The Twenty Minute VCCerebras CEO, Andrew Feldman on Why Raise $1BN and Delay the IPO & Why NVIDIA’s Worried About Growth
EVERY SPOKEN WORD
125 min read · 24,945 words- 0:00 – 1:19
Intro
- AFAndrew Feldman
Things are moving at a rate that six, eight, 12 months out, everybody's unsure. It's so fast, it's so big. There is unbelievable demand and nobody knows where it will go in the future. The question of depreciation is how much faster are future generations than the current generation? That's the actual question on depreciation. People often say we don't have enough power in the US, and this is strictly wrong. We have plenty of power. It's in the wrong places. Risk comes in financial markets where people fundamentally underestimate risk. No company ever went bankrupt by paying extraordinary people too much.
- HSHarry Stebbings
Ready to go? Andrew, dude, it is so lovely to have you back on. I so enjoyed our first show. You put up with my naive questions enough to agree to do a round two. Man, I must be charming. (laughs)
- AFAndrew Feldman
H- h- Harry, I'm, uh, uh, I'm okay with any questions, naive or otherwise. So I'm, uh, ha-happy to do it anytime. I, I read your, your, uh, your LinkedIn posts, your Twitter posts. Uh, I, uh, I'm rooting for your mom. I mean, it... All good. All good.
- HSHarry Stebbings
Dude, you are too kind.
- 1:19 – 4:17
Why We Did Not IPO and Raised $1BN From Fidelity
- HSHarry Stebbings
Uh, listen, I wanna start with the billion dollar raise that you just announced yesterday. Um, can you just talk to me about the billion dollar raise, why it's important, why now, and what it means for the company?
- AFAndrew Feldman
Look, it was the largest raise ever done in, in our category. Uh, it was done at the highest valuation and with the, the premier investors. So, uh, at, at late stage investing, you're looking for, uh, the likes of Fidelity. They are the, uh... What would the English call it? The sort of Oxford or Cambridge of investing, right? (laughs) I mean, they, they are the, uh, the premier, uh, public market investors. And when they choose to lead a round, uh, it, it brings the, uh, Wall Street a great deal of confidence. And so, uh, we were really happy to partner with them and with the treaties to lead the round, and then we, uh, were able to get enormous participation from Tiger Global, from Valor, from 1789. So that's point one. I think point two is that, um, w- we, we now have sort of the dry powder to, uh, to really push and to take the opportunities in front of us, uh, to build out our manufacturing to, to the scale and scope we want, to add new data centers. We added five this year in the US to add more data centers. And we have more big ideas, right? I think incremental improvements, uh, uh, m- make-believe gains achieved by dropping from, from, you know, 8-bit to 4-bit, uh, th- those aren't gonna get us to, uh, to the promised land in AI. We, we need... We've got real work to do as a community. And I, I, I think this, this funding puts us in the catbird seat for that.
- HSHarry Stebbings
On the Fidelity side, it's actually interesting. I had Brian Halligan, the CEO of HubSpot on the show recently, and he, uh, actually taught me the importance of specifically getting Fidelity in, both pre-IPO and when you IPO, just because of the signal that it sends. And I didn't actually realize as a venture guy-
- AFAndrew Feldman
(laughs)
- HSHarry Stebbings
... the weight that's placed on Fidelity.
- AFAndrew Feldman
As an early stage venture guy, you're like, "Who are the public guys?" (laughs)
- HSHarry Stebbings
Uh, okay. Fidelity, like, wha- whatever, T Rowe, sure, they're all the same, right? And I was like, "No, no, they're not." Like, Fidelity are the monster in the room.
- AFAndrew Feldman
They're the monster.
- HSHarry Stebbings
And the importance of getting them is, is very high. Can I ask you, dude, like, why not go public? It was... 'Cause it was rumored that you guys were gonna go public. Why do this pre-public round?
- AFAndrew Feldman
We, we still have every intention of going public. I think it's very common, uh, i- in a late stage to do a pre-IPO round, i- if you can get it done very quickly, if it doesn't distract you, and, and keep moving. I, I think there were so many opportunities in front of us that, uh, g- gathering the capital so that we could continue to prosecute these opportunities was sort of a no-brainer.
- 4:17 – 13:22
Analysis of Chip and Compute Landscape Today
- AFAndrew Feldman
- HSHarry Stebbings
You said about the real work to be done. I think it's quite difficult for everyone who's not really in the market to understand what the hell's going on given all the news that we see. Can you help us just-
- AFAndrew Feldman
(laughs)
- HSHarry Stebbings
... with the lay of the land in the last three months, of where are we at now? What's changed? Let's start there.
- AFAndrew Feldman
The first thing, Harry, is we, we, we are in a, a, a stage of the market wh- where the, the claims are enormous, (laughs) r- right? Where, uh, tens of billions of dollars are being done here and there, and nobody's reading the fine print that it's over five years and it's up to. This is the, the, the great sort of CYA word in, in, in marketing history, is it will be up to 100 billion over five years. Right? Well, up to means it could be 30, it could be 12, could be 40, right? (laughs) It won't be bigger, right? You, you could pick a lot of big numbers and it won't be bigger than. And so I, I think as you read these deals, I, I think, you know, you, you have to really think about the timeframe over which they're being done, you have to think about whether, uh, anybody is actually counting, right? Lots of people are saying they're gonna, you know, gonna bring hundreds of billions of dollars of jobs to the US and this and that. I mean, in eight months, has anybody got a little, little spreadsheet, like nine jobs plus one factory? I mean, who, who, who holds anybody to account? And the a- the answer is nobody. A- and so I, I think that's number one. I think number two, wh- what this signals more than anything is that there is unbelievable demand and nobody knows where it will go in the future. That it's so big and happening so quickly that, that they don't know.I mean, we, we have customers coming to us and saying, uh, "We, we would like between 5 and 40 million queries per second." Well, how do you not know by a factor of 35 million queries per second where your demand's gonna be? How do you, how are you unsure by an order of magnitude (laughs) of, of what your queries are gonna be? And the, the answer is, things are moving at a rate that, that six, eight, 12 months out, everybody's unsure. It's so fast, it's so big, and so I, I think you should think about sort of these announcements as sort of options on the future. Right? That, that's really the, the, the way to think about it, is, "In an unknown environment, how can I take an option on the future? I don't know if I'll use it all, but I'll pay something for, for the future, uh, rights to, to, to, to have some capacity." So that, that's a w- a way to think about it.
