from drones. It is doing hundreds of things, but many of those, the vast majority of things were already being done, perhaps inefficiently by people, but were being done. I think the most exciting part is when suddenly we're doing things that weren't done. And so I would begin with that discussion with, with, uh, with everybody where I share the view that, that this isn't a little thing, uh, that this is the rise of something profound. I, I would say that, uh, in those environments, uh, again and again, uh, have emerged new hardware platforms. And that is the history of the emergence of new hardware. It is very hard in the status quo for new hardware to emerge, right? Uh, nobody has made a meaningful threat on AMD or Intel in the x86 world for decades at a time. And then there was a threat that emerged with the rise of discrete graphics. right? And boom, they got no share. AMD acquired a company to get share, but Intel had approximately zero share. There emerged a rise of a new workload in, uh, the moving of data, uh, in IP networking. Intel could have done that. They had zero share. There emerged a new fundamental way of doing compute where it's in your hand, not in a laptop, where it's powered by a very small battery where weight mattered a great deal, new companies came in. And so I, I would point first to the size of the market and second to the history of large dislocations producing great new companies and having really deleterious effects on the status quo. I mean, Intel has spent what, $15 or $16 billion acquiring company and has approximately zero share in this market. You know, I love Lisa Su and what she's done at AMD is extraordinary. She is, I think, one of the top 3 CEOs of the last decade. They have zero share. So I would first say the size of the market and second, the structure of the opportunity. And finally, I'd say that what we've done is not build a part that's incrementally better. I don't think that's the, the way you build a great company. I— we've built a part and a technical trajectory that is profoundly better, that involved deep work. We— it involved invention on the dimensions of chip design, of system architecture, of power, of cooling, all domains that are fundamental to, to building big computers. And that the proof is in the pudding, and that our customers, not only are they saying nice things— words are cheap— but they're publishing under their name in top journals results that are mind-boggling.

Great.

Thank you, Andrew Feldman. I gotta tell you, it's been such a privilege just having you share a lot of those insights and wisdom. And I dare say that, you know, 10 years from now, we'll look back and listen to this interview and realize just how profoundly prophetic, uh, you were today. So with that, I will turn this over to our audience for Q&A.

Thanks so much, Joe. Thanks so much, Andrew. This has been fascinating. We're getting a couple of questions regarding the chip itself. And obviously we've already established that you've created the largest chip ever. And as you're in hardware, it's great for show and tell. So maybe you should show it again to our audience. But Hong Chan is asking, Could you please elaborate why is the big chip approach better than chiplet approach like Tesla Dojo?

Uh, so first, I, I think that the question begins with an assumption that you need a lot of chips. And also buried in the question is the idea that communication is really important, because for those who don't know, Uh, Tesla opted for a similar approach. They take a wafer, they dice it into individual chips of about 800 square millimeters, but then they place those on interposer, on another piece of silicon, and they place, I think, 64 of them, uh, forgive me if I get that wrong, right next to each other. On a single platter, right? And when they're done, uh, they come up with, uh, a, a part that has 64 little chips that are next to each other, very, very close. Now, uh, that approach has pros and cons. Uh, uh, we, we know Elon well. He visited us early before he started that project.

Um, the problem of placement of those chips is very, very hard.

You mechanically have to pick them up and put them down exactly where you need them, and that's historically been a very hard problem. You still have to leave the chip boundary to communicate. So when they communicate, they communicate with Assertis, and Assertis is a, a technology that goes from the electrical domain into the optical domain and pushes traffic out the edge of the chip through the interposer to the edge of the next chip. Now, the SerDes takes a great deal of space and power, so their chips run vastly hotter and use a great deal more power than ours. They also get less performance. They get less performance because they had to allocate real estate on each chip for those I/Os, for the SerDes blocks. And SerDes are very large. And so they made a different set of trade-offs.

They built a single-purpose machine.

They built a machine for, uh, computer vision as it relates specifically to self-driving cars. Um, they recognize the importance of big chip, uh, of having all sorts of compute on a single platter.

We did it on a single piece of silicon.

They do it on two pieces of silicon. The first is cut up and the second serves as sort of a platter onto which the little ones are stuck. They're, you know, they're still buying a ton of GPUs. So I assume that years down since they started the project, it is yet to deliver sort of what they what they wanted. Other than that, you know, there are a lot of different ways to skin a cat. Ours, from all data they've shown, is higher performance and less power. But I'm sure there are all sorts of thoughtful architectural reasons that they chose their path.

Okay, amazing. We have a kind of similar question regarding the chip. From James Long. James says, I believed Innovium was building the kind of chips you're talking about, um, hence I invested in Innovium. How is— are your chips different from the ones Innovium, now Marvel, were— are building?

