Cerebras CEO Andrew Feldman: We're 10–20x faster than GPUs and 75% of our 2027 data centers will be outside the US

Nov 17, 2026 · Full transcript · This transcript is auto-generated and may contain errors.

Featuring Andrew Feldman

Speaker 3: he is.

Speaker 1: We are very lucky to be joined by Andrew Feldman from Cerebras, the cofounder and CEO. And he has the chip with us.

Speaker 3: Always. He's got the Keeps it on him.

Speaker 1: Let's put this right There

Speaker 11: you go.

Speaker 1: We'll keep it on you.

Speaker 6: Great to

Speaker 1: see you. You gotta get a suit pocket that's specially sewn to tuck it in there. I think that's

Speaker 11: That was the the best idea.

Speaker 6: Yeah? Yeah. You're gonna do it

Speaker 11: to me last night. Okay. I made a note.

Speaker 6: I sent it my wife. Fantastic. I said, I need

Speaker 1: a tailor can do it in your suit. They can sew it in there. Right. You

Speaker 11: can get pants. Look. One one of the things that is so great about being a hardcore guy is that you get show and tell. Yeah. That's right. Show and tell. Guys, there's nothing. Yeah. There's nothing for them to show.

Speaker 1: But Yeah.

Speaker 11: This is the largest chip in the history of the computer industry. The fastest AI chip. Yeah. This is our wafer scale engine.

Speaker 1: Where does this begin? Is it... Does it start with sand? It does... Are you buying the the wafers and just telling TSMC don't cut them up? Or how does that work?

Speaker 11: It it it starts on the TSMC side

Speaker 1: Okay.

Speaker 11: With sand.

Speaker 6: Okay. Sand.

Speaker 11: Right? And they create a molten goo. They pull out an ingot. They ingot it sliced. It comes out as a a a round disc. Yeah. The disc is 12 inches in diameter. Yeah. They etch our transistors in it. Sure. And then we cut the largest square Yeah. That we can fit out of.

Speaker 1: But you're cutting a square out of a circle. You're by definitely losing a little bit of surface area.

Speaker 11: We are.

Speaker 1: Is there any hope for a true wafer scale circular chip, or is that not functionally possible for some reason?

Speaker 11: I I think there were a bunch of technical reasons why we why we chose to to remove the ears Sure. Off the the edge. Okay. Yeah. Square's a nice way to to to to place a chip. It has nice characteristics.

Speaker 1: With cooling and whatnot. And you vertically integrated that. Right? You have your own liquid cooling stack. Is that correct?

Speaker 11: That's correct. So we don't just make cool chips.

Speaker 1: Yeah.

Speaker 11: We we we build the the entire system. Yeah. The server into which the chip goes.

Speaker 1: So I need to bring a powered shell and you can handle almost everything else?

Speaker 11: You need to you need to bring a powered shell and we'll take care of else. We'll roll in the compute. Yeah. We will deliver the API to the customer.

Speaker 1: And so

Speaker 11: you can just go to us and and get fast fast fast inference, the fastest on Earth.

Speaker 3: Talk about the current moment and how Cerberus fits in specifically with cyber. I mean, I think it's on everyone's mind. Everyone's freaked out about it. Everyone's trying to spend money to put themselves in a better position. What what role has Cerberus played in in arming defenders?

Speaker 11: Yeah. So one of the the the things about about cyber is time matters.

Speaker 1: Yeah.

Speaker 11: Right? Right? A lot of damage can be done in minutes or or even seconds. And so what you want is the minute you are alerted to an issue, you wanna take action quickly. Or better yet, you you wanna be able to prevent these things from happening without imposing these enormous taxes on us.

Speaker 1: Yeah.

Speaker 11: Right? There's always a balance.

Speaker 6: Your engineers like, why do I have to

Speaker 11: log in 11 times to to to to my work? Why does it have to be... You want all of that behind the scenes and you want it done so fast that it annoys no one.

Speaker 4: Yep.