- HSHarry Stebbings
Given that it's so fast, it's so big, how do you think about planning for that uncertain-
- AFAndrew Feldman
Brutal.
- HSHarry Stebbings
... future?
- AFAndrew Feldman
It's brutal. I mean, I, I think there's a very interesting question about, uh, in extraordinarily, uh, rapidly moving environments, what, what the right planning cadence is, right? (laughs) A- a- and what you really need is good planning changing rules rather than good planning, (laughs) right? If you could only s- I mean, we have to make big bets. We're making five- and seven-year investments in, in, uh, data center capacity. We, we are, uh, making, uh, hundreds of millions, now on term of billions of dollars of bets in the supply chain. Um, y- you, a- and those are not three-month bets, right? And so, um, I think what you need to do is use different rules that have historically been used. You, uh, plan more frequently. You have a shorter, uh, view. You take options on the future. Uh, uh, and, and if the future moves against you, you, you, you lose the premium on the option, right? You, you, you pay a little price to, to secure some capacity, and if you don't use it, you just, you go, "All right. That, that was sort of a way to manage uncertainty about the future."
- HSHarry Stebbings
What do you think the chances are that you are still underestimating even your wildest of demand expectations?
- AFAndrew Feldman
100%. (laughs) I've been wrong... I mean, heh, look, if, if you would have said, uh, a year ago, (laughs) two years ago, pick, pick a time at which it would've been conceivable that, uh, OpenAI would get the valuations they're getting. Pick, pick a time, right? Wouldn't have been conceivable, right? Three months ago, six months ago, nine months ago. When was, when was it a reasonable, like, the day before? (laughs) Right? (laughs) A, a, a, a- and, uh, I think that's true with the demand we're seeing. It's true with the valuations on companies we're seeing. It's true with, um, the, the rate of ideas entering the community.
- HSHarry Stebbings
How much of this do you think is sustainable? Everyone lobbies the, "Oh, it's not sustainable. A lot's experimental. It's not enduring." How much do you think is sustainable?
- AFAndrew Feldman
Y- you know, I, I would say this. I, I would say that there are always grumpy people who, who are like, "It'll never work," and, you know, you know, "E- e- you'll never beat Goliath." A- and, you, you, you know, truth is, most things don't work, and most of the time, Goliath wins, (laughs) right? But there's no alpha in that. There's no money f- made for you or me be- betting on the biggest of the big dogs to continue not to lose. I mean, how, how uninteresting is that? A- and so, I, I think, uh, what, uh, uh, of course, right? If NVIDIA keeps growing at the rate they're currently growing, 11 years from now, everybody on Earth works for them, (laughs) right? It's, do the math, right? I mean... However, uh, uh, is it possible that our economy looks very different in five years? Is it possible that, uh, the things we value, uh, are very different? Um, that we have reorganized around AI, we've seen a major bump in, in labor productivity, we have, uh, uh, benefited dramatically, and the economic pie is much larger? I, I think that that's not only, uh, likely, it's almost certain.
- HSHarry Stebbings
You mentioned that if NVIDIA continues to grow the way that they do, everyone will work for them.
- AFAndrew Feldman
Well, uh, the, you, you, you just, you know, you, you keep doubling at that rate (laughs) and you multiply the, the, the... You, you can't keep doing that.
- HSHarry Stebbings
To what extent is it just completely unshakeable at this point for them? Like, where the scale and the size of money, it's like, you know, Jonathan Ross from Grok said on the show they will unwaveringly get to tr- $10 trillion within a five-year timeline.
- AFAndrew Feldman
I hope he's long on them, then. Um, I, I, I don't pick public market stocks. I, I, I think, uh, I don't like public, picking public market stocks. I think in the public market, you can lose money on good companies, you can make money on shitty companies, and that, for me, doesn't sit well, right? (laughs) Uh, a- a- a- as an entrepreneur, as a David in the battle with Goliath, I, I wanna make money when, when we build a great company, period. But, um, can they continue to grow? I think we are seeing some things that big companies do as they begin to worry about growth, I think. Use your balance sheet more and your technology less. Right? Th- this is something that historically large companies have done as they feared for their, their, their technical prowess.
- HSHarry Stebbings
And when you say that, you're kind of referring to investments in your OpenAIs-
- AFAndrew Feldman
Yes.
- HSHarry Stebbings
... for $100 billion-
- AFAndrew Feldman
Yes.
- HSHarry Stebbings
... and ElevenLabs and everyone-
- AFAndrew Feldman
Yes.
- HSHarry Stebbings
... in between.
- AFAndrew Feldman
You start buying business-... as opposed to winning business. And I, I think that we saw that with Cisco, uh, who, who emerged as a dominant position, you know, '99, 2000, 2001. Um, that, that has been a, a, one of the strategies. Another strategy you, you see is this predatory pre-announce, where y- y- you announce B300s before anybody can get B200s. You start talking about Rubin before B200s are, are, are technically, uh, finished. Um, you don't talk about, uh, the, the field failure rate of your products, which are massive, uh, uh, rather y- y- you, you sort of keep talking about the future in an effort to, to convince people to wait, t- to make a good decision, uh, uh, rather than go with technology that's, that, that's, uh, better and present. And so I, I think the- these are the strategies of, of very large companies using their strengths. Um, a- and I, I think that's what you're beginning to see unfold with, uh, with NVIDIA.
- 13:22 – 15:57
NVIDIA Showing Signs They Are Running Out of Ideas
- AFAndrew Feldman
- HSHarry Stebbings
Specifically on, uh, uh, I, I do wanna talk about actually kind of the speed of chip development-
- AFAndrew Feldman
Yeah.
- HSHarry Stebbings
... uh, specifically. But speaking about 100 billion into OpenAI, how did you analyze that? For, for me reading that, I didn't really know how to analyze it. It's so unprecedented.
- AFAndrew Feldman
Well, I, I, I think it was designed for nobody to understand it. If we, one w- w- wants to make something very clear in an investment, "We've invested this amount at this valuation. The deal's done now." If you wanna make something more difficult, "It's up to this amount over an unbound specified amount of time, but no valuation given or a valuation specified, but it can change." (laughs) Right? It, it wasn't designed for you or, or other analysts to, to anchor (laughs) o- on different things. And, and that's a, a very reasonable thing for, for, for both of them, but it's just, it's not an analyzable, uh, thing. What, what, what... Beyond the fact that, that NVIDIA has chosen to try and lock up a portion of, uh, OpenAI's demand, uh, by investing in them. Th- that, that's about as much as you can say.
