Cerebras IPO: The Tech Breakthrough That Could Change Everything

Josh: There's a chip being built right now in California that the founders of OpenAI Josh: believe is so important that getting access to it could decide who builds AGI first. Josh: In fact, the co-founder of OpenAI just testified last week that he and Ilyas Josh: Zizgever calculated it would take 15 years to build AGI until they discovered Josh: this hardware that could get them there in just five. Josh: The name of this company is called Cerebris and they're actually going public tomorrow.

Ejaaz: For the longest time, it has been a unanimous consensus of who the king of AI is. Ejaaz: It's been NVIDIA and they've own the entire moat on AI, AI chips specifically. Ejaaz: And this moat has made them famous and pretty valuable. Ejaaz: They are worth $5.3 trillion, the most valuable company in the world. Ejaaz: But with success, there's a double-edged sword. Ejaaz: And it highlighted a bunch of constraints that NVIDIA GPUs couldn't really fix.

Ejaaz: One of these constraints is something known as inference. Ejaaz: GPUs that NVIDIA create aren't as hyper-optimized towards that. Ejaaz: And that is the issue that Cerebris focuses on specifically. Ejaaz: They create these chips which generate lightning-fast inference. Ejaaz: What this means is if your AI models run on these Cerebrus chips, Ejaaz: they are lightning quick.

Ejaaz: And in the current AI industry, we're heading towards a world where inference Ejaaz: is valued and matters way more than AI training for models. Ejaaz: In fact, it's estimated to be 10 to 50x as large by JP Morgan in their latest report. Ejaaz: So the point is, if Cerebrus' chips can fulfill that constraint found for inference Ejaaz: for AI models specifically, this could be the first real threat that NVIDIA faces on their own.

Josh: Okay, so let's talk about briefly, before we get into the IPO, Josh: what makes Cerebris so special, so different? Because clearly there's some novel architecture here.

Josh: Perhaps we can use this describing it as the pizza problem where Josh: every chip you've ever owned it kind of starts the same way it takes the Josh: silicon wafer about the size of a dinner plate here Josh: that i have for scale so it's this big giant thing Josh: and it gets cut into a series of tiny little chips that goes into your iphone or Josh: your ipod or your macbook or whatever it is but cerebris Josh: decided to question that they said what if you don't cut the pizza what

Josh: if the whole pizza is the chip and that's exactly what they Josh: did they just took that giant wafer and instead of chopping it Josh: up into smaller chips they just created it all in one big Josh: thing and they had this breakthrough that allowed them to fit Josh: 1.2 trillion transistors on Josh: a single wafer as opposed to someone like nvidia who currently Josh: has 21.5 billion and we could see the Josh: discrepancies on the screen here it is a huge amount of

Josh: additional compute that you could put on a single chip now what does that result Josh: in because the information doesn't have to travel as far the chip is able to Josh: access it all so much faster so they put all the memory on the chip all the Josh: processes on the chip and the result is just significantly faster AI that works Josh: for the entire stack between inference all the way up to pre training. Josh: And this is a really huge win for for a lot of the industry.

Josh: It's a big breakthrough. Ejaaz: And it's not just a theoretical concept, is it either? Like so OpenAI, Ejaaz: about probably like six months ago, invested $10 billion into their company, Ejaaz: and they acquired the rights to basically take over all their chip designs and Ejaaz: use it for OpenAI models specifically. Ejaaz: And we've seen that manifest in their existing models. Ejaaz: So OpenAI has this like world leading coding AI model, right? It's called Codex.

Ejaaz: But what most people don't know is there's a fast mode for Codex. Ejaaz: It's called a GPT Codex Spark. Ejaaz: And it is absolutely rapid. There's like near zero latency. Ejaaz: It's running on cerebrous chips. So the concept isn't a concept anymore. Ejaaz: It's proven at scale to work with these big models.

Ejaaz: And if you are someone that is working on the frontier of software engineering Ejaaz: and, you know, every iteration cycle, the speed at which you design and ship software matters a lot. Ejaaz: Or if you're in financial services and you need to get that product out or make Ejaaz: that trade, you need the fastest chips.

