Intel Corporation (INTC) Earnings Call Transcript & Summary

Good morning. Thank you for joining us. I'm Tim Arcuri. I'm the semiconductor and semi equipment analyst here at UBS. And we're very happy to have Intel with us. And with Intel, we have Sandra Rivera, she's the EVP and GM of the Data Center and AI Group. So before I turn it over to Sandra, I'm going to read the safe harbor. Today's discussion includes forward-looking statements, which are subject to risks and uncertainties. Please refer to Intel's SEC filings available at intc.com for more information on the risk factors that could cause actual results to differ materially. So thank you, Sandra. And I'm going to turn it over to you for some prepared remarks.

And thank you, Tim, and thank you all for joining this morning. So as Tim shared, my name is Sandra Rivera, and I lead the data center and AI organization at Intel. I'm in the role now since July, so just a few months, but I'm no stranger to the data center organization. I have been part of the team that built and grew the networking business for Intel over the last decade into what is now a $5 billion or $6 billion very successful business for Intel with a lot of growth opportunity. So I wanted to start out just by setting a little bit of context in terms of the tech industry that we find ourselves in today. And clearly, we are in a time of great transformation and opportunity. And we see that the pandemic, of course, has accelerated all of the trends that we had already been seeing in terms of the move to cloud architectures and cloud business models, the pervasiveness of computing and ubiquitous computing everywhere, the need for connectivity to move and store and process all of that data. And of course, the continued growth of AI, which is a workload that impacts every single customer segment, every product, every deployment going forward and still very much in its early ages -- early stages. So when we look at the continued growth of data, the exponential growth of data, we know that we have to have computing to process all the data and make it something of value. And so the demand for computing has never been greater, that's accelerating in time. And it's against that backdrop that with Pat coming back to Intel, we have a lot of transformation happening at Intel as well. New CEO, new strategy and recommitment to being integrated device manufacturers, something we call IDM 2.0. The new structure that Pat has set up with fixed business groups to lend more focus and get back to greater disciplined execution around those 6 businesses. And of course, a new strategy, which is anchored on, as we call it, the 4 superpowers of AI: pervasive, connectivity, ubiquitous computing and cloud to edge infrastructure. So the fact that the data has to flow between what is typically a highly aggregated infrastructure to a disaggregated infrastructure and we're removing the compute to that point of data creation and data consumption. So our strategy has been anchored on leadership products, which is what customers value. Regaining process leadership, which you've heard us talk about the investments that we're making to get back to leadership on the process side and the fact that all of the process nodes that we're driving over the next 4 years are on track or ahead of the milestones that we've established. Intel 7, Intel 4, Intel 3, Intel 28 and Intel 18 are all tracking from a process technology perspective, which, of course, as a business unit general manager and a P&L owner allows me to take advantage of bringing process technology internally. But also what we talked about is with the IDM 2.0 strategy, we have a more modular architecture in terms of our -- the design of our products. And we're able to take advantage of both external foundry capacity and capabilities, as well as our own foundry capabilities and design the leadership products that our customers expect from us. And that level of flexibility and agility does give us a lot of opportunity for innovation as well as building products in a more agile manner moving forward. So IDM 2.0 is internal capacity that we're investing in leading process technology that we're anchored on, but also the external capacity and the increasing use of external foundries. And then, of course, the third element of that is standing up our own foundry capability, which gives us an additional route to market with customers that want to introduce their own IP into products that we manufacture. The second point that I want to make is just around the embrace of open platforms and architectures. That is a differentiated capability that Intel has, that is our DNA. We invest in ecosystems, we invest in industry-shaping standards. We lead many of those standards on behalf of the industry, and we believe that lowering barriers to entry, increasing market participation, just accelerates the rate of innovation. So open is very much the differentiated approach that we have to the market. And then lastly, the investments that we're making for manufacturing at scale. And you've heard Pat and [ George ] [indiscernible] the significant investments that we're making for more capacity as we see the continued growth of data, the continued demand for computing and the opportunity that we have to satisfy a lot of that requirement for our customers. So let me just pause there, Tim. That's really what I want to share with you.