- HSHarry Stebbings
Totally get you, and I'm glad that it's meant to be confusing, 'cause I was confused looking at it, going, "What price was this? How much are they buying?" Like...
- AFAndrew Feldman
I, I don't know if it was meant to be confusing. I think it was meant to be-
- HSHarry Stebbings
You know, my mother goes shopping, and I ask her, uh, "It's a lovely dress, Jules. How much is it?" "Well, it doesn't matter. It doesn't matter."
- AFAndrew Feldman
(laughs) You call your mom-
- HSHarry Stebbings
I go, "Whoa!"
- AFAndrew Feldman
... by her first name? Hold on, wait a sec. Let's go back to the important thing. You call your mom by her first name?
- HSHarry Stebbings
Oh, yeah. Jules.
- AFAndrew Feldman
Okay. You don't call her Mom?
- HSHarry Stebbings
No.
- AFAndrew Feldman
I've never called my mom Shirley. I mean, I, I, well, I could-
- HSHarry Stebbings
(laughs)
- AFAndrew Feldman
... (laughs) never call her... I mean that would be just Mom, or something else. But not-
- HSHarry Stebbings
No, no, no. But when I, when I get the, like, price on application, I'm like, "Oh, Andrew."
- AFAndrew Feldman
(laughs)
- HSHarry Stebbings
Oh, dear. Oh.
- AFAndrew Feldman
(laughs)
- HSHarry Stebbings
That, that's what this felt like. That's what I was like, "Really? Uh, th- no? Nothing there?"
- AFAndrew Feldman
I, I, I certainly don't think it was... The truth is, is there may be, and there are likely, a huge number of moving parts that make it impossible to clearly, uh, describe without giving out more than they wanted to give out. But-
- HSHarry Stebbings
You mentioned pre-announcements there, like B300s, B200s, timings of such. Are we thinking about chip depreciation in the right way? I just, you know... Again, I just did a show with Jonathan. He's like, "Hey, we actually think about them on, like, an 18-month time cycle to maybe two years." And I was like, "Wow, that's, it's quite quick."
- 15:57 – 30:49
The Real Questions to Ask on Chip Depreciation
- HSHarry Stebbings
Are we thinking about it the right way? And how should we be thinking about the amortization of chips?
- AFAndrew Feldman
We are in unprecedented water- waters. Um, I think, uh, people are clearly still getting value from H100s.
- HSHarry Stebbings
Mm-hmm.
- AFAndrew Feldman
And that's more than two years, right? So, uh, you know, uh, if you say it's a two-year depreciation, you're empirically wrong, right? I mean, they, they are, uh... And I think people are still getting value from A100s, though not on the cutting edge. And so that's closer to three or four years, uh, a- and could be as long as five or six. The question of depreciation is, how much faster are future generations than the current generation? That's the actual question on depreciation. Because what you're... With, with depreciation, you're, you're, you're saying, at some point, it's no longer worth using a part that's fully paid off, because there is a new part that's so much faster, uses so much less part, power, that it's better for me to retire it. That's the actual sort of, uh, the underpinning to the depreciation question. If I have a data center and it's 50 megawatts and I have this much capacity in it, at some point, even though my, my chips in it are, uh, have been depreciated and I'm running them at, at zero cost, right? Power plus zero depreciated cost, it makes sense to move them out, because the new one, the new chips are so much faster, so much better, use so much less power. I get so much more dollars per. And, and so that's the question. And, uh, w- if we don't, as an industry, uh, continue to build extraordinarily sort of better parts generation after generation, th- then people don't move from one generation to the next. They last longer, you depreciate them longer.
- HSHarry Stebbings
Where are we? And I feel very naive for our audience. Where are we in the performance improvement pathway for chips? Are we, like, in there, we've got 90% and we're at incremental gains? Or are we at the, we are still at the super early stage and we have 90% of the gains to be made?
- AFAndrew Feldman
I think, uh, the question is, in that case, whether you read people's marketing material or the a- actual performance results. I mean, certainly, the, uh, people's marketing material would lead you to believe, uh, that generation over generation, there are huge gains. Um, a little bit of engineering digging...... uh, p- probably leads you to the conclusion that you're getting 2, 2.5x, uh, per meaningful generation move, um, not more, right? If you compare apples to apples, 8-bit to 8-bit, 4-bit to 4-bit, uh, if you compare actual performance, uh, you know, you might have more FLOPS on the chip, but your memory bandwidth didn't improve more than 2x so you can't get to them. These chips are, uh, are- are a solution, and if you make one part fast, it's a system. You make one pa- fast and the other part you- you don't move, doesn't move as far forward, it becomes the new bottleneck. And you're k- it doesn't matter how- how many FLOPS your chip has, if you can't get data onto and off of the chip, th- those are wasted. And so, the, uh, the- the question isn't sort of, oh, how much, how much faster is the chip? It's how much faster is the solution? And that includes memory, which has, for inference, is the fundamental limiter for the GP architecture. And so, uh, it doesn't matter how much faster the chip goes, it ma- it matters how much faster the memory bandwidth is.
- HSHarry Stebbings
On this, I- I was chatting to a founder in the space and he said that what everyone fails to understand is that, like, although SRAM sounds great in terms of having memory, and SRAM is obviously, you'll describe it much better than me and hate me for this, but SRAM is obviously memory, like, on-chip versus off-chip.
- AFAndrew Feldman
Yep.
- HSHarry Stebbings
Um, seemingly great, but, uh, he said it's completely unable to handle scale. And so, although it may be quicker, for anyone who wants to do large scale, it is incapable, at present, of doing that, and that's a fundamental need and requirement ou- of any of the large providers.
- AFAndrew Feldman
Right.
- HSHarry Stebbings
Do you think that's fair? And how do you think about that?