Josh: So these chips are incredibly fast. I mean, we're talking 50 times Josh: more data on a chip than some of the nvidia chips but the Josh: result is that it's very expensive these chips are not cheap to Josh: make the good thing is is that money doesn't seem to be a problem Josh: when it comes to a lot of these ai companies and open ai Josh: has committed to what 10 to 20 billion dollars already Josh: and this is taking us to the headline news of the day which is the ipo the company

Josh: that everyone is super excited about in terms of accelerating the rate that Josh: we get to agi is actually going public and the numbers are pretty big i mean Josh: this is a big deal this is the first AI hardware company to really go public Josh: as we're going on this parabolic trajectory. Josh: And I think the numbers are going to get pretty staggering in a very quick way.

Ejaaz: Yeah. So to give some context on this, about a week and a half ago, Ejaaz: Cerebrus announces or basically files for their IPO. Ejaaz: And there's a prospectus that comes out, right, which explains basically, Ejaaz: you know, why they're raising and how much specifically they're going to raise. Ejaaz: A week and a half ago, that number was $3.5 billion that they were going to Ejaaz: raise. And the The price per share was roughly around $115.

Ejaaz: Something happened immediately after that, though, which is they became 20x oversubscribed. Ejaaz: So there were 20 times more people that wanted to invest in this thing, Ejaaz: qualified, institutional, etc., for their IPO, such that they had to revise the terms, Josh. Ejaaz: So this ended up issuing 2 million more shares at a higher expensive price per share.

Ejaaz: It's 150 bucks. So the total amount that they're raising is actually $4.8 billion, Ejaaz: call it $5 billion, $1.3 billion above what they were expecting to do, Ejaaz: which tells me that there is an insatiable amount of demand for this particular IPO. Ejaaz: Everyone's talking about SpaceX, they're talking about Anthropik and OpenAI's Ejaaz: potential IPOs later this year. Ejaaz: But this is one of the real IPOs that show that there's a huge demand in the West for AI companies.

Ejaaz: And it's exciting to see because this is the first of many.

Josh: Yeah, well, we have Cerebris kicking things off. Then we have the rumor for Josh: SpaceX, OpenAI, Anthropik, Databricks, perhaps even Stripe. Josh: That's a combined potential $3.6 trillion in market cap that could be entering the market. Josh: And we have this hot ball of money. We just filmed an amazing episode last week Josh: about the AI stack and where every company kind of lines up and where the opportunities Josh: are for this new era of AI investing.

Josh: And this sits very squarely right in the hardware era, right next to someone Josh: kind of like NVIDIA, right? Like they're not quite Nvidia's size. Josh: They're nowhere near Nvidia's size, but they're competing on the same kind of Josh: level, the same type of architecture that a company like Nvidia is just at a Josh: very different angle using these huge pizza-sized wafers.

Ejaaz: Yeah, so what I put and want to make on the Nvidia thing specifically is, Ejaaz: okay, so there's one framing of this where if the Cerebrus IPO is successful, Ejaaz: it's the first real signal that Nvidia's monopoly on AI chips is actually going Ejaaz: to be threatened, right? Ejaaz: And this might sound like a crazy thing to say, right? It's a small company, Ejaaz: they're raising $5 billion versus NVIDIA's $5.3 trillion.

Ejaaz: There's another company which proved that NVIDIA's mode might also be shakable, right? Ejaaz: And that is Google last week, who became the most valuable company. Ejaaz: I don't think they're currently the most valuable company, but they beat NVIDIA.

Ejaaz: And the reason why they did that was because they released these brand new TPUs, Ejaaz: their version of their own GPUs, which they train their Gemini model on, Ejaaz: which so many customers, including Anthropic and OpenAI, want to use to train Ejaaz: and inference their own model.

Ejaaz: So the point is, the ground is shifting beneath NVIDIA, even though they have Ejaaz: like decades and decades worth of experience building these GPUs, Ejaaz: there might still be valid, compatible threats against their particular mode. Ejaaz: The other thing I want to say is, and I mentioned this earlier, Ejaaz: inference versus training. Ejaaz: For the longest time, people have said AI training is where all the money is going to be.

