Systems Foundry in the AI Era

Patrick Moorhead:
Welcome back to the Six Five Summit 2024. This summit is all about AI, surprise. Two elements of it. First of all, the continued building out the infrastructure for AI that we need all the way from the edge to the data center and everything in between, and also the benefits that consumers and enterprises get out of it. And we’re going to be talking about one of my favorite topics here, which are semiconductors. Surprise! And also that we’ve seen semiconductors just pretty much make all of this AI goodness move forward.

One thing we haven’t talked about yet on the Summit is that you actually have to build these in foundries. I know, shocker there. It’s a really hard thing to do. One of the companies that is blazing new trails in AI and foundry is Intel, and I have Craig from Intel to chat all about this. Craig, welcome to the Six Five.

Craig Orr:
Hey Pat. Thanks for having me.

Patrick Moorhead:
Absolutely. So it’s been a huge year for Intel Foundry. You had your first customer event in February. You broke ground in Columbus. Can you talk about any noticeable big milestones that you can share with the crowd, just to get them up to speed?

Craig Orr:
Yeah, sure. So this year, like you said, we’ve had a lot going on. I think my wife is wondering why I’m never home, but that’s partly because of all the stuff we’re doing. And so we’ve had a lot of support. The thing that strikes me the most, at least, is we’ve had a lot of support from the customers, from our ecosystem partners, and from the government. We’ve had a lot of actual really tangible results along each of those dimensions so far. So we have customer wins across all of our technology platforms. We have, especially, the six wins on 18A. And we’re well on our way to becoming the number two foundry in the world.

We had announced at Direct Connect all of our technological progress and our future roadmap. So of the big five nodes in four years plan that I had announced a couple years ago, we have three of those done and we’re on track to deliver the last two, 20A and 18A, so we get back to process leadership here in 2025. Then with the ecosystem, we announced partnerships with Cadence, Synopsis, ARM, and we actually had all of their leaders at our Direct Connect event, which, if you’ve followed Intel for a long time you’d be really quite shocked that Rene Haas was standing on stage with Steve.

Patrick Moorhead:
Craig, I have been following Intel for 35 years.

Craig Orr:
Did you ever think you’d see that, right?

Patrick Moorhead:
Actually, no. But it’s great to see.

Craig Orr:
Yeah. And we had a great list of industry speakers. We had Sam Altman had Satya Nadella making a virtual appearance, Gina Raimondo, Secretary Raimondo, and then a bunch of customers as well. And then I think the final thing is, on the government side, we announced the CHIPS ACT funding, which was very exciting. It’s a total of like $44 billion, $8.5 direct, and then $11 in loans, and the 25% tax credit on the $100 billion investment. So we’re getting a lot done here in a very short time.

Patrick Moorhead:
Yeah. I have to tell you, when your CEO, Pat Gelsinger, rolled out five nodes in four years in the context of your IDM 2.0 strategy, I was like, “That seems aggressive. Let’s see where this goes.” And to your credit, you’ve been able to hit… you’re over halfway there. What I’d love to do is dive into your approach and your strategy. It’s a system foundry approach, which I think you had mentioned was really good for this AI area. Can you drill down on this system foundry approach and maybe compare it to non-system foundry approaches?

Craig Orr:
Yeah, yeah. One of the things that became pretty obvious from talking with our customers and from observing the trends in the industry is that the way that things have been set up until now has worked very well until now. But with AI coming online, we need a whole entirely different approach. It’s still very early in the AI era, if you will, but already it’s getting pretty clear that the way things are done today in the industry, the path we’re currently on, it’s not going to be sustainable from either a power perspective or from a sustainability and reliability perspective.

And so if you look at what you have to do to actually get to these leaps and bounds that are required to make AI commercialized successfully, it actually takes optimizing everything from the factory all the way up through the software. The foundry model today is you make something really, really well for chips, and maybe a little bit of packaging in there, and that has worked great. But if you want to get to the point where you’re keeping up with the scaling of AI, we have to do something dramatically different. And that’s why this system foundry pyramid that we use to express this starts with doing everything well that a foundry is supposed to do, because you can’t be a systems foundry unless you’re first a foundry.

Patrick Moorhead:
Right.