Speaker 11: And that's what we do. We we specialize in fast inference. Mhmm. And fast inference allows you to scrub traffic. Mhmm. Right? Removing threats. It allows you, if there's already been an attack, to quickly understand and recognize what it is and take action. But in all of these cases, speed matters. And we're not a little bit faster than than GPUs. We're 10, 15, 20 times faster. So we set the bar in how fast you can get answers.

Speaker 1: I saw recent, there was a viral launch from a company called Jeff and I think OpenAI, and I believe Gemini launched the decisions API. Are there any other new architectures that aren't just the the the the linear progression of scale scale scale, bigger transformer, LLM that is maybe a little spiky, but it can kinda do everything? We've seen diffusion models. We've seen these decision API models. Are there any new models or architectures that you're particularly excited about plugging into the Cerebras ecosystem?

Speaker 11: I I I think you're right. We... We've seen tremendous advances on the closed source Yeah. Closed side. Yeah. You know, both OpenAI and Anthropic. You've seen tremendous advantages on the open source side with with Kimi and and GLM, you've seen new types of models Mhmm. On these decision models.

Speaker 1: Yeah.

Speaker 11: I I think that there is a battle for the hearts and minds of the AI community about whether we want large models to do many things Sure. Or we want a collection of small models that each does one thing well. Yeah. And and that's a battle that's gonna continue and continues in in most healthy ecosystems.

Speaker 6: Yeah.

Speaker 11: Right? It's true for for many of the the the software landscapes. Right? You you can buy a package that does one thing very well or you can buy... Let's take security. Yeah. There there are specialized security vendors or there's others like Palo Alto that can give you 20 or 30 or 40 or 50 different software packages all for security. Yeah. And so there will

Speaker 6: be the same thing in

Speaker 11: the AI world. There'll be people who want one model to rule them all. There'll be people who prefer a collection of models, each specialist in a particular domain.

Speaker 1: Won't the economics and the sort of gravitational force of the economics lead us to one or the other? Like, either the one model becomes the cheapest token per watt, and you use it for everything. Or do we wind up in the ensemble approach, which I think is more likely? And there's a little bit of on device inference and then some wafer scale inference. And then there's some big GPU racks that are, orchestrating sub agents. And there's teams, and there's some programmatic stuff running on CPUs. There's a whole host of these. That feels like the likely outcome. So who would be staying in the monolithic model world? I... And I'm not saying that the labs won't have their own ensembles internally. Some of them might be running on Cerebras, but it feels like the end state here is use every bit of silicon. Use every power source combined.

Speaker 11: A 100 agree. Okay. 100% agree. I I think there is... Just like in in nature Yeah. In in business ecosystems, a constant battle between the generalist and the specialist. Yeah. Right? Yeah. The the generalist wants to to to sell you one thing to rule them all. Yes. Platform.

Speaker 1: Yes.

Speaker 11: One one chip. Yes. One whatever.

Speaker 1: The 10 in one shampoo that you can brush your teeth with.

Speaker 11: That's exactly right. Yes. And and there are others. And and in some cases, you you don't want 10 little tubs, 10 little tubes sitting in your bathroom. Right? When you travel, you don't wanna And and so they all have their place. Yes. And I I think you're exactly right that even in the background Yeah. Of the big models, what you're gonna have are collections of

Speaker 5: Of course.

Speaker 1: Different models. Because abstracting it, I... As a consumer, I want an... I want just a text box, and I can say, go build this thing. And it spins off sub agents and uses the best tool for the job. Sometimes it's just pulling an open source package. I don't even need you to write it. Just pull that code.

Speaker 11: That's right. Right? I I think that the guys who who were doing routers early on Yeah. Real... Really sort of got to the nub of that issue

Speaker 1: A 100%.