Ejaaz: And that's where NVIDIA has lived and breathed over the last decade. Ejaaz: But what has shifted recently is what matters more is what happens after the training. Ejaaz: How many people are going to be using agents that query the models. Ejaaz: Inference is where the money is going to be made, where the value of the opportunity Ejaaz: is going to be huge. And Cerebrus plays directly in that point.

Ejaaz: NVIDIA actually proved this because they acquired another company called Grok Ejaaz: for $20 billion, which does something similar to what Cerebrus did. Ejaaz: So it's validating the fact that Cerebrus is going to be valuable and now the Ejaaz: public can get exposure to it. It's cool. Josh: And I have some older benchmarks about how fast Cerebrus runs versus the rest

Josh: of the world. uh taking a model like llama for maverick uh cerebris was delivering Josh: at 2500 tokens per second which is more than two times nvidia's flagship blackwell chip Josh: And then if you lower it down to Llama 3.1, which is a $70 billion parameter Josh: model, it was 20 times faster than NVIDIA. Josh: So the actual rate that it's able to deliver these tokens is huge.

Josh: And one of the most important things that we're seeing right now is kind of Josh: reasoning, chain of thought, deep thinking of these agents. Josh: And when you're able to deliver tokens at a much faster rate, Josh: you can get from the question to the answer much quicker. Josh: One of the new things that we've seen in Codex recently that has become super Josh: popular is the backslash goal command.

Josh: Where you can set it a command it'll go off and think for 24 Josh: 36 72 hours of time to Josh: collect enough information to get to a single answer this can Josh: make that process significantly quicker and therefore make agents Josh: much more capable so the reasoning speed in terms of inference in terms of delivering Josh: reasoning it's so important and although currently i believe fast mode on chat Josh: gpt is about 1.5 times the cost you have to suspect that that cost is going

Josh: to go down at some point and that makes Cerebris a really compelling company. Josh: And I think that's what OpenA I saw when they made this huge investment in them. Josh: Now, Ijaz, I have to ask you, what are you thinking about this IPO? Josh: Is this something that you're interested in investing in? Ejaaz: It's a tricky one because my default answer is yes. Ejaaz: But if I want to like touch grass for a second to get off my eye holes...

Ejaaz: The IPO is kind of already happening. It's just happening in the private market, Ejaaz: if that makes sense. Like, think about this, right? Ejaaz: NVIDIA, when they IPO'd, I believe was, they IPO'd at around $800 million. Ejaaz: That's million. It's under a billion dollars, right? Ejaaz: And currently we have like Cerebras, which, you know, they have one major customer Ejaaz: as OpenAI. They've proved the concept fair well, but they're raising $5 billion.

Ejaaz: And this is all pre-IPO. So there's a lot of people that are making a lot of Ejaaz: money. we see this with the Anthropic secondary shares, OpenAI, Ejaaz: the same thing with SpaceX as well, before the actual IPO happened. Ejaaz: So the skeptics are saying, this is just going to be a retail liquidity event Ejaaz: where they're going to dump on retail and no one's going to really make any money after that. Ejaaz: I don't think there's a zero chance of that happening, but there is a chance.

Ejaaz: Now, the bull case for this IPO is one specific reason, which is there is going Ejaaz: to be an insatiable demand for two things, Josh. Ejaaz: One is I need to use this AI to do things for me, whether it's agents or whether it's me prompting. Ejaaz: Actually, I'm more bullish on the agent side, but we can tap into that as a second thing. Ejaaz: Second thing is Cerebrus's chips has something very unique that NVIDIA chips can't get enough of.

Ejaaz: And that's something called memory, specifically static random access memory. Ejaaz: It's a very unique type of memory, which Cerebrus installs into their chips. Ejaaz: And it's actually what makes them incredibly unique. Ejaaz: If you look at all the typical 99.9% of GPUs that people use, Ejaaz: they use this thing called high bandwidth memory. Ejaaz: It's composed of something known as dynamic random access memory.

Ejaaz: It's kind of like the sister to static random access memory. It's cheaper. Ejaaz: It's more scalable. It's the sexier thing right now. It's what all the major Ejaaz: memory manufacturers have focused on for the last decade. Ejaaz: And that's what's being installed by default into these GPUs. Ejaaz: But there's one issue. They don't deliver the information that the memory stores Ejaaz: very effectively or quickly enough to allow these models to run very quickly.