Thank you. Thank you, Sandra. That's great. Thank you very much. I guess I kind of wanted to start just with the new org and your new role. Last time we spoke, prior to HR, you used to run the networking group under Bob, and then with the realignment into, as you said, 6 business units under Pat, you're now in cloud and enterprise, the Xeon business, you're an FPGA and you own the AI ASICs by Habana. So really, I have 2 questions. First of all, can you talk about the realignment of the DCG segment? And second of all, maybe the different requirements from a management perspective for this segment versus networking, which you were running in the past?

Yes. Yes. So as I shared, I am not a stranger to the data center organization because I was part of the team. Of course, what is a data center, it's compute network and storage, and so I had the network being remit for many years. And that was really the business that I helped to grow for Intel for a number of years before Bob, our previous CEO, asked me to make the hard pivot to be the Chief People Officer of Intel and the Head of HR for what was 2 years, largely during the pandemic. So that was a humbling experience, I would say, an extraordinary learning opportunity, but just a huge amount of responsibility to take care of our people during the time of such uncertainty and anxiety. Then when Pat came in, he wanted to structure the organization in a more focused set of business groups. And the ask he made of me is to lead the data center organization, along with leading the AI strategy for Intel. So the data center organization, the data center remit is the core of the franchise, of course, the entire platform. It is the accelerators that we have for AI, including the Habana asset that we acquired 2 years ago, includes FPGAs as just a wonderful platform for innovation and particularly for algorithms that are continuously changing over time, whether they're security algorithms or networking algorithms and acceleration as well as AI acceleration reducing the AI, of course, with FPGAs as well. And then I am also responsible for the Optane memory organization as well as interferon. So all of that is highly anchored on the customer base that we have around cloud and enterprise. And for each of us, as you hear, we're actually today very focusing on our own met. But of course, when we look at the need for pervasive connectivity, I'm looking to the Network and Edge organization led by Nick McKeown to partner with me on that. When I look at a lot of the opportunities in graphics and graphics acceleration, of course, that is what Raja Koduri leads as part of his organization. And we partner very closely to ensure that we have the complete platform to offer customers leading solutions. But the thing that ties that all together really is the software, and that is a new organization that has up level to be the direct CEO report, bringing in a deep domain expert and professional, Greg Lavender, as our CTO, but also the Head of Software. And in that regard, Greg and I have partnered, while I am convening the Intel assets around the AI strategy, several of which I had in my portfolio, of course, but others in other parts of the portfolio, Greg and I are partnered on creating that single software stack for AI. Because at the end of the day, AI is largely a software problem and the need to onboard developers and to accelerate that time to value is really anchored in the software. So that's how we're structured. Our focus in my organization is around cloud and enterprise. And cloud is, of course, a set of customers that we're all quite familiar with, the Tier 1 hyperscale cloud service providers. But also, a set of strategic customers way beyond that top 7 or 8. Because cloud is really an architecture and a business model, and every single customer, whether you're an enterprise, whether you're a cloud service provider or whether you're a comp service provider, you're embracing the idea of shared resources, cooling, software extraction and these recurring revenue subscription-based business models. So cloud is just pervasive in every single customer segment. But we are primarily responsible for the cloud service providers and the enterprise where the workloads are quite diverse and quite broad. And of course, our portfolio of products are designed to address that very broad segment of customer design points that vary in terms of power, performance and price.

Great. I wanted to move on and ask you a couple of questions about the road map. So the first question is with Ice Lake ramping and Sapphire Rapids on horizon, certainly, there is some pent-up demand as Ice Lake has been late to the market, obviously. And of course, we've seen very strong CapEx commentary from some of your largest cloud customers. So my question is, can you speak to some of the market segments and the workloads that are going to drive the business into next year? And maybe what's your general outlook for next year?