Speaker 11: That traffic's gonna come in and and you you don't need, you know, you don't need a Michelin chef to make your breakfast. Yeah. Right? It's it's okay to have some eggs on toast. Yeah. And that... That's sort of... That's okay. And the same in the the the the network I model eat

Speaker 1: Michelin chef. Sure. You're sorry. Mentioned this to the guy who drives a Ferrari to the grocery store.

Speaker 4: You know,

Speaker 5: it doesn't really make sense.

Speaker 1: Don't do that. And

Speaker 11: don't drive it to Home Depot.

Speaker 1: That's what the first song

Speaker 3: was for. So question. So we've heard takes on the the Jensen exit from public markets investors, private markets investors, employees, things like that. The people that aren't going on podcasts and talking about this really broadly are are... Is like the people across the semi supply chain. What conversations are are those groups having? And how are they how are they thinking about heading into this next year with with what what seems, you know, at least very unlikely that Jensen's in the driver's seat?

Speaker 11: Well, you know, I'm I'm not in the conversations with with with NVIDIA and and their supply chain partners. And one of the things that...

Speaker 3: No. But I imagine there's some overlap.

Speaker 11: There is some modest overlap. But one of the things you do when when you're David going up against Goliath is you think really hard about how to not be in places where their size matters and supply chain is one of those. And so we've worked really hard to build a a supply chain that has as little overlap as possible. Sure. So for example, at TSMC, we're on the five nanometer node where we don't compete with them for Yeah. For capacity on three. We don't use HBM memory where they're one of the largest buyers of memory on earth.

Speaker 6: Yep. Yep.

Speaker 11: And so we... We've worked really hard strategically to to to to not have to duke it out in the supply chain where where where they have strength that it's hard to compete. We wanna compete where we're good

Speaker 1: Yeah.

Speaker 11: Yeah. And where we're fast and where we have significant advantage. And and that's no different if they're big, make them run. Right? It's true. Here we are in a basketball stadium. And... Right? I mean, if if you've got a small team, you better run.

Speaker 1: Yeah. And and

Speaker 11: and that... That's the same in in everybody trying to think about the other guy's strengths Yeah. And what those have... What those weaknesses are and how to play into their weaknesses. Yeah.

Speaker 1: That... Can I ask what? Yeah. That... So that running cadence that you mentioned, I'm interested, like... Ben Thompson was looking back on Jensen's tenure, and he was saying that when when AI started happening, he really scaled up the intensity, you know, six keynotes a year, annual release cycle. What does the... What what is your goal for for release cadence of new chips, features? Like, what is the cadence that you run the company on? Brian Chesky from Airbnb said that there's, like, an incredible value of just forcing the whole team to ship and have a keynote even if it's an ever evolving software product in his case. How do you think about the cadence that Cerberus needs to continue and win?

Speaker 11: Well, for first, look, you gotta take your hat off to to to Jensen. Yeah. Thirty three year run from zero to building the most valuable company on Earth.

Speaker 1: Yeah.

Speaker 11: Singular. That's right. That that is unparalleled. Yeah. Right? In in in our history. Yeah. And what what an amazing run and and sort of unquestionably one of the top two or three CEOs of the first quarter of

Speaker 1: the century.

Speaker 6: Yeah.

Speaker 11: Right? So... Alright. So we we set that that bit aside. I I think the the the the pace right now in the AI industry is is difficult for those outside the industry to comprehend. Yeah. I mean, it's very common to to go out for two dinners.

Speaker 1: Right? To to

Speaker 11: to have a dinner at six Yeah. And then meet someone at at a two business dinners. Right?

Speaker 1: Yeah.

Speaker 11: The the number of events that are being done,

Speaker 6: the the fact that that Saturday meetings at nine are very common Yeah.

Speaker 11: In the morning and in the

Speaker 1: evening. Yeah.

Speaker 11: That that everybody's on all the time. I mean, the the intensity in this industry right now is unbelievable.

Speaker 1: Yeah. Yeah.

Speaker 11: And I... I've been an entrepreneur for the past twenty seven years and we've never seen anything like it.

Speaker 1: Yeah.