Craig Orr:
And then we layer onto that the advantages that are needed to bring AI everywhere sustainably and scalably. And so the next level on top of that is actually a resilient and sustainable and secure supply chain. And then above that, we differentiate by bringing the systems of chips capabilities that Intel has developed over its 50-year history to bear. And so the combination of those three things really helps you optimize everything from the factory all the way through the software. That’s going to be critical if you want to keep up with the scaling that’s demand of us by AI.

Patrick Moorhead:
That’s great. And by the way, stuff like the compute networking and cooling, totally self-explanatory. Get that. What kind of software are you talking about here?

Craig Orr:
This can be multiple things. There’s the one API software that everyone talks about, this abstraction layer to help things plug in easily, especially on the enterprise customers. But there’s even things that, if you dig a little deeper, that are more differentiated. We have something called the AI NIC that Satya announced in the Vision event.

So what that is an open alternative to ethernet that is basically twice as power efficient at the same speed. But if you want to take that and use that, let’s say you’re a hyperscaler, you need to be able to go and say, “Okay, how is that going to perform in my system? How is it going to be integrated to my software stack?” And we can help with both of those. We have software that’ll help the customers simulate it in their own environment, what performance benefits they’ll get. We have also the capability to help the customers write code and put that code into their software stack. So that’s one example. But there’s other things too, like emulation software to look at, how do you accelerate being able to write your code before your chip comes out? And a whole bunch of other things that we’ll be announcing probably next year at Direct Connect.

Patrick Moorhead:
Yeah. Are you sure you just don’t want to spill it here?

Craig Orr:
Yeah.

Patrick Moorhead:
I’m just kidding. Yeah, we’re not a breaking news show. It’s more about thought leadership and strategy here at the Summit, but it is cool. And by the way, I also appreciate, and it used to be the wafer in the silicon was front and center, and the packaging was kind of an afterthought. Now it’s like you can’t do any of this AI goodness without having great wafer fab capabilities and also the packaging of all this goodness. So literally packaging is now a first order citizen. And things like EMIB and Foveros and the ability to package memory in multiple chips. I know you call them the tiles, but it’s definitely the industry is changing.

Craig Orr:
Yeah. We think that by the 2028-ish timeframe, probably 60% or 70% of the advanced node ships, like less than five nanometer, that ship will actually go into packages that are either 2.5D or 3D, or a combination thereof. Yeah, so it’s going to be pretty much everything at some point. And Intel, a lot of people don’t really realize this, but Intel has a lot of capability to do this. We have three to 4X the capacity of TSMC in advanced packaging, and then we do so many different varieties and configurations of these chips and are vertically-integrated, so we can influence everything from the substrate all the way up through the cooling, like we talked about. And so we’re able to… and we even, actually, have specialized equipment and so forth for assembling. So we’re able to put these really complicated chips together with really good performance, power, area, and especially really high yield.

Patrick Moorhead:
Yes. Yeah, it’s just fun. It’s not just AI, it’s chip revolution. Different types of developers adding different types of IP. It’s fascinating here. So I’m going to dial out a little bit on this and ask you kind of a 50,000-foot question. How does Intel Foundry look at the potential of AI? Sure, it’s going to be big. It’s going to be bigger than anybody even thought it could be, but how do you look at it through your lens?

Craig Orr:
Yeah. I think, in terms of a market size perspective, when we look at the data for, let’s say, the most advanced node at any given time, by the 2028 timeframe we see probably about 80% of all wafers shipped will have a significant component of that related to AI, and probably about 30% or 40% of those wafers will be exclusively dedicated to AI. So from a market perspective, it’s going to be very large. But the challenge that we see is it’s very early in the game. We haven’t even gone through the hockey stick for AI yet.

Patrick Moorhead:
Exactly.

Craig Orr:
But already we’re seeing the amount of computations you have to do to train a model double like every 10 months. The amount of power you use for these things next year alone will be greater than 60% of countries. And so if you put all those systems, that’s just for the chip itself, right?

Patrick Moorhead:
Right.

Craig Orr:
For the server chip. But if you put those into a system, it ends up being like three times the power consumed by the entire United Kingdom. So it’s just a staggering amount. And so if we want to be able to go to that rate where we’re improving this and keeping up with the amount of flops that are needed every year and the amount of memory bandwidth that’s needed every year, it can’t happen just by Moore’s Law. It has to happen… Moore’s Law, it’s great. It’s going to continue to fuel us, but it’s not going to get us 100X or 1,000X improvements over the next five years. So that’s why we’re doing this whole systems approach, because if you optimize all of those things together you can get there.