Yes. So Ice Lake is deep in its ramp. Right now, we will ship here in the fourth quarter as much product as we've shipped since its launch at the beginning of the second quarter of this year. So that ramp continues to accelerate with every single OEM, ODM and enterprise and cloud service provider launching Ice Lake instances or Ice Lake platform. So the ramp, while it was protracted, is actually going quite well at this point. But of course, we are on the heels of Sapphire Rapids, which will begin shipping in Q1 and ramp in Q2. And actually, it will continue to ramp throughout the year next year. And Sapphire Rapids brings with it many new innovations. We have DDR5 memory. We have PCIe Gen5 coming with Sapphire. It is designed for micro services, cloud native architectures so that we're offloading a lot of the infrastructure overhead within the CPU, allowing the course to be more efficiently used for the revenue, compute processing that our customers do in the cloud. And there's just a number of new accelerators that we're introducing with that platform. We have accelerators for crypto and compression, where it's critical in terms of data movement and data security. We have instructions for data streaming and dynamic load balancing and a lot of the ways that we're trying to make the infrastructure much more efficient and effective for our customers. And the one area that we have clear leadership that we introduced nicely, but are taking it to the next level at Sapphire is in AI. Through our AMX or advanced matrix extensions, we have a number of accelerator engines built into the CPU under the ISO contract, like the ISO instruction set. So it makes it very easy for customers to actually run AI workloads on the CPU they already have, and in a programming environment that is very familiar and frankly, ubiquitous and easy to use. And this is the way that we see the market playing out where, Tim, when you look at what is driving most of the growth of new compute cycles, it isn't that AI domain. That's true, of course, in the data center, but it's also true out at the edge of the network, where you will do typically model training in a centralized data center, but a lot of the inference both in the data center, of course, out of the edge happens on Xeon CPUs. So with Sapphire, we're introducing many new innovations, but this area around AI is one where we have clear leadership and where we see our customers wanting to do some level of AI in almost all of their workloads. And then one other big innovation that comes at Sapphire is the CXL XpressLINK interface, which allows you to attach all types of accelerators to the CPU platform. It's an industry standard that we led with. We see our customers and competitors and ecosystem adopting it. And we think that this also accelerates the rate of innovation and the opportunity to put more products and more chips and chiplet together to create a more comprehensive and compelling overall platform. So lots of growth happening in cloud architecture, lots of growth happening in AI and lots of growth also in security spaces where encryption and compression and having both software-based and hardware based on place for securing data and separating data, tenant data, from infrastructure -- from the infrastructure platform is important.

Got it. Since you brought up Sapphire, maybe I can ask a question about competition. Obviously, Sapphire is a big, big step up from Ice Lake. There's no question about that. Obviously, from what we've seen, it's very competitive with what your competitor has in the marketplace today. Of course, they will have a next-generation product next year. But of course, you have Granite coming in 2023. So can you just sort of zoom out and maybe characterize the competitive landscape?

Yes. So when you are in the business of high-performance computing and you have lots of segments and customers and design points that you're servicing and the depth of the relationships that we built over many decades, we have competition everywhere, Tim. And we take competition very seriously at Intel because, of course, competition is what, frankly, often makes you better and sharper and more focused. So we have competitors at every end of the spectrum. Our focus is on building leadership products because ultimately, that's what customers value. We know that we have leadership with Sapphire. We know that there is a high level of differentiation in terms of the feature set. And perhaps, as importantly, what customers value is the workload that they run, and the opportunity for us to work with customers in workload -- in optimizing their workloads to achieve the very best performance is something that requires a lot of software capability and deep engagements with those customers where we're tuning and optimizing both the hardware and the software together. And when we look at, for example, with Sapphire, with the AMX instructions, we know that between the accelerator hardware, the process technology entitlement and the software optimizations, we're getting 30x the level of AI workload performance versus previous generations. And we're many more times superior to anything that the competition has. So actually, they don't have any built-in AI acceleration in their CPUs. And this is what the customers value. And I will say that I've spent the last 5 months now, very deeply engaged with customers. I meet and talk with customers all the time. And they are very much counting on Intel. They appreciate the years of platform optimization work that we've done with them, the software resources that we bring to their organization, into their workloads and applications, the service and the support and the fact that they can just count on Intel. And we've had our stumbles, we've had our setbacks, but in every discussion and every dialogue, they're counting on Intel. Because if you look at most of the world's high-performance computing deployed in data centers is deployed on Intel CPUs. And there's a lot involved in terms of managing that fleet out in time where they're counting on Intel to service and support the fleet that they have, but also to continue to upgrade, enhance, improve and accelerate the capabilities. So competition is formidable. It's the reality of the world that we live in. It's a good thing. It makes us sharper. It makes us more focused, makes us frankly hungrier. And I feel very optimistic in terms of the leadership roadmap and our process technology leadership that we're going to demonstrate in the coming years.