- AFAndrew Feldman
Well, I- I... not only is it fair, it's the reason we went to Wafer-Scale, and so let me explain. What your friend said is strictly true, in that SRAM is blazing fast and low capacity. HBM is a flavor of DRAM. It has high capacity and it's very slow. Now, NVIDIA and all GPUs, including, uh, uh, AMDs, chose a big capacity memory that is slow because it's perfect for graphics. You don't have to go to memory very often. You can hold a lot, you don't go very often. SRAM is blazing fast, but it can't hold very much. So, the problem on traditional chips is, if you put memory on the chip... you are using space that could be otherwise used for compute. You have a fixed amount of real estate. And so, if you put half memory, then you have y- half your real estate's available for compute. And so, our idea was that if we, we sort of, if we built a chip that was the size of a dinner plate, we could stuff it to the gills with fast SRAM, overcoming the limitation of SRAM, which is it doesn't store very much, by putting a huge amount down, by using more silicon area. Now, if you're an SRAM solution today, in a normal-sized chip, and you're trying to do a trillion-parameter model, you use 4 or 5,000 chips. What a mess. You know how many cables that is? Do you know the- the- the impact to the AI? It's a horrible mess, right? And e- it limits you from doing things you wanna do with the AI, uh, like speculative decode, it has all sorts of painful challenges. On the other hand, use one of these, or two, or four, right? And it's simple, it's easy, and this is what your friend said exactly right, and the reason we went to build a bigger chip. So we could fill it with this fast SRAM, so we could get over the traditional limitations of SRAM, that it couldn't store very much, by using a lot of space, by using a huge amount of- of- of silicon area.
- HSHarry Stebbings
Mm-hmm.
- AFAndrew Feldman
So your friend is exactly right, you might listen to him again in the future.
- HSHarry Stebbings
Question to you. No offense, that seems a little bit obvious.
- AFAndrew Feldman
Yeah.
- HSHarry Stebbings
Like, okay, increase the real estate, shove more SRAM on.
- AFAndrew Feldman
Yeah. It does, doesn't it? Yeah.
- HSHarry Stebbings
Is it, is it as obvious as it seems?
- AFAndrew Feldman
No.
- HSHarry Stebbings
Am- am I missing something here?
- AFAndrew Feldman
Well, what- what- what- what we were missing is that for 75 years nobody could do it. That building a bigger chip had proven impossible before we did it. Nobody in the history of the compute industry had been able to build a chip bigger than about 840 square millimeters, in the 75-year history of the compute industry. And many people had tried and failed. And after we did it, uh, Elon tried at Dojo and they failed. Um, and so, uh, it's really, really hard. And, uh, our strategy, uh, had never been done before, never been successfully, uh, sort of yielded, and so, uh, while it was obvious, uh, it- it was hard.
- HSHarry Stebbings
So, when we think about, like, where the market is today in terms of training inference, do we agree then that actually NVIDIA's chips are much better for training than yours are, but yours are much better for inference than theirs are, and the market splits in that respect?
- AFAndrew Feldman
No, we're faster on both, but the, uh, the software challenges in training are real.
- HSHarry Stebbings
What does that mean?
- AFAndrew Feldman
It- it means that when a new model is built, all right, and everybody reads about it in- in- in a publication-It, it was done on a GPU. Everybody to, to train it, they, they take the recipe that was originally done for the GPU, and they have to move it to the recipe for their hardware, whether that's a TPU, whether that's an AMD GPU, whether that's a- a- a- another, uh, dedicated chip like ours. Y- you have to move it. And that's a harder software lift. In inference, the truth is, nobody cares about CUDA, nobody even cares about PyTorch, all right? Um, what they want is an API. (laughs) A- and so, the, it, it's literally 10 keystrokes to move from, uh, uh, a GPU-based solution on, uh, OpenAI OSS 120B to our solution. It's 10 keystrokes. That's it. It, it, it's, it's nothing. And so, I, I think the answer is, is that, uh, while we are faster at training and we are faster at inference, it's easier to demonstrate inference, right? You just put up a side by side. To show that you're faster than 1,000 B200s, you know, you gotta get 1,000 B200s. (laughs) Y- you need to train them for, for six week- train the model for four weeks or six weeks. You gotta stand up a cluster of our machines. You gotta... It's a bigger lift. So, I, I think, uh, uh, the, uh, the market right now is easier to, to move people off GPUs in inference and m- the, uh, number of people doing inference is vastly higher than the number of people doing training.
- HSHarry Stebbings
Can I ask you, when you look at the inference market today, how has it developed in a way that you did not expect?
- AFAndrew Feldman
I think it's really hard for the mind to wrap itself around geometric growth or exponential growth, right? I think the, the, the, there is nothing confusing about the greater growth of inference. The greater growth of inference is the number of people who use it, times the frequency of use, times the amount of compute needed per use, right? It, it is three different variables multiplied by each other. The problem is they're all growing fast, and that produces some mind-numbing effects. More people are using AI. Once they start using AI, they use it more frequently. And what they wanna do with it is bigger and more complicated, so it uses more compute. And so, you, you have three variables, the size of the market's a product of the three, all growing fast. And, uh, we knew that going in, we see that, and it still takes your breath away.
- HSHarry Stebbings
I didn't think we've seen anything yet. (laughs)
- AFAndrew Feldman
I agree with that. No, no, no. I agree. I, I think, uh, the reason I am 100% sure that, that we are underestimating the market is because of that premise.
- 30:49 – 36:42
Energy Requirements for AI: Is it Feasible?
- AFAndrew Feldman
- HSHarry Stebbings
If we're gonna see that transition, uh, that you mentioned there, the data, sorry, the, the energy requirements are just insane. I mean, Sam said a trillion-dollar spend. He needs the energy of Japan or more-
- AFAndrew Feldman
Yeah.
- HSHarry Stebbings
... to be blunt.
- AFAndrew Feldman
Yeah.
- HSHarry Stebbings
Is this feasible, Andrew?
