AI chip startup Fractile raises $720M Series C to build chips with 20x memory bandwidth advantage over GPUs

Speaker 1: We're running a little late today, but thank you so much for coming on on such a big day. Congratulations. How are you doing?

Speaker 7: It's it's great to see you. Yeah. We're all having a fun time here at Fractile. Yes. Fractile's been on this mission for four years now and things things things are going pretty well. So it's been an exciting day getting to announce a little bit more about what we've been doing as well.

Speaker 1: $720,000,000 raised. Is that a Tony Hawk's Pro Skater reference?

Speaker 7: It's... We did we did $2.20 before. So, yeah, hopefully, we can go for what's next? $10.80? Yeah. Could do a 900 and

Speaker 1: with $12.60, McTwist, something fun. Yeah. Take us through what what the money is for. Is this is this all... Do you need this money to put down your first order? Are you still in r and d? Introduce the company. Tell us about the strategy.

Speaker 7: Definitely. So Fractal is... We're a chip company. We've been building chips to run the world's largest models faster, orders of magnitude faster if we get this right than any other chip on the planet. And this is a pretty full stack endeavor, so you have to do custom silicon, custom software, you have to build these kinda huge power hungry racks as well that these chips go into. Mhmm. And, yeah, the capital lets us do a lot of this faster. It lets us paralyze a bunch of things as well. You know, we're heading into a world where if you wanna, you know, secure capacity on production that you might be doing in the second half of next year, even the first half of twenty twenty eight, you know, you've gotta move today. And so, you know, we're already walking factory floors. We're already thinking about kind of, you know, where that production even in the second half of twenty twenty eight might be built. And so this is really where kinda having capital today. Let's just do things a little bit faster. And then, obviously, it lets us expand what is, you know, already an incredible team. And, you know, we're building a team kind of pretty globally, you know, in in The States, in The UK, in Taiwan. So it's exciting to be able to do a bit more of that as well.

Speaker 1: How much are you in the camp of, like, we are at the end of history in terms of architecture, how bitter lesson pilled are you Yeah. How much of a transformer, you know, absolutist are you?

Speaker 7: Yeah. Great question. Yeah. You know, I think there's two ways. If you... You're crazy enough to see a, you know, $4,000,000,000,000 incumbent and say, you know, we should we should challenge this space. Mhmm. There's sort of two ways you go after it. Right? And one is hyper specialization and you're like, right. Is it... Know, think it's gonna be this workload and it's gonna be this workload forever and we're gonna bake that into the chip. I think that's pretty tricky. You know, chips take a long time to build. Mhmm. Even if we can start designing them in, you know, weeks other than years as, you know, AGI comes and kinda holds a pen for us at all, they take months to get through the fab. And then you have to ramp that production. Your end customer has to find that rack useful for three to five years or the CFO is gonna like blow a fuse because it has to be amortized over that period. And so you can't actually specialize without imposing a bunch of risks. And so kind of the other...

Speaker 1: And that's... That's...

Speaker 2: And that that clearly was a very popular pitch like you know over the last 10... The specialization right when you think about you know Cerebras and and Grok and some of these other players and like maybe you needed to do that at that point to to have a compelling pitch of like, well it didn't feel like things were... It did like certainly you know eight years ago it didn't like feel like things were moving so fast and so the specialization pitch maybe was like landing a bit better and then now investors it feels like have like not a lot of faith that like the the thing that you decided to specialize on four years ago is like the right thing for the next Right. Four years.

Speaker 7: Right. So I think now we see, you know, and with the benefit of that hindsight, there are some things that are sticky and predictable. Models of whatever shape and size will keep getting bigger. They seem to be increasingly hungry for a particular thing that you can design your chips around, which is more bandwidth to memory, the ability to load weights from memory faster, the ability to load today, you know, state, the KV caches that power these models from memory faster. And that's the kind of thing where if you can find a way to build a chip that's, you know, in some way a lot better than something else that is out there, but is placing a bet on a kind of... It's almost like a macro property rather than a micro property. You know, Fractile's big focus over the last couple of years has been, I guess, this sort of secret project, which we're now sharing a little more about to take memory. So this DRAM memory that everybody knows is in, you know, acute shortage today. And rather than having it as like a separate piece of silicon, this is what every single chip today looks like. You've got the kind of processor that's running math and then you have this DRAM memory, HBM memory that's off to the side. Rather than having it like that, we stack it on top. Mhmm. And by doing that and then rearchitecting various things, you have this this huge increase in bandwidth, which translates to a huge increase in speed. You know, on this first gen chip, we're projecting 20 times more bandwidth than you would get to HBM on a on a on a GPU or XPU.