Great. And then just as a quick follow-up to that. When you sort of think about just the Architecture Day. One of the key highlights I thought was the bifurcation of the x86 roadmap with single-sided and multi-sided workloads relying on different architectures, really a hybrid approach, of course, with the P course and the E course. Can you talk about that? How does that help your roadmap? And how does that help you better compete as we were talking about that? Is that indicative of a sort of a bigger shift in terms of how you think about the product roadmap? Can you sort of frame that against what the competition is doing?

Yes. Tim, it's an excellent question because, as I described, our customer base and the breadth of workloads and design points that they have is very extensive. And it really requires us to be more heterogeneous in terms of the platform and the core CPU offerings. And so when we talk about P cores, or performance cores, or E cores, efficient cores, we actually have demonstrated our ability to make some match cores on a single product, which is the Alder Lake product, the client product, that we've announced recently. We are going to have those same assets to bring to bear on the server roadmap side. I'm not making any new announcements here today, but we will talk much more about how we're leveraging those innovations and how we are able to more specifically target the power price performance bands that customers have, which again, are varying and wide. The -- of course, the powerful strategy behind that is that it's the same ISO contract. It's the same software tool team. It's all of the benefits that customers expect from designing Intel platforms which is very much drop in and accelerate and use many of the improvements and enhancements in the overall architecture. And that's one of the things that we're seeing the huge excitement around Alder Lake. And then you'll see that in the server roadmap in the future as well.

And does that allow you to customize more? Obviously, one of the big trends from the cloud and from the network and from everybody really even in the auto segment, customers want customized solutions. And so does this help you provide more customized solutions? And maybe can you give an idea of really what portion of the mix is custom versus standard?

Yes. So customers increasingly want a level of customization around their platforms. We've done that with the Xeon roadmap for many years, mostly anchored on different frequency bands, different core counts, different power envelopes. And so we've been offering that for a long time, and well over 1/3 of our SKUs are custom SKUs for customers that want that level of customization in terms of the standard roadmap. The fact that we have a more modular architecture going forward and we have different cores to choose from, and we have a chiplet or tile-based approach that we're utilizing to instruct our own SoCs, our own products, is also a benefit that we can offer to customers going back to the IDM 2.0 strategy. So IDM 2.0 does allow us now to not only leverage our internal process technology, but also process technology from external foundries. And then the Intel foundry service allows us to build highly customized products for customers that may want to build solutions or may want -- and this is really where so much of the interest is coming for IFS is, given that we are going to put our own innovations, our own IP in the foundry, customers that are building their own products, Tim, have this wonderful opportunity to utilize and introduce their own differentiated IP with our IP and then have us manufacture it within our foundry service, leveraging, of course, the process technology that we have internally. So between the modular architecture or more chiplet and tile-based architecture and the IDM 2.0 strategy that allows us to leverage the foundry for building custom products through our advanced packaging technology, we feel that we have many more routes to market and a much richer value proposition for customers that have differentiated IP that want to mix and match with our own differentiated IP.

Great. Let's double click on IDM 2.0. So obviously, you had talked about -- one of the premise is to say you can -- you would rather use internal, but there's the flexibility to use external foundry as well. So can you just talk a little bit and zoom out about the benefits and the opportunities that, that creates for you as the GM of a product-focused business? I mean I'm sort of curious about the logistics around that. When do you have to decide whether you're going to go internal or external? Does that change the roadmap 2 years out? Do you have literally dual-track teams designing internal, external? Can you just walk through some of those things?