- AFAndrew Feldman
Yeah. It's feasible. Um, is it desirable or good for society is a different question. It, it's feasible. Pe- people often say, "We, we don't have enough power in the US." And this is strictly wrong. We have plenty of power. It's in the wrong places, (laughs) right? It's, it's not where we have people or where we have fiber optic cable, right? We have a ton of power in West Texas in, in natural gas. We have a ton of power in, uh, Upstate New York in hydro. We have a ton of power in lots of places. We don't have people there. (laughs) And so, uh, the problem is, is one of a mismatch between where all the power is and where the people are, or where the buildings are, or where the telco fiber is that we need to get data to and from the data center. So, that's the first observation. The second observation is sort of one of a community, and that's that to the extent we consume this extraordinary amount of power, we have an obligation to deliver amazing things, right? And that's on all of us, right? I, I, I, I think we have an obligation to deliver drugs that they're more efficacious, to deliver better healthcare, to, to make, uh, aging less painful, make the looking after of aged parents or sick parents. Uh, I mean, you go through society's ills and woes, um, if we are gonna consume this amount of power, the burden is on us to deliver value for it. And if we use it and don't do that, th- th- then it's n- not a gain for society. But if we-
- HSHarry Stebbings
Do you think that's controllable? Like, you know, creating Gibli or Gibli images, I get it wrong, like isn't particularly value inducing, but it churns a huge amount of, you know, compute and energy. Can we control that? That's like-
- AFAndrew Feldman
It's a very hard question, Harry.
- HSHarry Stebbings
... hey, you get what you
- 36:42 – 39:27
Mag7 Value Concentration: Feature or a Bug
- AFAndrew Feldman
Um-
- HSHarry Stebbings
D- do you worry about the concentration of value in Mag 7? They now make up more of the S&P than they pretty much ever have done in history. And that concentration of value is very real. If AI hits a speed bump in any way, the market could derail significantly, and the multiplier effect of that is felt by everyone.
- AFAndrew Feldman
The risk there is that, not that they consume that much, th- th- they are that much value. Th- th- that's not the risk, I think. They're that much value because the future economy, th- we believe the future economy, uh, will reward that. I think the issue is, is that people then think the S&P is a safe investment, or a safer investment than it might be. The risk is the mismatch, right, in the, uh, in the mental model people have. Right? Ri- risk comes, uh, in financial markets where people fundamentally underestimate risk. When risk is priced properly, all right? The- then what happens is, uh, your outcomes are not surprising. But, you know, i- if people continue to think the S&P is sort of a- an index of- of the global economy, or, a- and it's not. It's 30% or 50% seven companies, um, then they're exposed to sector risk that they weren't signing up for. They thought they were diversified, and in fact, they're heavily dependent on a very narrow sector. And that's a risk. Do- do- do- do you know what I mean?
- HSHarry Stebbings
I totally get you.
- AFAndrew Feldman
Yeah. Th- that- that seems to me to be a- a- a- a challenge i- in the new w- world order. And all the advice that sort of, th- the- the pundits give, you know, diversified portfolio, when the world changes and that portfolio is not y- your, you keep holding it and it's not diversified anymore 'cause of consolidation, uh, then there's real risk.
- HSHarry Stebbings
Do you think the risk is priced properly when you look at NVIDIA at four and a half trillion?
- AFAndrew Feldman
Look, I think they've proven themselves to be the- the greatest company the first quarter of the 21st c- right? 21st century, right? I mean, th- they- they've proven themselves to be, uh, uh, the, uh, extraordinary company in the first quarter of the century and I- I- I don't know if four trillion's right, but I think, um, uh, a very big number, maybe it's too low, is right, because of what they've achieved.
- 39:27 – 43:18
Talent is the Bottleneck and Trump Makes it Worse
- AFAndrew Feldman
- HSHarry Stebbings
When we look at, we- we said before about the insatiable demand, th- demand that we cannot predict or anticipate, what are the bottlenecks today in your mind? You know, a- again, we had Jonathan Grock on and he was like, "Actually, you know, I had someone come and demand five times the supply that I have in total, and I only, that was from one customer." Um, supply is mine. How do you think about the bottlenecks that we have to reach the insatiable demand that you mentioned?
- AFAndrew Feldman
I think if you go back to planning, if you've got customers demanding five X your capacity, I mean, you probably didn't get your planning right, right? (laughs) You probably should have, should have planned better. I think there are bottlenecks, uh, at every level. Th- that are meaning, re- re- m- real and meaningful. I- I think, uh, I think th- the first one is expertise. I- I think, uh, we have, uh, fundamental limitations in AI expertise. We're not making enough AI practitioners. We're not making enough data scientists who understand, uh, uh, data pipelines. We are not, we're n- we're just, our universities aren't minting enough. And our- our challenges in the US with immigration, uh, don't help that. Um, we, uh, have historically sucked the best and the brightest first on J-1s to come to our schools and then to H-1s to stay. If that is not our policy, we need to make them. Right? I- if the government decides that that is not the way they wanna build a workforce, instead they wanna build it, uh, out of people who live here, w- we need to do a better job of training those people. We need to do a better job of teaching them in K through 12, we need to do a better job of educating them in our universities in order to- to- to make the number of, uh, of, uh, of engineers we need t- to meet this demand. That's a bottleneck. And it's why the- the best and the brightest are- are getting such extraordinary compensation.
- HSHarry Stebbings
Is the war for talent completely out of control? You're seeing your Zucks of the world spend hundreds of millions on one person. Do you think that's a blown-up anomaly or do you see the war for talent being unprecedented?
- AFAndrew Feldman
There are engineers who have skills that no number of other engineers working together can achieve. There are scientists who do, who have ideas and who have brains that are, ca- can't be replicated by lots of other talented people working together. Ought they to be paid more than w- world-class soccer players? I have no idea, (laughs) right? Um, maybe, maybe not. Uh, but-
- HSHarry Stebbings
I mean, inherently, yes-
- AFAndrew Feldman
(laughs) Maybe.
- HSHarry Stebbings
... from an economic rationale standpoint and yes-
- AFAndrew Feldman
May- maybe they do. That's right. Maybe.
- HSHarry Stebbings
... the value generated from a chief scientist at OpenAI, if they add $50 billion of enterprise value to pay them a billion dollars is worth it.
- AFAndrew Feldman
It may-
- HSHarry Stebbings
Like, it's worth paying-
- AFAndrew Feldman
It may well be, and that- that, that's what we have to think about. I mean, we've, I don't know, we- we paid Charlie Sheen $2.5 million an episode for Two and a Half Men. (laughs) $2.5 million an episode, right? I'm- I'm pretty sure that- that there are lots of people who are, uh, uh, whose net productivity to society (laughs) is above that. Um, we're pulling-
- HSHarry Stebbings
And he still spent it all.