Well, so in Innovium, uh, that's— well, I'm glad you invested in Innovium. They worked out okay for you. Uh, Innovium was building a switch chip, and that's a different thing. They weren't building a computer. They're not a processor. They were a— if I remember correctly, uh, they competed with Broadcom in building a communication chip. And a— I used to build switch chips. That was the first 15 years of my career. Um, a switch chip has one job. It, it takes an address from a packet, it does a lookup, and it sends it off, uh, in the direction prescribed by a table. And, uh, they built, uh, big fast switches, but their, their chips were smaller than this, right? So that they were traditional chips, uh, to do switching. Uh, what we've done is, uh, we've taken the elements of compute called the core, and we put 850,000 of them on a giant wafer, and we've linked them, each core to its own memory. And each one of those cores is a real computer. You could run an OS on it. You could run separate software on each of those 850,000 cores. Now, that's not the, the way we recommend doing it, but you certainly could do it that way. Um, and that's different. That this is a, a, uh a processor, whereas they were switch chips.

Okay, great. I'm going to ask Stephen, who's holding his arm up to ask a question, to come over and ask it for you live. So let me just pull him over.

But they sound so smart, Irene, with that English accent. I mean, everything sounds better with an English accent. No, it does. I mean, even silly stuff sounds so sort of thoughtful.

Great. I'm glad that's what comes across.

Am I on?

Stephen, please ask your question. Go for it.

Brilliant. So I'm actually already an investor in Sirubus.

Well, thank you.

Pretty excited about your prospects. Would love to learn a little bit more about differences with Zambonova, you know, and would like to learn a little bit more about customer acquisition.

Sure, uh, SambaNova, uh, is, uh, a company that has chosen a very different path, uh, a path not focused on, on performance, um, and has chosen to sort of deprecate their development of hardware underneath, uh, the delivery of sort of a combined sort of software offering. As I said, there are lots of paths in this market. I think our view is we've never seen performance data that makes that strategy the right approach. And this is a performance-intensive market. We haven't seen, and in our discussion with customers who've done testing, performance is vastly different. But, you know, they— Chris Rhee, who's one of their founders, is an extremely smart software guy, and they have a talented team. And, you know, we don't see them often, But it's a big space. And I, I think, uh, there is, uh, in, in vectors we care about in the performance of, of large language models. I, I think there's, uh, plenty of customer data to show that, that we're head and shoulders faster. Um, I, I think they chose a, a less innovative strategy. They built chips of approximately the same size as everybody else's. Their, uh, approaches to gang several processors together. This is a pretty straightforward approach that's been done before. On the other hand, they're smart guys, and, and I'm sure they're, they're getting, uh, decent results from, from that approach. Customer acquisition, uh, certainly wasn't hurt by the, the last 4 months of, of, uh, sort of frenzy over ChatGPT. Um, you know, maybe a year ago, uh, you'd be in working with a small group in a large company, and the group might be led by a wizard. Uh, I mean, not with the, the sleeves wizard, but a wizard who has some, uh, sort of is recognized as, as a technology guru. and suddenly, every board member in every large company is asking, you know, the, the Global 500, you know, what's your AI strategy? How's AI going to impact our business? Projects are being kicked off at fast-moving and slow-moving companies of, of every size. And so this is true domestically, this is true internationally. There is an extraordinary push this minute. For these large enterprise customers to have AI strategies, to think about more generally how to use data assets, which were expensive to obtain, how to use them to create differentiation, to create durable advantage, and to find value in them. And I think for those of us who work in this domain, data is the new oil. Right? And these companies have extraordinarily rich data assets, and frequently they don't exactly know how to turn them into value. And that, that's something that's very much top of mind, Stephen, across the world as we see engage with large companies. Um, and so, uh, at the same time, the venture community has poured vast money into startups doing AI work. And so while we're working with sort of Global 500s, there's sort of a plethora of little startups who have raised $10 or $20 million who need $5 or $7 million worth of training done. And they are interesting and fun to work with, and they're pushing the AI to new levels. Whereas the larger companies frequently want others to invent the AI and want to rely on their data to provide insight. Did that answer the question?

Yeah, that is an excellent answer. And then just a real quick follow-up from that. So I know you mentioned Qualcomm and its sort of position as mobile technology took off. So Qualcomm had, you know, these strategic partnerships with Samsung and sort of other handset manufacturers. You know, what are the key strategic partnerships that will drive Cerebras growth?