Yes. So maybe I'll just zoom out for a moment and just remind everyone that we have used external foundries for many years. We have an excellent relationship with global foundries, with TSMC, with Samsung. They've been building any number of our products or different chiplets or tiles for many years. And today, it's probably as much as 20% of our overall volume. So we do have that capability. In terms of as a business leader and being responsible for the P&L and the roadmap, we are always looking years ahead in terms of what the process technology is that we want to use in our products. Some of the tiles or chiplets benefit from being on leading-edge process technology, and we will always buy, source that internal leading-edge process capability. And particularly encouraging to see that, as I said, we're on track for Intel 7, 4, 3, 28 and 18a. And so that would be my normal bias because we just have a better margin structure that way. But there are tiles and chiplets are part of the solution that may not meet the latest process technology and where we can benefit from capacity from external foundry. We make those decisions following that tick-tock model that we've had historically where we will want to be consistent, perhaps with some of those tiles, we don't need to change them or redesign them to a new process technology because that's not where the big performance increase is going to come from. And so we look out in time. We are designing our products 2 years ahead, and in some cases, longer in terms of the IPs. But it doesn't mean that we're double investing, it just means that we take a look at the overall architecture and the construct of the product that we want to build and then make decisions on where [indiscernible] the different IPs, the different tiles in terms of the foundry process. So it does take, of course, foresight and planning because silicon development are long and expensive campaigns. And so you are looking out in time. But it's not a dual-drive strategy, meaning that we don't have parallel teams working on the same IP on different process technologies.

So just to clarify, so for something like Granite, so there are not multiple teams, one Granite team working on internal and one Granite team working on external. That's not the case.

No. But what you would have is you would have a team working on the compute tile that would be running on -- typically on an internal process and then maybe an I/O tile or a memory tile or another IP tile that would be targeted for an external value process.

Okay. Great. Let's move on to some of the segments and some workload questions. So you own the AI strategy, so Intel is known for having the broadest portfolio of silicon assets and AI. Can you speak to some of the traction that Habana in particular is seeing in the marketplace?

Yes. So we have a very broad AI portfolio. As I shared earlier, with Ice Lake, we had our DL Boost, VNNI instructions and hardware accelerators with Sapphire Rapids. We take that to a whole another level with the advanced matrix extensions, the AMX accelerators. We have FPGAs. We have GPGPUs. We have computer vision accelerators. And of course, we have the Habana AI accelerator assets. And in some ways, the choice of the underlying hardware is something that we are rationalizing is one of the areas that I'm diving deep into with Greg to understand how broad the portfolio needs to be in terms of the hardware architectures. Do we have an opportunity to focus in on fewer and then, of course, build the complete software stack on top of that. And that's the work that's underway right now. And it's actually progressing very well because everyone is excited about just the pervasiveness of AI workloads across all segments, customers and products. When we look at the overall AI market, there is classic machine learning, there's training, there's inference and there's, again, even within that, a very broad range of design points. You hear mostly about deep learning training and, of course, the hundreds of billions of parameters, models that need very dedicated computing acceleration to actually train and it takes a bespoke effort and architecture and products. But many of the models are not $500 billion parameter models, they're much, much simpler, much, much lighter weight. And you can do a lot of model training in the CPU that you already have. So if you look at the workflow for AI, it's data prep, and that typically is done on a CPU, on Xeon CPU. If the model training, and depending upon the size of that model and whether or not you need peak throughput and roofline performance 100% of the time, you'll choose the right accelerator or a CPU for that. And then it's the inference or the deployment of that model. And then, of course, there, again, Xeon has a big footprint. And then there's the retraining, right, that happens in that overall loop. When you do need peak throughput or high levels of acceleration that you need consistently over a period of time that it isn't something that you're doing a batch offload process overnight for like fraud detection and some of those types of applications, then that is where you're served by a bespoke accelerator. And that is where Habana is really targeted. And the success that we're seeing with Habana, of course, as being the world's largest public cloud service provider, AWS, launching the DL1 instance last month -- or actually 2 months ago, where they are offering the DL1 Habana based instance at 40% better TCO than the alternative solution, than the market leader they have deployed in their infrastructure. And we're getting a lot of traction from customers that are interested not only in, of course, the TCO benefit of having this as an alternative, but also the fact that, again, Intel is much more about open platforms and accessibility and lowering barriers to entry. And much of the success that we're seeing with the Habana accelerators, the Gaudi product is a result of us upstreaming into all of the major standard frameworks, the TensorFlow and PyTorch. And then moving forward, we'll be expanding that. So there is an appetite for alternatives to what has been largely deployed for that dedicated accelerator. And then complementing, of course, the Gaudi Habana accelerators is the Xeon portfolio that can do both the training and the inference for many, many of the workloads and the models that we already have today.