Speaker 1: Mhmm.

Speaker 7: And that doesn't have to be... It's a, you know, it's a bet that is sort of informed by the workloads, but it doesn't have to be kinda hyper specialized. You know, we sort of see all sorts of potential futures where this remains a kind of key, you know, uplifting capability. And so I think, you know, long term, some bitter lesson stuff like stuff scales, things are getting bigger, memory bandwidth requirements are getting higher. What you don't wanna do is kinda build a chip that's the exact kinda shape and size of the network that you're going after because by the time you've even sort of, you know, told the next person about it, you've already you know, the world's already moved on.

Speaker 2: Makes sense. Anthropics IPO materials imply that chips will be a commodity. Is that okay? Can you still build, you know, a trillion dollar business in that in that world? I I can I can make the argument that that you can, but I'm curious your view?

Speaker 7: Yeah. You know, I think I think what becomes more of a commodity is what it looks like to squeeze the best possible performance out of any given chip. Right? And so, you know, I think to some extent, this is a a huge gift to challenges, companies like us that the... You know, I think in those IPO materials, Anthropic said, you know, it used to take six to twelve months to get a workload to run on a new chip that we'd onboard. With the most recent chip we did, this took us three days. Yeah. And so what you start to see is when you look at the chip, in the end actually, one of the only things that kinda matters is, you know, has it got that set of bets right? Has it not been too opinionated? It's still useful for the workload. And then... And does it have some, like, fundamental uplifting property that means that it is just always gonna be better once I throw my model at optimizing, you know, optimizing something for it. And I think that's kind of where we've where we've focused is, you know, some of the kind of more unusual stuff that we do at Fractile is the sort of pavement pounding of constant flights out to to Taiwan to East Asia, working with the supply chain to build sort of these like fundamentally new building blocks. You know, when we started doing DRAM stacking, it felt like nobody else was was doing anything that looks like this. We still think what we're doing, nobody else is doing anything like this. And once you kinda crack that building block, you can then deploy it. And so I think, know, in some sense bring on the commoditization because what it means is you can juice so much more from these kinda core advantages. I think they come through more.

Speaker 3: How are you

Speaker 1: thinking about cyber security? I imagine that you have ongoing conversations with the big labs about selling the product. But are you are you interested or able to get access to cyber security models early because we've been hearing like rumbles about what if there's a p two p p two p to shell like flaw in a CPU or a GPU. Right. Right. And that feels like really really important thing to get right. I'm sure you're taking it seriously, but what has the nature of the process been like for you?

Speaker 7: Yeah. So I think two two separate things on that, you know, one is we are in a scary time. And I think this scary time sees chips become in some ways sort of, you know, our schools of defense. Right? Yeah. We we need to... You know, if you go to the hugging face attack, you know, it's the kind of somewhat heroic tale of defense.

Speaker 1: Right?