Speaker 11: The raw intensity and the drive right now. Yeah. To the specific question Yeah. Of does it make sense to have have an event

Speaker 6: or do... You know, my

Speaker 11: dad's a professor and he used to say if if there weren't conferences, nothing ever would be finished.

Speaker 6: And that you could...

Speaker 11: You can use these events to actually drive behavior.

Speaker 6: Yeah.

Speaker 11: Right? To to force clear thinking. Yep. To to to force releases. Yep. To to to use the cadence of of the public events Yeah. To help motivate the team and drive the team and and and create hard deadlines. And I I I think that's something that that we've seen across the industry. Yeah. We're producing a new release about every year for the hardware for the next several years.

Speaker 1: That's

Speaker 11: great. We're doubling performance in each one.

Speaker 7: Yeah.

Speaker 11: Yep. So you you get fast in a hurry that way.

Speaker 1: Sure you're also very reactive to where the technology is moving, what the labs want, what the buyers want, the different architectures. Like, you can change and evolve Of of as you as you iterate.

Speaker 11: Of course. I mean, the the the hardware is the the way to think about it is maybe your your your muscle and the the the software is the brains. Yeah. And you gotta bring them together to meet what your customers need. Yeah. And, you know, week to week, month to month, they don't even know

Speaker 1: Yeah. Yeah.

Speaker 11: What what they need. I mean, the landscape is changing so fast that that their requests change within a week and certainly across several weeks.

Speaker 6: Yeah. Yeah. That makes a lot sense.

Speaker 3: Last question for me. We talked about the arms race in in cyber. What's your... What what should we know about what's happening in trading? The market right now moves so violently off of individual headlines or what what individuals here are are posting. And I imagine it's the same thing. You know, these shops wanna be faster than... It's PVP. They gotta be the fastest. Otherwise, they're gonna... You know, returns will be compressed.

Speaker 11: Yeah. I I think one of the things that that sort of is is now common in the the markets is the speed of diffusion of information is is unparalleled.

Speaker 3: Yep. But you need to process that information to make a decision.

Speaker 11: That's right. I think sometimes that information is bogus Yeah. As we've seen. And the the gap is people want information processed and good decisions made as fast as they can. And right now, what we're getting is whipsawed back and forth. Right? Every burp or hiccup is causing wild fluctuations. Yeah. I think one of the things we can do and and products like ours can do when you're fast is get that information, synthesize it, come up with recommendations that are thoughtful and interesting. Yep. And not just knee jerk. And I I think that's one of the roles AI will play in the markets in the future.

Speaker 6: Yeah. Great.

Speaker 1: How much demand are you seeing for Cerberus in The United States versus anywhere else?

Speaker 11: You know, our our demand right now is is overwhelming. Yeah. Yeah. Overwhelming. Yeah. And we're constrained by the number of data centers we can put up.

Speaker 1: Okay.

Speaker 11: And so... And and even in our... The part of our business where we sell hardware

Speaker 1: Yeah.

Speaker 11: Our customers are constrained. By the number of data centers they can put up or they can rent. Sure. And and so we are seeing across The US and Europe and starting in Latin America, in Asia, just overwhelming demand for fast inference for getting answers in less time.

Speaker 1: Where's... Where... What is interesting in Latin America? I mean, we've heard a little bit about what's going on in Europe. I'm much less familiar with what's going on in Latin America. What countries are excited about build out there?

Speaker 11: I I think we are embarking on a project in Brazil.

Speaker 1: Okay.

Speaker 11: And I think they have... They've just changed some of their tax structure Yeah. And and made it easier to bring in equipment by lowering tariffs. Sure. And A

Speaker 1: lot of it's gotta be permitting. I mean, that that was the fallout of the midterms. There's a question about our permits gonna get approved. We have a bunch of other founders on the show who actually build the data centers and can can contextualize that. But I'm interested to see... It does seem like if America is slowing down on permitting, there's a natural reaction from the international community. For sure.