Yes. I mean obviously, you have a very strong position in inference. And I'm kind of wondering whether in training, you have an estimate of all the training workloads that are being done either on Xeon or on Xeon with your own accelerators. So is there some share of the workload where you could give us and say, well, we're still doing 95% of all training workloads are either on Xeon or they're on Xeon plus Gaudi or for some Intel accelerator. Is there some number [indiscernible]?

Well, I would say the majority of deep learning training in the data center is happening on alternative architectures. It is the opportunity that we have for headroom and growth, both with Xeon and of course, with Habana. And actually, in other ways, particularly if you look at HPC workloads and AI as part of the HPC workloads with our Ponte Vecchio solution that Raja is responsible for as part of the accelerated graphics organization. So for us, we have lots of opportunities for growth in the training because we are not the majority player today, but there's a lot of momentum behind having a choice, and particularly one that's anchored on open software programming models and leveraging the ISO OneAPI software environment. On the inference side, we have the majority position there today. We've estimated somewhere in that 65% MSS range that we enjoy today. And again, inference, both as well as out on the edge, and we believe that from all the estimates that we've seen in all the industry analysts is that the growth of inference, the volume of inference is going to outpace the training volume in the data center. How much that translates to in terms of revenue is not necessarily -- it may still be equal revenue buckets because inference is just a lower ASP. But the volume of opportunity for growth is the largest there over time. And we think that we're going to see much, much more innovation, again, the more we lower barriers to entry and make those products accessible to the market, both in terms of domain experts and data scientists or just subject matter experts in different industries, that we're bringing into the AI workloads [indiscernible] and off the shelf in terms of just the frameworks and the models that they're accustomed to programming.

Yes. I guess my question on training was really around what the catalyst is going to be to migrate those GPU workloads back to either Intel and maybe it's your discrete GPU, I mean, that will be the answer that's not part of the group. But sort of what's the catalyst going to be to migrate those training workloads back to Intel?

Well, I think the -- certainly, the ease of programming and having a software stack that's accessible, that's open, that isn't designed for any single architecture and a lot of the work that we're doing in OneAPI is to build out what is an open standard language, the sickle language that supports multiple architectures, including, of course, GPUs and gives customers more [indiscernible] about where they land their workloads and helps them achieve a lower TCO as a result. And competition is a good thing going back to some of the earlier comments. So Tim, I expect that there will be some migration in terms of ease of use. But if I really look out at where we are in terms of the maturity of AI, it's very early days, lots of growth opportunity. As I mentioned, we see huge volume and innovation happening at the edge. We see huge growth still, of course, in centralized data centers. And you talk to all of the cloud service providers and any enterprise customer how much AI they're incorporating into their own infrastructure, into their own workflows. We at Intel, particularly when we look at our TD, our technology development and process technology, lots of introduction of AI/machine learning, just to accelerate the capabilities that we have there and help us make better, smarter decisions faster. So I would say that in order for us to grow and to grow faster than market, it isn't necessarily moving workloads. It is very much -- there will be some of that, but it's very much just a growth opportunity where innovation. Again, we're at the early days of innovation and lots of opportunities that we can make available to more customers in an open standard language that allows them to get into the market at a very low bar in terms of the cost point.

