I worry, to your point, about the linearity of this. Does it sort of slow down next year before the 800-volt data center really takes off and we see a re-acceleration again in 28? It's going to be a job to keep track of all this.
Hello everyone. Welcome to another episode of The Circuit. I am Ben Behar.
Greetings programs. I'm Jay Goldberg.
All right, we've got a range of things to talk about today. Let's start with Analog Devices, sort of what do we say, a quarterly check-in with the analog industry. Sometimes there's pockets of optimism, sometimes people freak out. But anyway, they had a good quarter. Obviously we've talked about there will be a pretty consistent analog cycle, pull-in particularly around data center power and others, but we've noted it's been kind of lumpy I guess. But I'll let you start on the signals from earnings.
So the big sort of event that took place at Analog Devices, they reported earnings this week and industrial and consumer and all that stuff they usually do is doing well, which is important, but the most important stuff is they laid out a clear case for their opportunity in AI. And the analog companies have not participated to the same degree as memory and digital logic companies have. But analog is starting to become more important in the data center. We're using so much electricity. You start to need a whole new category of electrical systems. And we've seen this for a while now because people have been talking about the 800-volt data center, companies like TI and ON and Navitas and Wolfspeed, right? And ADI has been kind of quiet about it, and this quarter they came out and they said, hey, this is what we think our addressable market is. And we're seeing data center is about 10% of their revenue now, and it doubled in the quarter, and they think it's going to double this year and double again next year. And so they're now, you know, they're joining the party.
So interesting just that ADI is sort of the last to talk about it. They're probably the most conservative management team in the space. But also interesting was they scoped out the market for analog, which I don't think anyone else has really done yet. It was interesting. They shared a lot of information. No, I was going to say it was interesting, the talk on margins, where they are pretty high. I mean others have slightly lower margins, which I heard, and I was trying to listen to some of the investor questions also, which was poking at the idea of others having a little bit more pricing leverage to use if they're going to hold at lower gross margins, since I think they're signaling 74% and can hold. But yeah, the margins I think was interesting, which again is just an absurd conversation when you look historically at analog margins, but so be it.
So all right, so on the margin side, I think Analog Devices and TI tend to have higher margins than analog overall. Because they tend to sell directly to customers. Much of the rest of the analog space sells through the channel, where you're sort of competing on price much more heavily. I think there is this weird thing though where you have a lot of these traditional long-standing companies that are now joining AI, and sometimes the questions they get on the calls show people just trying to figure that out and just trying to use the old framework for analyzing these companies in this very new way, in this very new market. And so there's a little bit of disconnect. I think their margins there are a secondary story. I think they're going to do fine. The whole 800-volt data center story for analog is you're going to replace a bunch of sort of lower value parts, not quite commodity but almost commodity parts, with some very, very high-spec parts. Right? That's what the 800-volt data center is, your big transition, in addition to just being large quantities of these chips now. And so the long-term story for analog parts in the data center is this idea of we're selling a lot of high price parts, which is good news for them.
I think to me the really standout feature is they quantified how much analog content goes into a data center, and they said it's a billion to $1.5 billion of analog content per gigawatt, right? And so we have Nvidia saying 50 going to 100 billion per gigawatt to build a data center, and you know, what's that, 2% of that goes to analog, which is small compared to Nvidia, but is very meaningful for companies like ADI and TI. I also really appreciated this: they didn't talk about TAM, total addressable market. They kept referring to their SAM, their serviceable addressable market. Right. And the difference being TAM is like everything that goes into the space. SAM is everything that they can sell into. Right. And so you may not have a part you can sell into the TAM, but the SAM is what you actually can sell into. And I think the knock on ADI up until now has been their SAM is relatively small in the data center. And I think they thought that too. What they said on the call was over the last few months, they did a deep analysis of the market and they realized their SAM was double what they thought it was. They also have an acquisition in there that brings them deeper into it. And so I just like, everybody throws out these big TAM numbers. And ADI, again a conservative management team, a lot like AMD, but ADI's management team threw out the SAM because they're implicitly saying like, hey, we can throw out these big crazy numbers too, but we're just going to talk about what we can actually address. And it's bigger than they thought. So, it was good news for them.
Yep. SAMs are the way to go for those companies who are listening. SAM is a better number than TAM. So we always get this question in market models and data, which is like, what's the real market opportunity? Not how much is — you know, it's nice to say capex is going to be 10 trillion dollars, but you know, what are you actually selling into? Much smaller pie, let's not mislead people, but I get it. Yeah, and I know, and I think again, right, we did a power report a couple months ago just looking at the sheer pull of attach to power as all these data center racks get bigger and pull more power, but also just like you said, 800 VDC, the transition that's going on in just serving power from grid to facility, there's a lot of analog in there.