Speaker 11: I I I think... And you've got Chase coming on Yeah. Crusoe.

Speaker 1: Yeah. Yeah. Exactly.

Speaker 11: One of the most interesting startups Yeah. And doing really, really cool stuff. Yeah. I think I think he'll tell you the following, that for data centers for 2027, it it... They've already been permitted.

Speaker 1: Yeah. Of course.

Speaker 11: Of course. They've been permitted. Concrete's being poured. Totally. Totally. Bulldozers are moving. Things are being deployed. So the the the... What we saw in the the elections

Speaker 1: Yeah.

Speaker 11: Will have an impact in '28.

Speaker 1: Sure.

Speaker 11: It usually takes fifteen to eighteen months from the time a permit is is is delivered in order to build a a facility. So Yeah. Right? If permits start to slow, what we thought would be built in The US in '28 gets pushed out.

Speaker 1: Yeah.

Speaker 11: And that's a problem for sure.

Speaker 1: Yeah.

Speaker 11: We we were a bit ahead of that. We we saw this ground swell coming Mhmm. And we began sort of contracting for data centers outside The US. In '27, about 75% of our data centers will be out of The US. Woah. Yeah.

Speaker 1: It's high.

Speaker 11: Wow. Yeah. It's high. Canada Yeah. It... Canada really likes data centers being built.

Speaker 3: They were they were, you know, some politicians there I saw kinda celebrating the midterms being like, this is our opportunity to get a piece of this industry.

Speaker 11: I I I think that's...

Speaker 3: And it pains me.

Speaker 11: Yeah. It pains us too because part of it the... I I don't think the the data center industry did ourselves any favors.

Speaker 1: Of course not.

Speaker 11: Right? We we we could have gone in as a community and been good to to to local.

Speaker 6: Yeah. The funny is everyone was like, hey.

Speaker 3: The data center word is unpopular. Let's call them AI factories, which we saw was the dumbest possible.... Of any kind. Why would you want a factory in your back?

Speaker 1: You don't wanna live next to a toothbrush factory.

Speaker 3: You had... We had supercomputer right there.

Speaker 6: I know.

Speaker 3: We could've just called them computers. We're creating a computer center.

Speaker 11: Right.

Speaker 10: Yeah. And you...

Speaker 3: It's in your town. It's gonna give you a bunch of tax revenue.

Speaker 11: Right? It's gonna give you tax revenue. The construction jobs are large and meaningful. Yeah. Totally. Alright. The tax base changes. Yeah. And while you're there, why don't you ask us to build a football field for the high school or pay for the high school's power for for the next ten years? I mean, those are things that we could be doing Yeah. For for these communities. And I I I think there was an opportunity missed by the early wave of data center builders in

Speaker 1: the Totally. It was just go. Go.

Speaker 3: Go. And and

Speaker 1: and it's been 20 of no pushback.

Speaker 12: That's

Speaker 1: right. All the pushback was on the chemical plants, send those overseas. The the industrial factories, send those overseas because smoke stack. Don't want the smoke sack.

Speaker 11: Compared to those, right, these are the cleanest factories on earth.

Speaker 1: Totally. But but the data center builders got got complacent.

Speaker 11: Such a bad job Yeah. Of PR marketing.

Speaker 5: A 100%.

Speaker 11: And in community engagement that we... An opportunity was missed

Speaker 6: Yeah.

Speaker 11: Across the board. Yeah.

Speaker 1: Yeah. No. No. A 100%.

Speaker 3: Are are you are you hopeful that that the data center mix for '28 and '29 for Cerebras could shift some market share back to The US if some of these, like, new approaches get that... Get dialed in?

Speaker 11: I think they should. Yeah. I I think we we would like data centers in The US to serve US customers. Yeah. And I I think I know guys like Chase and others who are doing this, are working really hard with the with the local communities. There's a community that's a few hours outside of of Austin where we're sitting today and, you know, we were thinking of a data center there and I I called the mayor and and I said, look, if if you don't want us to to to work with this data center developer and you don't want this in your town, let's not do this. Yeah. Right? This is just that easy. Let... That's right. I understand. Let... Let's see if we can can can can do something that's good for for us and for you.