Yes, I think that most investors are obviously very -- in fact, I'm acutely aware of your competitors roadmap versus your roadmap. But interestingly enough, I get asked many if not more questions about the silicon efforts from your customers, as I view about the relative roadmap of your competition versus you. So can you talk about the silicon efforts at your customers? How much of a headwind is that? How much does it keep you up at night? And obviously, in some ways, they want customized solutions, which, of course, now through IFS, you can offer those and you can -- and you're licensing x86 through IFS as well. So just broadly talk about do you see it as opportunity? Or do you see it as a headwind to you?

I mean, most definitely, it's an opportunity. And I think the days where any one of us can project out in the future and know exactly how things are going to pan out and call the ball in terms of what are the right -- what is the right architecture for any particular market or product or segment. We know that there is no one size fits all. And so us being able to engage with our very smart, very innovative customers allows us to co-architect and engineering -- co-engineer products for the future. And so the fact that several of the largest customers are building their own products, for us, gives us an opportunity to, again, co-architect, co-engineer, co-innovate. We've done that on merchant silicon products. We just announced the IPU, the infrastructure processor unit at Intel On with Google as the definitional customer. But that is a merchant silicon product that will be available to all customers, and there's a rich roadmap of follow-on products to that initial Mount Evans product introduction. So we have opportunities to partner with our large, very capable, very innovative customers in that way. And in the case where customers want to build their own products, they are still typically wanting some level of IP from industry or from Intel. The fact that we're putting our core differentiated IP in the foundry is now a new capability that they didn't have before. They're actually quite excited about that. And as you know, we announced, even in the spring when we launched the Intel foundry service with AWS as one the key customers that is coming into the foundry to take advantage of the packaging IP and packaging technology that we have. So Tim, I would say that for us, having additional routes to market, whether that's co-engineering, codesigning our merchant silicon, mainstream roadmap and then offering that to other customers, but the definitional customers have, the time-to-market advantage of working on that with us from the onset or having a different route to market with that combination of their IP, our IP and us to be able to manufacture products for them, just opens up another route to market, another opportunity for us to grow with them. My goal always with all our customers, is to just say yes. Say yes to capacity, say yes to innovation, say yes to our products and architectures and capabilities, and that is something that our customers value, our service, our support, our software and our ability to deliver for them, particularly over the next several years where capacity is so constrained and where they're looking for a partner they can count on, they can count on Intel.

Just as a last question. So obviously, next year is shaping up to be a good year in terms of enterprise IT spend and we all see what the cloud companies are all announcing for CapEx next year. That's going to be great as well. So my question is, how should we think about the cloud versus enterprise mix next year relative to this year? It seems like cloud is going to keep on growing very, very fast. But it seems like enterprise could become a bigger relative piece of the mix next year than it was this year.

Yes. So I'll answer that in 2 ways. First of all, we do see the cloud service price accelerating. I think they've been very forthcoming in terms of their CapEx spend and they're continuing to build out, and we will participate in that, of course. But we also know that there's a lot of enterprise pent-up demand, and frankly, money that's been sitting on the sidelines that will come into the market. So we expect cloud to grow next year, those customers to continue to grow, not just the top ones, but of course all of those strategic customers, all those cloud native companies that are coming -- that have come to the market in recent years or coming to the market, and then the amount of enterprise IT spend to also be healthy. But I do want to [indiscernible] this notion that cloud is an architecture that every single customer is embracing. And that's why you hear about multi-cloud solutions and hybrid cloud solutions and the build-out of the enterprise edge and how we facilitate that processing of the data at that point of data creation, data consumption at the edge of network. And so cloud as a deployment model is going to continue to be pervasive for all of the different customers. And this is an area where we are very focused in our own portfolio to ensure that we have the most efficient architecture that offloads a lot of the infrastructure overhead, whether it's network acceleration, storage acceleration and patient acceleration and then frees up the cores to do the high-performance computing that's required for the actual data sets that are coming into the platform.

I was on mute, sorry. We've run out of time. But Sandra, thank you for your time. We really appreciate it. Thank you so much.

Tim, thank you for having me.