And it was interesting too, I don't remember if I said this so sorry if I'm repeating everybody, but there's a real interesting way that kind of these analog play for safety. So like when somebody is detaching a rack, so that circuits can actually be managed and shut off accordingly. And so there's a whole new sort of path on what it's looking like for these very specific applications and serviceability and safety also, as the entirety of a cluster and whatnot starts to pull far more energy. So yeah, I mean I think the content opportunity is big. It's good ADI put a little bit of numbers on that. But yeah, analog is going to be a beneficiary.
Yeah. Right. And I think for ADI, my assumption had been that they really served from the edge of the data center where the utility connector where the power comes in, or you know, if you're doing behind the meter or whatever, where the electricity comes in to the wall of the data center itself, right? And then it turns out they actually had content inside the data center. And now what they're saying is they bought this company called Empower that actually brings them actually to the package level. You get the sub-volt conversion stuff. Right. And they talked about optical too, because everybody loves optical, and they seem pretty excited about their optical opportunity, and again they're not sitting next to the transceivers and the lasers, which is I think where the biggest volume is. They're on the control path. So, timers and power management and telemetry kind of stuff, which is again not the massive opportunity, but still meaningful for them. And leaves the door open for them to do another acquisition.
Yeah. I think what's interesting about this, like, they aren't the only example, there's a handful of others kind of in this space, but I would also again say in things like FPGAs, like my report last week around FPGAs and talking with Lattice management, even just conversations with Infineon, NI, like it's interesting that they are sometimes hearing about new opportunities they didn't know about as customers are going deep in the weeds trying to solve engineering problems for compute racks or for power or for liquid cooling, right? And so I just think it's interesting that they'll be like, 'Oh, we just heard of this new application that, you know, we didn't know we solved for X, Y, whatever.'
And I say all of that to say we're in such this moment where there's a lot of engineering challenges kind of across the board, and it's creating interesting opportunities that are somewhat sometimes unpredictable by some of these vendors, because they're being used in a unique way that perhaps they didn't slot for or design for, because an engineer is like, 'Oh, I could use a power PMIC for this, or I can use an analog sensor for this.' So I think that's just interesting in, you know, call it this kind of very technical engineers' dream to solve some very hairy problems in building complex data centers and compute and all of that stuff.
Yeah. I mean for me the really interesting thing is you're seeing the analog companies' data center revenue really running. Like both TI and ADI are doubling their data center revenue, and that's before we get to 800-volt data center, right? Nobody's really building an 800-volt data center, it's all prototyping, right? It's just the design decisions are getting made today to solve those technical problems. The real buildup for 800-volt data center isn't until 2028, and we're already seeing these companies' data center numbers run. I worry, to your point, about the linearity of this. Does it sort of slow down next year before the 800-volt data center really takes off and then we see a re-acceleration again in 28? It's going to be a job to keep track of all this.
This is why I go to all these conferences and I stick my head in these compute racks so I can look at — when I send you all these pictures of microcontrollers and PMICs and analog sensors and capacitors — because I'm like, 'Look how many there are now.' But you're absolutely right. It is. Anyway, I just love the engineering moment that we're in, that people are trying to solve new problems with existing technology, being super creative about it. And I just, you know, again, it's interesting that it's a little bit more unpredictable. Like some of it's predictable, but then some of it's like, 'Well, how are you going to solve that problem?' 'Well, I'm not sure.' And engineers are just being creative, and some company can all of a sudden benefit from that they didn't see coming. So I think that's interesting. That was the only point I wanted to make on that.
Um, all right. So apparently AI has solved cancer, or cancer has been solved. There's now a vaccine, so I don't have to keep going down to the guy down by the river who keeps offering me the cancer vaccine for 20 bucks a shot, right? I'm just kidding. It's always my joke when I'm like, 'Yeah, I just got the cancer vaccine from this guy down by the river.' People are like, 'What?' I'm just kidding. All right, you had a point to make on this, so I'll let you go.
So there was big news in pharma this week. Moderna, the company which created the COVID vaccine seemingly overnight, has announced it came out with a vaccine for melanoma, for skin cancer, and it is novel in that there is a degree of personalization to their treatment, where every patient gets a slightly different combination of chemicals. And there is a heavy amount of compute involved in sort of calculating all that for each patient. And so the headline was that Moderna used AI to cure cancer. And if you go back and actually read their press releases, and I think they did this with Merck, they go out of their way to not call it AI, right? They call it bioinformatics, which is an older term. And I think there was a misperception that they actually used some big AI training model to do this. And that's not the case, right?