Patrick Moorhead:
Yeah. In fact, I was on your stage this year at CES 2024, and I was very clear on the beginning of a super-cycle really kicking off the end of this year and going through ’25 and ’26 on the client side. And we’re already seeing the massive investments on the data center and the build-out. I’m really glad you brought up power. The more that I dig into this, literally, Western Virginia is the home of many hyperscalers data centers, and they literally don’t have any more power on the grid to do anything more. That literally says, we put even big EV stations there or more housing, if we don’t do something different, we will not have enough power. And it takes 10 years to get, say five years to get a new power source online.

And that was probably the best and freshest example of this power. And it’s wild to think the impact that you at Intel Foundry can have on this overall power equation. When you look at how many chips that you bring out and the sophistication with leading nodes, and your ability to optimize that. You have the ability to have a huge impact on that power draw holistically.

Craig Orr:
Yeah, definitely. And the power is 100% true, but I think it’s even bigger than that, too, in the sense that we also have to not just address the power that comes from running the chips but the power that’s used to create the ships. So if you look at where they’re all created today, they’re all created in countries that, not only is it dangerously centralized in terms of supply chain but, the countries where they’re made, they don’t have access to renewable electricity. And so as a planet for our countries, for our economies, for our companies, for the planet, we have to get much better at doing this.

Patrick Moorhead:
Yeah. And I know we could rat-hole and talk about the decentralization, but it’s real. And Intel, at least from my vantage point, is a huge part of fractalizing this investment in foundries. So let’s get down to the tip of the spear here on differentiation. You talked about your strategy, but I’d love to talk about differentiation. Where are your points of differentiation with your competitors? And I would say, on leading edge, you’ve got two competitors. And then on the middle of the stack that you’ll be addressing yourself over time or through partners, there’s a lot more competitors in there. So give me the elevator pitch here.

Craig Orr:
Sure. Yeah. I think there’s two ways that I look at it, at least. One is in terms of, how differentiated is each aspect of the traditional foundry offering? And then how is it put together in a way that makes it more differentiated?

Patrick Moorhead:
Yeah.

Craig Orr:
And so if we start with that system foundry triangle concept, pyramid concept that we have, the bottom of it is that world-class foundry offering. And so you look at the process technologies. I think we showed in our investor day, Intel 3 is coming up to be competitive. 18A, we think we get back to leadership. 14A, we think we’re in leadership position. That’s where we introduce high-NA. 18A is where we introduced with the industry-first backside power and gave all-around together.

So I think, on the technology front, we’re getting back to process leadership. On packaging, we talked about how much more capacity, that 3X to 4X capacity we have. We also do so many more different configurations of packaging, and are able to do so at the really high yields with really large sizes and really good cost. So I think at that level we’re differentiated, but where I think things really start to shine is when you move up to the next levels of the triangle. If you look at a resilient and sustainable supply chain, who has that? No one.

Patrick Moorhead:
Yeah.

Craig Orr:
And so we’re going to be doubling the amount of capacity that’s available in the world, outside particular regions. And then we’re also doing that very sustainably. We have 99% of the power we used last year was from renewable sources already. So that’s a differentiator. And then once you go up even further to actually thinking about the systems of chips’ capabilities like being able to know how to design something, how to help our customers architect something. How does the compute and the memory and the network need to be in balance and mapped well to the software? Who else can help you do that? Other companies, the model is more like, “Okay, here’s your PDK. If you follow the rules I put in here, it’ll yield.” So I think that’s where the differentiation comes in. Sure, we have the differentiation at the process level, at the packaging level.

Patrick Moorhead:
Right.

Craig Orr:
When we start to put everything in together from the factory network up through the systems of chips capabilities, that’s where it gets really special.

Patrick Moorhead:
Now, I appreciate you pointing that out. And because it’s somewhat of a newer concept, it’s important to be talking about that over and over and over, kind of like we are on this video right now. So that’s the good part about it, is I have always thought that balancing the supply chain was critical. In fact, I may or may not have been an advisor to the previous administration on what became the CHIPS Act. Not violating any NDAs, DOD, but I was involved in that, and I thought it was crystal clear for me, too. And plus, I spent a lot of time traipsing around China and Taiwan and Asia, so I’m very familiar with the region and the pluses and the minuses here.