- AFAndrew Feldman
(laughs) He- yeah, I saw the, I just saw the show on, was it Netflix or Prime? What- what a, what a sad story, uh, of somebody who was so self-destructive and so talented. But, you know, should we be paying soccer players or basketball players? I have no idea.... and I, I don't spend a minute worrying about whether we're paying extraordinary people too much. I think no company ever went bankrupt by paying extraordinary people too much. You go, if you wanna go bankrupt, pay mediocre people too much. (laughs) Right? Uh, that, that's how you, how you mess up. But n- nobody's ever struggled by paying truly extraordinary
- 43:18 – 55:50
Evaluating the Data Centre Economy: Many Will Lose Money
- AFAndrew Feldman
people too much.
- HSHarry Stebbings
What's the other bottleneck? You said expertise is one. What's another?
- AFAndrew Feldman
TSMC can't build fabs fast enough. I think the, the truth is, is that these are both for, for TSMC and, uh, and Samsung. These fabs are th- the most amazing manufacturing plants on the planet. And, you know, these are $30 billion, $50 billion factories. Um, and the, th- their ability to build them quickly enough is, is very much limited. Um, I, I think that in turn limits and keeps the supply, uh, below where it would like to be of chips, not just our chips or NVIDIA's chip, but everybody's chips, below where it might otherwise be, keeps the cost up. Um, I think, uh, uh, th- right now, there's a shortage of, of data center capacity. Um, and I, I think, uh, there's a huge amount of investment that has gone into that. Um, there's a lot of words, but w- where are these giga- gigawatt facilities that everyone's been talking about, everybody's committing to? Where are they? Well, they're not up yet.
- HSHarry Stebbings
How long does that take?
- AFAndrew Feldman
You know, somebody like for Elon, who's the fastest in the world and, and maybe the best at building plants and l- large construction projects, it takes six months, eight months. And for the rest of the world, it takes a year and a half, um, maybe longer. And-
- HSHarry Stebbings
Are we investing enough in data center buildout? I mean, it is one of the most insanely, um, hot categories now, in terms of investment properties. I, I'm coming from, like, a pure Wall Street mindset.
- AFAndrew Feldman
Oh, yeah.
- HSHarry Stebbings
Like, every Wall Street guy wants to be in data centers.
- AFAndrew Feldman
Yeah. It, it, (coughs) it has a structure that they really understand, right? It looks like a bond to them. Um, it looks like a piece of real estate. You, you get a tenant. They pay rent every month. Um, you, you can loan against that. You, you get an investment-grade tenant that's basically a bond. Um, it, it has the advantage of, of falling into a category or a pattern that is really well-understood in the debt market and in the capital markets. Um, and that's an advantage. And I, I think CoreWeave and some of their sort of financial engineering and innovations there help the world see that. Um, uh, I, I, I think, uh, you know, li- like many things, lots of people will enter. The smart will make money. Uh, the less sophisticated will, will lose money. And I, I think building data centers is not for everyone.
- HSHarry Stebbings
How will you lose money building data centers?
- AFAndrew Feldman
Look, I, I think if the best can build them for eight million a megawatt, uh, and you're spending 12 or 14, that's how you lose money. You lose money because it, the, it begins as, can you get access to low-cost power? It then continues to, once you have access, can you get permitting? Does that take long? Or can you, do you have real access that gets you fast permitting? Once it becomes a construction project, can you keep control of your costs? All right. (laughs) Once it's finished, all right, can you keep good tenants in it? Um, the, the ways to lose money in property are large and many. I- i- it had, you know, th- there's no, uh, there's no free lunch there either. And so, uh, I, I think, uh, when you are trying to go unbelievably quickly, you, you, i- it's harder and harder to, to be disciplined and not make mistakes.
- HSHarry Stebbings
To what extent is it important to be fully horizontal? You know, we hear about Zuck wanting the data center buildout to be just immense in terms of size and scale. To what extent does it need to be horizontal versus vertical?
- AFAndrew Feldman
It is completely unclear, right? The, the most successful two companies to date, uh, OpenAI and Anthropic, n- neither are vertical. OpenAI used, uh, Azure, 100% infrastructure for years. And Anthropic has used a combination of AWS and, and Google. And so neither are vertically integrated to date. Now, whether that's the right strategy going forward, whether they'd make those decisions again, w- uh, who knows? But it's clear that it's not the only strategy. There are plenty of working models where you are not fully integrated from, uh, chip through system, through data center, through software, all the way to the top.
- HSHarry Stebbings
Totally understand that in terms of the different models there. Again, sorry to cite it, but it's, it's kind of handy having just done it. Jonathan said that you would definitely have OpenAI and Anthropic build out their own chips because then they would have control of their own destiny.
- AFAndrew Feldman
Uh-huh.
- HSHarry Stebbings
Do you think OpenAI and Anthropic build their own chips, so they don't have self-reliance on NVIDIA in the way that they do today?
- AFAndrew Feldman
I think that there is a, a long history of software companies failing to build chips. The list is, is very large. I think, uh, whether, uh, OpenAI can do it, uh, whether they can do it through partnership with other vendors, with Broadcom, with smaller, more innovative companies is an open question. Um, but I, I think that, uh, you know, companies at the size of Microsoft have been unable to deliver.... uh, chips, right? I think, uh, uh, that there are plenty of examples as you look across the FAANG, uh, group where chips were tried. I mean, probably the most successful is Google, and they're 10 years in, right? Maybe longer. You know, soft- modern software does not fit well in a chip-making framework. I mean, weekly sprints don't work well on two-year-long projects. Um, you know, "Move fast, break things often" is not the way you think in the chip world. The way you think in the chip world is measure twice before you cut once, because your bugs cost you six months and tens of millions of dollars. And so, it, it's a very different mentality. And where there's been success, it has frequently been acquired. Apple got into the chip business through buying P.A. Semi. Amazon got into the chip business through acquiring Annapurna. Um, uh, Google acquired the talent from a collection of companies, uh, and then set it, uh, uh, in a BU that was aside and under somebody who, who had sort of enormous respect in the organization -in ORS, had a 10 or 15-year view. Um, th- these are things that, that have been challenging in, in many companies. Chip building is an MBA nightmare, right? Your, your analysis that says, uh, "Look, Intel had in, you know, between 2000 and 2010, some of the world's leading architects, the world's leading fabs, and proved completely unable to build a working cell phone part." And you ask yourself, "Why? They had everything they needed." And, you know, you do an MBA chart and it's like, you cannot... it's impenetrable. And the answer is, this is really hard. And the very small sort of mental model differences produce tremendously different results. How did every leader miss the largest compute market, right, in the first part of the, the 21st century? They... How did AMD miss it? How did G- how did ev- I mean, how did ARM win it? (laughs) Right? A- and all the leaders missed it. And then you say, "All right, may- maybe there's something in the guts here that I don't understand," (laughs) right? You gotta really get in there, and it's not on a PowerPoint. It's not in a two-by-two. It's not at some sort of consultant level. It is deep in the DNA of the small number of people who can build these things. And, uh, you know, we're lucky at Cerebras, we've got one of, one of the top six or eight teams in the world, and, uh, other startups don't.