Ejaaz: Now, one solution to this is static random access memory, which Cerebris has Ejaaz: focused on for their entirety of their existence. Ejaaz: And they have perfected the design and installation method for their particular Ejaaz: chips. There's a reason why it looks so much larger. Ejaaz: You know, you used that dinner plate just now, Josh, versus the NVIDIA chips. Ejaaz: It is more expensive. You mentioned that the cost to use the things is more expensive.

Ejaaz: But think about how much more value you could get by using quicker turns of your AI model. Ejaaz: You could end up having a super fast quant algorithm to trade your fund or to Ejaaz: deliver access to a product for millions of your different customers. Ejaaz: And that in itself might be worth billions and billions more than having a slower memory component.

Ejaaz: So that is the clear distinction between them and what makes me the most bullish Ejaaz: on Cerebra's IPO specifically.

Josh: Yeah, as I was reading about it earlier this morning, the SRAM versus DRAM conversation, Josh: one of the easiest ways that i was able to kind of differentiate between the Josh: two is an sram is much more similar to an ssd a solid state drive whereas dram Josh: is kind of similar to a htd a hard drive a hard disk drive something that has Josh: an actual needle where it has a lot more capacity but it's a bit slower meanwhile sram

Josh: So long as there is power provided to it, it will retain that memory and information indefinitely. Josh: So in DRAM, one of the big problems that happens is there is this cache memory Josh: and it has to continue to refresh itself over and over because it's a very short shelf life. Josh: With SRAM, that shelf life is infinite and it can be stored as long as there's Josh: power and it can be recalled very quickly without that constant refresh.

Josh: And that's the huge breakthrough. So it's a really compelling product that I Josh: think is very impressive and very necessary in this world. The only real other Josh: GPU architecture we've seen that has had success is Grok. Josh: And we saw what happened with them. They had an amazing exit and they are working Josh: now directly with NVIDIA. Josh: So this is the direct comparison. And for that reason, I think I'm pretty bullish Josh: on the company because...

Josh: Essentially, Cerebrus is tied to OpenAI success. We know that both founders Josh: are actually invested personally. Josh: Greg Brockman and Sam Altman are both personally invested in the company. Josh: OpenAI is invested in the company. And OpenAI has every incentive in the world Josh: to scale with them and to grow with Cerebrus. Josh: That aligned with the fact that this is the first AI IPO we've really had in Josh: this chain that's going to kick things off.

Josh: I expect there to be a lot of excitement, a lot of enthusiasm so long Josh: as the market continues to go up so long as these this token demand continues Josh: i see no reason why cerebrus wouldn't follow in that path and actually be somewhat Josh: successful i think i'd love for everyone who's watching to share their takes Josh: are you investing in the ipo why why not this is uh it's going to be a big conversation Josh: over the next coming weeks as we go through this another

Ejaaz: Thing is like um okay it's one thing having a good product right it's another Ejaaz: thing like getting the right distribution now them signing this like 20 billion Ejaaz: uh partnership chip with OpenAI is good, right? You get access to, what do they have now? Ejaaz: Almost like a billion weekly active users or something crazy like that. Ejaaz: The other thing is their chips are integrated into Amazon's Bedrock platform.

Ejaaz: I don't know if you knew this, Josh, but Amazon basically like owns the entire Ejaaz: enterprise compute market because they give access to every enterprise that Ejaaz: wants to spin up an AWS instance or get access to specific tools through AWS, Ejaaz: their Amazon web service. Ejaaz: Now, they run Cerebrus chips themselves there. So Cerebrus now has a distribution Ejaaz: mode through that, right?

Ejaaz: So the point is, distribution is solved, the ingenuity of going zero to one Ejaaz: with SRAM is also solved, and they have the consumer mode being funneled through OpenAI. Ejaaz: So all of this sounds incredibly bullish, but I do want to touch on the bear Ejaaz: case for a very brief moment, right and i mentioned the first one um earlier Ejaaz: on which is the valuation.