Speaker 7: You spin up another model and you use that to fight the agent that is is trying to hack into you. And Yeah. In these kind of... In in these adversarial games, any asymmetric advantage, we see this in anything that looks like this, is a is a huge advantage. Right? And so something like speed, which is what we're delivering, you know, the ability to take a Frontier model and run it 20 times faster than anything else can becomes completely vital to that kind of counter defense. And we don't yet know how kind of asymmetric asymmetric this ends up being. You know, we may really need to have that speed in order to keep ourselves safe. So I think that's a... You know, that's something which we now see ourselves as sort of, you know, potentially quite a vital part of in future. You know, another part of that is, you know, this sort of test time compute scaling, right, is such a core part of capabilities today. The longer I run a model, the more it can do. And to to some extent, if I'm a bad actor, I'm willing to wait. Right? I'm willing to wait six months to have this thing crack into... Know, if you look at the AI Safety Institute, you know, based out of The UK, they have this cybersecurity benchmark, and I think it was Mythos was the first one to crack it. Mhmm. And it got there by generating a 100,000,000 tokens. Yeah. And if you're generating a 100,000,000 tokens at 50 tokens per second on a GPU, it takes twenty three days. Yeah. But if you're doing that at 1,500 tokens, you know, it's kind of like the work of a working day. And so speed and the speed of the chips we have in serving these models kind of becomes the core factor. And and a scary thing is that labs are having to release models that they haven't necessarily had enough time because there's never enough time

Speaker 2: Yeah.

Speaker 7: To play all the way to the end. You know, what does it look like if I'm belligerent enough that I'm gonna wait for this frontier model to run for kinda six months to go do something? And so you also wanna be able to kind of compress that red teaming cycle by running on faster hardware, for example. Yeah. On the safety side at the hardware level, you know, there's this whole world of formal verification Yeah. Which, you know, like, has always been the preserve of kind of like the the brainiest people, you know, in any team on software and hardware. We're seeing huge returns now to leveraging AI to do that. And I think it's a very key part of how we, you know, basically squash all of the bugs, safety bugs in any hardware.

Speaker 1: I love it. Well, congratulations, Jordan. Do you have anything else? Gong. I wanna hit the Gong.

Speaker 2: And... And...

Speaker 1: You really had to scrape the couch cushions to get this round done. Did anyone did anyone like big come into Any this

Speaker 7: have Thrive. Night speed, red point behind us, bud.

Speaker 1: Burners round.

Speaker 2: Capital g. UK sovereign fund. The government. The British business bank too. Wow. That's an elite line on the cap table.

Speaker 1: Killer roster. I'm so happy that you have great partners for this project. It's so important. So thank you so much for coming on the show.

Speaker 2: Yeah. Thanks for breaking it down. Really impressive progress.

Speaker 7: Thanks so much. Thanks very much.

Speaker 1: Hopefully, talk

Speaker 2: to Cheers to the team.

Speaker 1: Have a good weekend. Week. It's not the weekend

Speaker 2: yet. I'm

Speaker 1: going to bed.

Speaker 2: This guy needs a nap.

Speaker 1: Let me tell you about Figma. Agents meet the canvas. Your AI agents can now create and modify your Figma files with design system context and let me tell you about the New York Stock Exchange. Wanna change the world? Raise capital at the New York Stock Exchange. Is there anything in the timeline that's urgent? Breaking news? Anything we need to talk about before we wrap this Veterans Day show on 11/11/2026. Jordy, did we get to the bottom of hobbies? I think we defined it pretty well. It's a hobby

Speaker 2: I think we nailed it down.

Speaker 1: The guy knows your name at the store that sells the stuff for the hobby, then it's a hobby. Otherwise, you're casual. Right? You agree?

Speaker 2: I think that's, you know, given the the decline of the significance of physical retail, I think there's a lot of arguments against that point.

Speaker 1: If you're getting retargeted constantly by the e commerce website Yes. I imagine that the surf shop in Malibu, they know your name. No? Yes. I think so.

Speaker 7: Yes.

Speaker 1: Anyway, thank you for tuning into TBPN. We'll see you tomorrow. Hopefully, we're feeling better. We'll be back no matter what at 11AM Pacific. Leave us five stars on Apple Podcast and Spotify. Sign up for our newsletter at tvpn.com.

Speaker 2: And in honor of Call of Duty Foundation Yes. Which you can support in the game Yes. By buying... The Valor Pack or you just go give

Speaker 1: directly. Never normally been a skins guy. I just sort of play the game for fun. I am a casual. I would not call it a hobby. But now...

Speaker 2: I went to the Call of Duty office....

Speaker 1: Know my name so I guess I'm going to be a hobbyist. Blue ball. But I'm definitely getting the code pack. Call of Endowment. See you tomorrow. Bye.