- HSHarry Stebbings
What does, what does that market look like, do you think, in 10 years time? I know 10 years is a huge amount of time, given where we're at. But in 10 years time, is it a monopoly market with one taking 90%? Is it like-
- AFAndrew Feldman
W- which market?
- HSHarry Stebbings
... cloud, web?
- AFAndrew Feldman
Which market?
- HSHarry Stebbings
Ch- the... Specifically, the chip market.
- AFAndrew Feldman
W- which part of the chip market? Are we talking about AI silicon, or we're talking about silicon in general, or we're talking about-
- HSHarry Stebbings
I, I would say silicon in general.
- AFAndrew Feldman
Yeah, absolutely not one takes 90%. You know, even at the... at Intel's strength, uh, they had dominance in x86 and zero market share, you know, on the cell phone, and, uh, uh, i- and almost no share in the switching market. And, uh, Broadcom had dominance in the switching silicon market, which is a, a form of processor in silicon, uh, and no share in x86 or, uh, other forms of compute. The, the... Yeah, it, it will not all accrue to, uh, one or two companies.
- HSHarry Stebbings
How do you think about the importance of margins today, as a business as Cerebras?
- AFAndrew Feldman
I think the reason we were able to raise at a higher valuation and from better investors and, and more money is 'cause we had them. (laughs) Right? And others who were out looking for mar- looking for, uh, m- money had negative margins. And I, I think, uh, as you prepare for being a, uh, sort of a, a credible public company, um, I think y- you know, people do look at your margins. A- and I, I think, uh, that, that's a really important part of, of moving from being an idea to being, uh, a, a real company.
- HSHarry Stebbings
What are NVIDIA's margins today?
- AFAndrew Feldman
Extraordinary. Some of the, the highest in history for a hardware company.
- HSHarry Stebbings
How do you think about that? Is that just pricing power which-
- 55:50 – 1:01:36
Three Changes the US Could Make to Beat China in AI
- AFAndrew Feldman
- HSHarry Stebbings
Final one, just in terms of geography, you know, DeepSeek obviously had their moment, and it kind of solidified the concerns around China. How do you feel about China today as a pressing concern towards the US in terms of the race towards AGI between the two? Do you hate the way that it's posited as, like, China versus the US, the AI race? How do you feel about that?
- AFAndrew Feldman
I think it benefits neither, the position we're in, right? I, I think, uh, the Arms Race certainly didn't, uh, help either the US or, or Russia, uh, in the late '80s and '90s. We, we both spent money we wish... on weapons we wish would've been spent on infrastructure, people, or, or other things. I think, uh, we will be much stronger if we can find ways, uh, to peacefully, uh, engage. Uh, before these issues, we, we knew the guys at, at DIDI and ByteDance and, and, and Ali and Baidu extremely well. They're talented engineers try- trying to build cool stuff. I think our governments are at loggerheads, and, uh, that's a problem. And I, I think, uh, we, of course, uh, made choices, um... We had a huge opportunity in, in 2019 to do a deal in China, and, and I decided to pass because I, I didn't think it was, uh, the right thing to do. Um, long before, uh, uh, the, the Department of Commerce limited exports to China, I, I didn't think it was right and I was concerned about how the technology would be used. Um, but I, I think, uh, the realpolitik right now is that they're better at making drones, they're better at making robots, uh, they are... Their government has an extraordinarily aggressive policy in AI. For years, they, they backstopped their venture groups, right? So if you lost money in an AI company, the government would make you whole. Um, imagine that, Harry. Imagine how much money you could make if the government of the UK, right, offset some of your losses from AI companies that didn't work out. Um, and I, I think, uh, you know, we have real work to do in the US and, um, you know, they spent decade-
- HSHarry Stebbings
What, what, what work do you have to do that you haven't done? Like, what would you like to see done?
- AFAndrew Feldman
Oh, I think, uh, you know, China thought long and hard about, uh, their power infrastructure and their, their form of government allowed them to, to plan strategically. Our decentralized form of government has left us with sort of a patchwork of power infrastructures, um, you know, where even if the federal government wants to support you, there are local regulations, like at the, at the city and county level of towns, right, that can interfere with a project and, and set, uh, a project back billions of dollars. I mean, Samsung built a fab in Texas and they had to change the design of the fab because of a local fire ordinance, and the US government worked for years to get deployment of billions of dollars in Texas and a local fire ordinance set them back eight months, 10 months, and caused them to redesign the fab. I mean, right? That, that's a problem and, uh, uh, that's a challenge that, that we have to, uh, sort of collectively work through. I, I think we have the premier universities, we have historically drawn talent from around the world, um, you know, if, if you look at, uh, CE- the great CEOs in our industry, Jensen, Hock Tan, Lisa. I mean, you go down the list at, at, at, uh, Sundar at, at Microsoft, I mean, they came or their parents came. I mean, we gotta take that really seriously.
- HSHarry Stebbings
You don't buy the whole, "Well, actually, a, a load of people actually just abused H-1s and we'll just move to O-1s which people were using anyway, and the average salary for an H-1 was $120,000 and so it's a good thing and people will just use O-1s?"
- AFAndrew Feldman
We have H-1s and we have O-1s. I, I am sure that in every government program there's an ama- there's abuse, right? I mean, I'm, I'm not saying there's no abuse. Um, was there more abuse in the H-1 than in, in other areas? Um, i- I, I, I don't think so. Having the best and the brightest come to your universities, and once they benefit from sort of our great institutions to want to stay and contribute first with a J-1, which is a student visa, and then enter the H-1B lottery through the, through the, uh, approved process, right, to get a green card and become citizens, and this is how my parents did it, right? I think it's one way to bring an extraordinary amount of talented people to the US.
- HSHarry Stebbings
Is there anything else you'd change? You said the power infrastructure and the permitting around it.