By the way, certainly a lot of those geopolitical elements we saw, the US government gets serious about it with the CHIPS Act. But I want to ask you about the future here. What does the future look like? How do you define success? And how should we measure your degree of success over the upcoming years?

Craig Orr:
I think, for us, ultimately, the future, we’ll be judged by our execution. So do we deliver on the technology milestones that we set out? Do we deliver on the performance, the power, and area that we promised our customers we would deliver? And then as a result of that, we’re going to see customers more and more coming on board. Foundry is one of those things where essentially, as you know, you’re going to a customer and say, “Hey, bet your company on us.”

Patrick Moorhead:
Right.

Craig Orr:
If we are not in this together and if you can’t trust us, then essentially their business will fail. And so we have to continue to earn that trust over time by saying what we’re going to do and then actually doing it. And so we’re really focused on that. And so far we think it’s going well. We have a pipeline of business awards that are about $15 billion, and we believe we have a solid, credible path to get to about $15 billion of annual run rate by 2030, which, that’s from external customers alone. And you combine that with the internal, and I think we’ll be squarely in the number two foundry spot that we set out to be.

Patrick Moorhead:
That is so audacious, Craig, and so exciting. It’s funny. It’s easy for me, being on the pundit side, to watch what was going on. But if you’ve read anything that I’ve talked about related to Intel, I saw the opportunity there. It was crystal clear. And I think I was in the industry for over 20 years and it was funny. I was actually Intel’s biggest customer once. And okay, I was one of your biggest competitors, too, but you get to know the industry and you see the signs. And what I’m seeing is very positive.

The thought that 18A could be not only the most cost competitive, but feature and performance competitive, I saw that on your last Direct Connect slide, is huge. And then moving into 14A is pretty wild and not that far off. Some people think, oh, you can crank out a new bank of software like a month, but to build a foundry, get process, it takes like four to five years of planning to get something like that going.

So literally, the decisions that you all made years ago are just coming into fruition now. And the signs of success that I’m seeing, obviously everybody wants to see more. They want to know who it is. Which customer? How much? By the way, ironically, your competitors don’t have to share any of that, but they want that from you. But I see the signs on the wall. It was funny. 10 years ago people were telling me, “Semiconductors, they’re a commodity, and anybody can make them and anybody can do this.”

And then the pandemic hit and we’re like, “Oh, wow. Diversified supply chain is super important.” And it wasn’t just the bleeding edge tech that mattered, it was always the specialty and the lagging edge. And then generative AI hits and it just amplifies this and literally is… semiconductors are the currency of building out AI and all of this stuff that software runs on. So I’m super excited, Craig. You seem super excited about this, and I look forward to anything, your next divulgement of information. So I will be listening.

Craig Orr:
Yeah. I’m so excited about it, too. And I think customers are getting excited about it. The ecosystem’s getting excited about it, and government’s getting excited about it. They’re all behind us. But you mentioned what happened during the pandemic, and that was for $0.30 chips disrupting our whole supply chain. Imagine now that AI’s going to be in everything we do. What happens when all those chips, who currently come from one vendor and one foundry, are not available? And so I think people are going to start caring a lot more about this. And Intel, as we demonstrate that we are a real foundry, that we’ve been executing, the conversations have been shifting from, “Can you execute?” To, “Okay, how can we do this? Or how is it going to work? And what are the logistics of it?” So the progress is there and it’s being recognized by the whole ecosystem, so we’re really happy to see it.

Patrick Moorhead:
Yeah, it’s fun to watch because it’s so big, it’s so bold, it’s so audacious, and so important, all wrapped up into one. So Craig, thanks for coming in here and speaking to the Six Five Summit 2024.

Craig Orr:
Hey, Pat, thank you for having us. We’re super excited.

Patrick Moorhead:
Yeah. Excellent. So this is Craig and Pat. We are talking about foundry services and the importance that it brings to the whole AI ecosystem and the whole AI story. I love chips. Hopefully you can tell that. Stick around for some more semiconductor action here in 2024. We have semiconductor manufacturers. We’re talking memory accelerator chips, CPU, GPUs, everything under the sun. Take care, and thanks for tuning in.