Speaker 1: Yeah. Yeah. What is the sweet spot for a Cerberus data center? We hear a lot about these massive data center campuses, a gigawatt. Those are the ones that make the headlines. But Right. Is 10 megawatts viable? Is that economical? Like, what what... Where is the curve? What is the shape of the curve for delivering value?

Speaker 11: That's a really important question. For training

Speaker 1: Yeah.

Speaker 11: You want a huge amount

Speaker 1: Yeah.

Speaker 11: Of computing power in one location.

Speaker 1: Yes.

Speaker 11: So the the gigawatts scale facilities Yes. Are generally for training.

Speaker 1: Yeah.

Speaker 11: For inference, 50 or a 100, just fine. Yeah. We can put them in different locations. Sure. In fact, there are benefits to having them in different locations because the the latency, the Yeah. Time the transport time Yeah. To get to and from the data center is reduced.

Speaker 1: Yeah.

Speaker 11: Yeah. And so fifties and hundreds, one fifties, those are all two hundreds. Those are all great for for inference. We do not need these mega campuses that are are are so large and so power concerned.

Speaker 1: Yeah. Yeah. It does seem like there's also... Don't I know. Maybe it's entirely misinformed, but there's something that's probably more reassuring to the average American voter of just like, oh, well, I have a PC over there or Mac mini, and, you know, it's running. And maybe I have, you know, a small data center that's whole... In my town that's hosting my Netflix and that's more relatable than the boar is over there in West Texas. It's scary.

Speaker 11: Think that's true. I think there are a couple of things. I think they're afraid of AI. Data centers are the physical manifestation Yeah. Yeah. Point at.

Speaker 1: Totally.

Speaker 3: Yeah.

Speaker 11: Right? And I don't think we've done a good job of of sharing all the benefits AI can bring.

Speaker 1: Yeah.

Speaker 11: I I think we've A 100%. There... There's been a focus on on sort of the fear and I

Speaker 6: I think that was a mistake.

Speaker 11: Yeah. I I think what we wanna explain to people is that you get a world class tutor for your kids

Speaker 1: Yeah.

Speaker 11: Sitting right there on chat GPT.

Speaker 1: Yeah. That's...

Speaker 11: Right? You can teach your your your kid how to to to work on the the problem for homework that in in a way that they understand it best. You can just keep asking it to try teaching it to me in a different way.

Speaker 1: Totally.

Speaker 11: Right? And there are amazing things that that that can be delivered. And and instead, I think we've in part managed to to create fear.

Speaker 1: Yeah.

Speaker 11: And that the dislike of data centers, think, is the sort of the something to point at. You can't point at software, but you can point at a building.

Speaker 1: Yeah. 100%.

Speaker 11: And that's that's that's unfortunate.

Speaker 1: Yeah. Well...

Speaker 6: Hey. Thanks guys for having me. Appreciate it. It was great to to see you guys. Thank

Speaker 3: you, guys. Do it

Speaker 1: again soon. Alright.

Speaker 3: We need our own we need our own chicks, by

Speaker 4: the way.

Speaker 3: Yeah. Yeah.

Speaker 11: Can we keep one at at your offices? Yeah. Yeah. We'd love

Speaker 1: to have one. Alright.

Speaker 11: We'll we'll organize one for you.

Speaker 6: Fantastic. Thank you now.

Speaker 1: Great to see guys. You. Let me tell you about MongoDB. What's the only faster than the AI market, your business on MongoDB? Don't just build AI. Own the data platform, and that powers it. I believe we have Keller Cliffton. Surprise guest. From Zipline coming on, or do we switch it up? We're doing Keller's.

Speaker 6: Keller's here.

Speaker 1: Welcome back to TBPN. Talked to you a couple weeks ago.