Ejaaz: Is pretty steep, right? The fact that they were 20x oversubscribed and they Ejaaz: just revised the terms in the last five days to give even more access is kind Ejaaz: of, I'm just like, so typically, just for those of you who don't know, Ejaaz: for the average IPO over the last, I think, five years, the price on IPO pops up between 30 to 80%.

Ejaaz: So if you're a retailer, right, that wants to get involved here and like buy Ejaaz: this thing, you might be buying it for like a significant premium. Ejaaz: And these things tend to dump a little bit after the first couple of days. Ejaaz: And then, you know, whatever, it figures itself out. So the valuation is pretty steep. Ejaaz: The other thing I want to say is that OpenAI is also building their own silicon.

Ejaaz: They have a deal with Broadcom, MediaTek, and a bunch of these other intermediaries Ejaaz: to design their own chip. Ejaaz: And Sam has specifically said that one of the major cases that he's trying to solve is inference. Ejaaz: So I'm kind of looking at this and I'm like, okay, so you've invested in this Ejaaz: company, you're going, you know, you're running their chips, Ejaaz: it's doing the thing, but you're also building your own custom silicon.

Ejaaz: Now, granted, that's probably going to take years to try and figure out. Ejaaz: But is the idea that you eventually acquire kind of Cerebrus, Ejaaz: but they're IPOing, so I guess you can't really do this and merge them together? Ejaaz: Or are you going to eventually ditch them and they're like an intermediary solution? Ejaaz: And so Cerebrus might end up like losing this customer going down the road.

Ejaaz: I don't know, but those are the two things that sound out for me.

Josh: That's fair. Yeah, there's also they're trying to compete against the CUDA moat, Josh: which is i mean very powerful and very very strong Josh: as we know they also are i mean they're trading devaluation point Josh: 51 times revenue a lot of these larger ai companies are trading actually at Josh: fairly low multiples right now a lot of the major players in the mag 7 are are Josh: no more than like a 20 times revenue this is at 51 and there's also one interesting

Josh: component where there is actually a memory ceiling currently on the chips where Josh: they're able to fit 44 gigabytes on a chip and currently, Josh: that's unbelievable, but it requires a lot more Josh: Innovation if they're going to continue to scale to provide Josh: inference for multi-modality and for video Josh: models or for ultra long context there's going to need to be Josh: continued innovation and i think that the ceiling is not quite as clear like

Josh: we're not quite clear where this can go to we know it's incredible now i'm not Josh: sure people really know where that's going to go to that's perhaps a longer Josh: term problem that's not something that would really impact the company short Josh: term but it is something to be noteworthy of. Josh: I mean, this is, it's an amazing technology. It's like in your computer, Josh: if you have your RAM and your hard drive,

Josh: as separate components, this merges them into one. And now everything is so Josh: much faster because of it. Josh: And the unlock that has for generating tokens, which is the single most important Josh: thing that we do, is huge. Josh: So I think as a company, Cerebris is something to be very excited about. Josh: This is going to be a fascinating use case or a fascinating case study as they Josh: go public this week to see what that type of impact has on the market.

Josh: And overall, I'm feeling really excited for some bullish innovation. Josh: I mean, Cerebris is a huge unlock in the world of AI. Josh: Sam and Greg saw it first. The rest of the world is going to see it this week Josh: as they go public. And we'll see what the downstream effects of it are. Ejaaz: I mean, it's the first domino of many, right? I think another obvious reason Ejaaz: why it's getting a lot of attention is like, it's AI summer, right?

Ejaaz: We've got a bunch of cool IPOs that are coming up. Ejaaz: You know, I've got a graphic here, which is actually severely outdated. Ejaaz: It's funny, it values Anthropik's IPO as being around $230 billion. Ejaaz: And I'm like, have you looked at the secondary market for their shares recently? Ejaaz: They're selling at like at a trillion dollar valuation. People are literally Ejaaz: leasing their homes out to be able to like fund like a purchase of Anthropic secondary shares.

Ejaaz: It's pretty crazy out there. But the point being is you've got SpaceX. Ejaaz: You've got Open Air, you've got ByteDance from China, you've got Anthropic that Ejaaz: are all planning to IPO this year. Ejaaz: There is a lot of money that's going to be entering the market. Ejaaz: Some of it will be net new coming in from retail, coming in from funds that Ejaaz: don't typically invest in these things. Ejaaz: But a lot of it is going to come from alternative types of companies.