I mention this because I think there are a lot of people who assume that AI is going to come up with novel drugs. And I actually think that is going to happen, but it hasn't happened yet. And I think that's important, like, and you can read that as a 50% glass. Is it half full or half empty? We haven't solved it, but maybe we still can. And it makes me think of this company I met when I was in China earlier this year. I went to see this company called Insilico in Shanghai. And Insilico is interesting. It's only about five or six years old. And they make a big deal about their use of AI for developing new drugs. And what they've basically done is taken all of their data, all the data they can, and fine-tuned some LLMs to do sort of every part of the drug processing pipeline, right? But even they are not using AI to come up with new drugs.
Yeah. Right. Yeah, they're enhancing the process of refining it and sort of getting it to approval stages. But it is so interesting, like, this again is a company that is using AI very, very heavily, and they're just using it to enhance the process, right? Which is not sort of novel — the process existed, they're not coming up with new ways to fight a drug. And they've done incredible things, like they're a five-year-old company, they have 30 drugs in their pipeline right now in development, which is a big number for a small company like that. And so, one, I think Insilico is very interesting, but I think it also shows the limits of what AI is able to do today, right? We're not at the point where AI is curing cancer.
I thought it was interesting that everybody that I follow who, you know, I thought was a technology, semiconductor investor, all of a sudden was a biopharma expert. Around Moderna. I was like, they're putting out reports. I was like, 'Wait a minute, didn't you just write about GPUs?' But actually your point is valid. Obviously people are working on how AI can be used to help cure diseases and cancer. There's optimism that that happens someday. Obviously that will play a role, but yes, I think again we have people getting ahead of their skis, to use the term lightly, in some framing.
But you know, I think that will become more of a story in pockets over the next few years about how compute, simulation and running that research can maybe make time to trial faster or something. But yes, AI is not solving all of the world's problems yet. Instead, people are still arguing about data centers showing up in their hometowns and fighting them like it's nobody's business. So sentiment on data center builds is not good.
Yeah. And I think that's part of the dialogue is that AI was so heavily hyped up in its early years in 2023, and people use ChatGPT to make silly memes and then kind of thought that was it. And I have to think if we start to get more meaningful advances because of AI, that will change people's tunes. But we haven't had that yet. And I'm optimistic, right? I'm optimistic like, again, these companies are barely scratching the surface of it. They're using Ampere and Hopper. What happens when they get orders of magnitude more compute in some of these big systems, right? And I think you look at some of the people who are really, really leading that and very, very advanced in their thinking about AI, you know, like some of the guys from Google who just left. I think they think that's all possible, right? And you know, those people were pretty deep into this. So I think it's possible. I just don't know when.
Yeah, I know. I went to, years ago, when Amazon used to do this event called re:MARS, which was sort of just a meeting of the minds of everything automation, robotics, AI, etc. Someone from big pharma, I can't remember who, but came in and basically showed like what they do to go through drug discovery today, and it was expensive, it was cumbersome, yes they threw a lot of compute at it, etc. But the whole point was, you know, what happens when we can speed this up via compute and AI and intelligent simulations grounded in real world science, etc. So that was years ago, so like it's always been talked about.
But to your point, like it's just we're in a weird moment with AI sentiment. You know, I don't know how long it takes to get over that. But it's, you know, you hear about these tensions. It's getting talked about a lot. Data center builds being stalled and whatnot because of counties getting angry. And I'm witnessing this in our backyard as Amazon's building three multiple-hundred-megawatt data centers. And so I'm on these Facebook community boards and I see them sharing like notes from meetings. They're all angry. 'We're going to run out of water. You're going to pollute our ground. You're right next to a pepper farm.' Like anyway, it's interesting. So the joys of this moment of AI in a buildout.
Okay, let's talk about custom ASICs. There was some news around custom ASICs. Marvell and Google have an enhanced relationship that was good for the stock, but also people again misinterpreted this news as a little bit different. I guess there was some knock on Broadcom, which was not the case, which we'll talk about. But if you read the news very clearly, it was definitely — this deal is more for IP and custom work around the periphery, not necessarily the TPU itself, which is what Broadcom does, and then does with MediaTek and other vendors in that design. But that's not bad, right? Being in the periphery, bringing in the optical. I noted very quickly, I think it has a lot to do with CXL around this and disaggregated memory, in kind of the report that we wrote a couple weeks ago about why DDR4 reuse is kind of an immediate TCO unlock. Google's got a lot of DDR4. Marvell's the best positioned in the full stack for CXL. But yeah, people were like, 'This is terrible for Broadcom.' It's like, it's not terrible. It's also a huge growing market in dollars for custom silicon, but you know, Marvell's playing an important role, but it is more on XPU attach, if you will, to use a technical term.