Yeah, I think the structure of the market is obviously very different. They are an embedded component provider and we're a system solution provider. And I think you're absolutely right, Stephen, in thinking carefully about what are the partnerships that a company like ours needs to develop in order to support rapid growth and global distribution. I think those are very much the right questions to ask. I don't think it's necessarily The same as other— the embedded handset market. I, I think there are, for all intents and purposes, only 10 ways to get your cell phone chip to market in the world, with about, what, a third, a third of the market being one company, Apple, right? And if you then throw in Samsung, now you're at half the market. So there are really a very small number of ways for them to get to market. Um, I think one of the beautiful things about the cloud is we can provide services, uh, with our U.S. facility. We have partners in Europe who, uh, fall within the European data regulations where we can provide compute today for anybody in the world. You can come, you can jump on the cloud, you can get the benefit of this, but by the, by the day, by the week, by the month, by the model. And so I think that changes the dynamic rather significantly and makes sort of the traditional partners whose job it was to distribute your product, it forces them to find a different way to add value. right, to engage with, maybe they are specialists in tuning models for the finance community. Maybe they are specialists in the analysis of medical records, which is a very interesting domain for AI. What you would like to do is to be able to take medical records, this is, plethora of medical records. And this is almost always done outside of the US because of HIPAA and all sorts of requirements in the US. But in countries that have nationalized healthcare, you have big datasets of medical records. And what you'd like to be able to do is to predict, uh, the, the next illness based on medical records, right? That's the generative aspect of that problem. And there are groups that are engaged in that research. It's a very hard problem. Doctors are notoriously bad at filling out medical records, but there's been more standardization. There are tremendous opportunities there. There are places that have genetic data and have medical records. And what you'd like to do is tie together the genomic data, the phenotypical data, the health outcome data, and make predictions again on future health challenges and what you might be able to do preventatively. And so I think, Stephen, that the structure of partnerships in this domain may much more look like software partnerships where rather than distribution partnerships, they are people who have the ability to use AI to provide unique insight and then have your customers and their customers want a combined solution of their software and our hardware on-premise solving a particular problem day to day with their data.

Andrew, thank you so much. We are at time. I wanted to make sure that we don't keep you too long, but I think we have— there are so many more questions that are still coming in. And people want to answer it. So if all of you are on the audience want Andrew back for another show of part 2 of this AI environment, just let us know in the chat, say yes, we'd like Andrew back, and we'd like him back tomorrow or next day, and we'll schedule it with Andrew when he can come back. Um, what, um, if you want to close us out with perhaps one or two sentences, I think Andrew Looking into the future as a futurist, what perhaps do you see the major change going to be using AI? But I mean, specifically, you are the rock that allows everyone else to generate AI stuff. But what, as a futurist, what would you say to look out for?

Well, first I'd say that today we're the pebble aspiring to be the rock. I think we are sort of a pebble aspiring to be a boulder. Um, I, I think we don't have to look out very far to see a tremendous impact across white-collar professions. And I think many of our innovations over the past 25 years in, in tech have made, uh, have gone after blue-collar labor. Robotics in particular, gains in efficiency have sort of reduced the role of technicians or increased the role and reduced the number. I think generative AI will change what it is to be a lawyer. It will change what it is to be a doctor. It will change what it is to be a wealth manager. It will change a whole set of professions that have remained in many ways immune to technical changes. We practice law approximately the same way we practiced it 200 or 300 years ago. I'm interested personally in how it impacts teaching. We teach approximately the same way we taught 2,000 years ago. And there are very few social institutions that have remained unchanged. Our families aren't the same as they were 2,000 years ago. If you look at the bedrock of our institution, our religions aren't the same, our values, but the way we teach children actually is not very different from the way Aristotle taught Alexander the Great. And that is a bummer. And I think it is— we teach without data. We teach in a bizarrely— I mean, it's before medieval way. And I think AI will attack some of these institutions and transform them, these institutions that have remained sort of immune to change. And I think that is a really interesting thing.

So much. So thank you, Andrew. That's so much to explore, and we've got a resounding yes. A lot of people want to see you back, so we'll coordinate with you for that next part too. But Joe, Irene, thank you all for, for putting this together for us. Thank you for all— thank you to all the attendees. You've got great questions in there. We had a couple hundred people who registered for this event today. Some of you were able to make it, some of you were not. We'll send you a note with the recording so you can review some of this conversation. Respond back to us on that email and let us know some of the questions that you would like us to prepare Andrew Feldman for, for part 2 of this show. So be engaged and let us know what it is that you want. We did have a couple of questions from investors. They're talking about, they'd like to know a little bit more about your market position and some of the questions that Joe perhaps touched on a little bit earlier, but more in-depth questions. Andrew, we'll send this back to you and your team so that you've got a heads up for our next time around. Joe, do you wanna say a couple of words, Irene, and then we can close this?

Just thank you again, Andrew. It's a real privilege having you here, and hopefully you'll come back sooner rather than later. Perhaps your interviewer then will be a bot who will be immensely more intelligent than me and will be armed with an English accent to boot.

As long as it has Irene's accent, I think we'll be fine.

Thank you so much, Andrew Feldman. There's a bunch of questions that didn't get asked, so I'll send them over.

All right.

We appreciate it. Thank you guys for making time to listen to us and learn a little bit about the AI market. Be well, everybody.

Thank you so much.

Bye. Thank you.

Bye-bye.