Ejaaz: Now, we've seen signals of where this is going to come from. Ejaaz: Whenever, for example, whenever Anthropic announces a new product or feature Ejaaz: that replaces a certain segment or industrial sector, you see the top market Ejaaz: cap stocks of those sectors dump as they've tweeted the announcement out, right?

Ejaaz: So it might be like SaaS companies are cooked or whatever that might be, Ejaaz: but it'll be interesting to see the market dynamics because I think there'll Ejaaz: be a lot of money sloshing around into these big IPOs. Ejaaz: And then the major question that skeptics have is, Ejaaz: They don't think there's enough money. And this might be, you know, Ejaaz: the popping of the bubble metaphorically or financially. Ejaaz: We might see a bunch of these like crash after the liquidity event.

Ejaaz: That's more of the bad case, but I'm choosing to remain bullish on this because Ejaaz: I think the demand is pretty insatiable, right? Ejaaz: Today, OpenAI announced like a huge partnership with enterprise for deployed engineers. Ejaaz: Anthropic did the same thing five days ago. So they're generating highest revenues. Ejaaz: Like Anthropic hit 45 billion ARR. they're aiming to hit 100 by the end of the year. Ejaaz: These aren't made up numbers. So, you know, I'm excited about it.

Josh: Unbelievable businesses. And yeah, I think to your point, there may be some Josh: liquidity issues as we get to IPO number four, five, and six. Josh: This is the smallest of all or among the smallest of all of them that's happening. Josh: So I'm not exactly concerned about that yet, but it's going to be a crazy year. Josh: And this will certainly set some sort of a precedent for how the market considers IPOs. Ejaaz: Are you buying some of this tomorrow? I have to ask you.

Josh: You know what? I think it'll depend on how it trades. I probably won't buy it Josh: at the open. Maybe wait a little bit. Ejaaz: You're going to wait for that 30 to 80% premium? Josh: I think the answer is yes. Like, I feel like I should, I should participate Josh: at least in a small amount in an IPO like this, because even if it is overpriced Josh: on the short term, it'll reprice itself over the next few weeks.

Josh: And I feel like six months from now, this company will produce better products than it does today. Josh: And the same will also be true 12 months from now. So I'm not much of a trader. I'm more of an investor. Josh: This is a compelling investment. And I could also see a world in which it does Josh: get acquired or a majority of it does wind up getting bought out by another Josh: company. It's just incredibly important infrastructure.

Josh: It seems valuable. We saw what happened with Grok. We might have an opportunity Josh: to actually participate in it again with Cerebris. Josh: I'm feeling bullish. I'm feeling like I want to get in the mix. Josh: I think I want to get involved a little bit. Ejaaz: Same. Hey, I'm down to buy NVIDIA at under a $10 billion valuation. Ejaaz: If it even achieves one of 10% of NVIDIA's market cap, that would be a win. Ejaaz: So yeah, I'm excited to see how this plays out.

Josh: Yeah. Assuming that thesis holds up where we continue used to need tokens and Josh: those tokens were able to generate a profit, the people who can distribute and Josh: produce those tokens faster than everyone else will win. Josh: And I think that is a valid enough thesis to have me convinced. Josh: But to the listener, to everyone out there who has just watched 20 minutes of Josh: this episode, do you agree? Josh: Are you participating in Cerebris? Is this overpriced? Is it underpriced?

Josh: Is it stupid? Should we even be investing or considering? Josh: Let us know in the comments down below. If you enjoyed this video, Josh: please don't forget to share with your friends. Ija, any final parting thoughts? Ejaaz: I'm actually curious whether the listeners of the show enjoyed this episode, Ejaaz: we do various different types of investment ones, the ones that we did last Ejaaz: week on the AI info stack got so much attention.

Ejaaz: And we're going to do a follow up episode on that just we heard we read all Ejaaz: your comments, we're going to do it. Ejaaz: But on something like this, like, is this something that's interesting? Ejaaz: Do you want to hear about upcoming IPOs? I'm curious to hear from you guys.

Ejaaz: But aside for that, I think we're all good.