- AFAndrew Feldman
Power infrastructure is, uh... It ends up at, at the local level, which is not necessarily where big ideas, uh-... uh, and sort of strategy is well knitted together. I think we, we have starved our universities of compute. If you wanna do interesting training work at a university, very hard to get enough compute to do that. We're just not set up for that. We have power, people. Those are two dimensions. I, I think the Trump administration's done a good job in, in generally relaxing some of the regulations that were painful.
- 1:01:36 – 1:20:11
Quick-Fire Round
- AFAndrew Feldman
- HSHarry Stebbings
Andrew, I wanna do a quick fire with you. So, I'm gonna pummel you with quick questions (laughs) and you gotta give me your immediate thoughts. That sounds-
- AFAndrew Feldman
Y- you know, that's hard because I only have long answers, Harry, so. (laughs)
- HSHarry Stebbings
It, it's totally fine. You'll be... Hey, honestly. Um, wh- what do you believe that most around you disbelieve?
- AFAndrew Feldman
We will have peace in the Middle East in our lifetimes.
- HSHarry Stebbings
Why do you believe that?
- AFAndrew Feldman
Because I believe, uh, having visited and spent time now in, in the UAE, in Saudi, in Qatar, I, I think the returns to being, uh, to moderation, the economic gains. Uh, someone said, we're, we're too busy to hate right now. We're too busy building. I, I think those have been writ large so clearly in the UAE with the rise of Dubai and the UAE e- in, in return for making peace with Israel, in return for a more moderate, uh, position. I, I, I really believe that, that, that, that is the path to the future.
- HSHarry Stebbings
How much of your revenues are from the UAE?
- AFAndrew Feldman
I think in the S1 it says, uh, uh, and that was for, uh, maybe the first half of '24. We haven't published others, but a lot. I think 75, 80%.
- HSHarry Stebbings
I mean this in the nicest way. Do you not have to say nice things about it then? Like if someone's giving me-
- AFAndrew Feldman
Oh, that's, that's a very fair question. No. I, I went there to do business as a Jewish guy, uh, b- before we, uh, before we had any business done, right? I, I, what I found surprised me. And we, we don't do much in Saudi, and I think they're making great strides. And we don't do anything in Qatar right now, and I think they're making great strides. So I, I don't think it's just... It, it may well sort of be colored by the fact that I, I spend time in, in Abu Dhabi and I spend time in Dubai and I spend time in Riyadh and I spend time in Doha. Um, and, uh, sure, it's colored by, by those things.
- HSHarry Stebbings
Why are your revenues concentrated there? Is it just 'cause they're, like, more willing to embrace innovation, new relationships, new vendors?
- AFAndrew Feldman
No. I, I think they, they bought so much, they consumed... And you know that the data I gave you was through the first half of '24. Uh, they placed such big orders, they consumed all our manufacturing capacity. I mean, it, it was... Th- they are building at such an extraordinary rate that, that through the first half of '24, they've consumed an enormous amount of our manufacturing capacity. And I, I think, Harry-
- HSHarry Stebbings
Did their orders exceed, did their orders exceed your expectations of their orders?
- AFAndrew Feldman
I think their orders exceeded everybody's expectations. You, you can be a professional salesperson in Silicon Valley for 20 or 30 years and, and not see a $500 million order. Really. I, I think you can go around the Valley right now and talk to VPs of sales or EVPS of sales at, at dozens of public companies who've never seen an order of that size. They were bold and they were early. And, you know, when we, we started doing bu- business with G42, nobody heard of them. Now everybody in the world's heard of them.
- HSHarry Stebbings
Do, do you think that was a resource planning mistake from you? I mean that nicely.
- AFAndrew Feldman
Oh. (laughs)
- HSHarry Stebbings
But like you said it-
- AFAndrew Feldman
Was it a resource planning mistake? I, I mean, they're all mistakes (laughs) in retrospect, right? I mean, they're... I- if we hadn't won them and we had the resources for it, that would have been a resource planning mistake. I mean, I'm, I'm in the business of, of making big bets and making lots of mistakes, Harry.
- HSHarry Stebbings
What was the biggest bet you've made with Cerebras that didn't work out?
- AFAndrew Feldman
My bets here have, have been pretty good. You know, we... To go to Wafer Scale, to solve a problem that, uh, nobody had previously solved... Gene Amdahl, one of the fathers of our field, failed. IBM failed, TI failed. Uh, everybody's failed at this. We had a period of about 15 months between about 2017 and, and early 2019 where we couldn't make one. And we were running a burn of about six million a month, seven million a month, and we stayed with it and our board stayed with it.
- HSHarry Stebbings
Did you have signs that it would work?
- AFAndrew Feldman
Yeah, we did. And it... We weren't running around like chickens without our heads. We were going through engineering process. Each failure was... You know, we did a full FA, a failure analysis. Each time we fixed the, the, the cause, we did another one, didn't work, did another one, didn't work. And each time, we got a little better. And we got better and better and better. Uh, and then we solved it. And when the founders... The first one that worked, the founders were in a tiny little lab that was a converted conference room, that for cooling, we had the windows open and we'd blown a hole in the wall so we could get external... a chiller outside and pour it in. And when we had it running, the founders stood there together and stared at the box running, which is about as interesting as watching paint dry. And we stood there and we couldn't believe it. It was like, "We have just solved the problem that for 75 years the smartest people in our industry had been unable to solve. And we have done it." And we stood there for like half an hour and it was, it was one of the highlights of, of my career.
- HSHarry Stebbings
Pretty cool. All right, I'll give it to you. All right, fair enough.
- AFAndrew Feldman
Yeah. So that, that was a big, big bet. That, that was a...
- HSHarry Stebbings
No, no, no. Fair enough. Uh, the six, seven million a month burn, I'm like, "All right, fair." That's, yeah.
- AFAndrew Feldman
(laughs)
- HSHarry Stebbings
I'm almost picturing, like, angels singing in this, like tears coming down your face.
- AFAndrew Feldman
You know what? It, it felt like that, and it was, you know, the brainchild of my co-founders, um, Gary and Sean, and JP and Michael. Uh, it was their invention, and, uh, it, it was a physical manifestation of their ideas.
- HSHarry Stebbings
Where are people investing today where they will completely lose their shirt?
Episode duration: 1:20:12
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Transcript of episode FdhAxqR1wVs
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome