Hello and welcome to this very special edition of Voices from the Valley. I'm Shireen Bhan. Well, we've been talking about AI and we can't talk about AI without talking about what's happening at AMD. My guest today has over four decades of engineering experience. He's been at AMD since 2011. Joining us now to talk about what's new and what's next at AMD is the CTO Mark Papermaster. Mark, it's an absolute pleasure to have you here on Voices from the Valley. You know, 2024 was a great year as far as AMD was concerned. You doubled your data center revenues, closed the year at almost 26 billion in revenue. But let's start by talking about what's new and what's next. You've had a big event in the last few weeks here in San Jose in the Bay Area. What are you most excited about?
Well, first Shireen, it's great to see you again and thanks for having me on with you today. It is absolutely an incredibly exciting time at AMD. We had a big event last week where we announced our next AI data center GPU, the MI 350 family. It's following the success we had as a new entrant in December 2023 when we announced our first foray to take on the market leader who's got dominant share of the GPU for AI computing. As you said, we had a banner year last year. Of that revenue you stated, 5 billion was data center GPU from virtually zero the previous year. So we went from zero to about 5% share. It's a great start and the event last week showed the world that we had committed to annual updates. We rolled out the MI300, did an update called the MI325, and here we are in mid-2025 rolling out our MI350. Not only that, we shared our future plans for our next two systems coming up in 2026 and shared that we're on track for our 2027 system. So we are here to compete. We're gaining share and creating an alternative. It's a very exciting time in AI as well as the rest of our portfolio which is growing quite nicely.
Mark, I want to talk about the partners that you've been able to onboard, whether it's Meta, Microsoft, OpenAI — in fact, Sam Altman was there at your event in early June as well. So take me through how crucial this era of partnerships is going to be and what it eventually means as far as the customer is concerned.
When you look at what we did as we grew that revenue from 0 to 5 billion US over the course of 2024, it was done through partnership with those hyperscalers. These customers are at the absolute cutting edge of AI. They understand the workloads. Many of them are creating their own large language models that have amassed vast information and it's evolved so quickly. So, Shireen, by deeply partnering with them we really understood their needs. We were testing our MI300 product with them as it ramped up. They said this is great but here's what we need one year, two years, three years, four years from now. Please build that in your roadmap. So those partnerships — and why we focus first on those largest of large customers — is that they informed us as to what we need to be where AI is going in the future. Now as we are in 2025, we're very focused on expanding to support the broad community. 2024 was all about ramping those largest of large customers and now we're attending to the community, getting regular software updates out, making it easier to run AI on AMD — not just on the big GPUs in the data center, but right down into your laptops or your Radeon gaming card.
But Mark, as you speak to the likes of Sam Altman and look at what's happening at companies like Microsoft and OpenAI, what are you picking up in terms of just how wide and large this opportunity is and where we find ourselves on the map today? Everyone's talking about potential. Everyone's talking about riding this wave. We've seen the kind of exponential numbers that OpenAI has been able to deliver in barely two years. So what are we talking about? What are you bracing yourself for in terms of growth?
Yeah, Shireen, it's a great question. When Lisa Su, our CEO, talked about there being hundreds of billions of dollars of addressable market, TAM, in 2028, people thought, 'Oh my god, that's just such a large number. How can this be?' But it truly is an insatiable demand for compute capacity. The reason, Shireen, is that at first all that computation was needed to train those big large language models. If you're running ChatGPT, Gemini, or any of these large language models, you probably already have it starting to be an indispensable piece of your life. When you have a question, even a really tough question, you can go to these generative AI applications and get very thoughtful answers. It's starting to really help people be productive. So what this is shifting from — all the spend on training those largest models — it's shifting over to running those models, getting those answers to you in ever more thoughtful ways as the AI algorithms evolve. That's why we see such exponential growth for this market going forward.
Mark, that brings me to the question of cost. What are we talking about with this exponential growth? One of the key constraints of course is the high cost. With the changes that we're seeing, with the advances that we've already made, what can we realistically expect in terms of taking cost out?
Yeah, cost and power is a huge issue because for those largest models it's very expensive. You have to think about what's the cost to be for the biggest companies to offer that general purpose generative AI application like ChatGPT. I think we'll end up having a handful of those — the hyperscalers we have today, and sovereign nations will establish some of their own because they're going to want natural language generative AI. So we're going to see that, but it won't be a huge number, and those costs get amortized across millions of users. OpenAI just crossed $10 billion in revenue, and that tells you people are really using and paying for the service. The costs get spread over a massive amount of users. There's also innovation that's going to drive down the cost. Even though so much cost and power of GPUs and CPUs power the engines underneath this capability, the cost per token — a token meaning the question you ask a generative AI — is actually dropping dramatically. We're figuring out how to bring applications into much less expensive applications. Think about an application that helps you be more effective at running these broadcasts. It could run right on your PC. You weren't asking about all the questions in the world, but it was specific to the broadcast you're doing, the information you need. You likely can run those AI applications right on your PC to assist you in post-processing. So there's a lot of innovation going on to drop the cost, and we're seeing that bear out in the industry.
Well, yeah, that is the hope at this point in time. Speaking of collaborations, I want to understand from you how you are seeing this collaboration that AMD has with Humane. It's a $10 billion commitment being made and I think this is an issue that many countries including India are grappling with at this point in time. When we talk about the need for sovereign models, how do we fund sovereign models? Take me through what the bet is here.
Well, we're very excited about our joint venture. It's actually a joint venture that we did with Humane. It's going to be investment with partners to really grow that compute capacity. Humane and the Kingdom of Saudi Arabia have power. They have energy sources. They have the willingness to invest, as do we. So it allows us to create a partnership to really build out that compute capacity to address that insatiable demand going forward. I do think we're certainly involved in discussions like that — maybe not the magnitude of what we just announced with Humane, but exactly discussions like that we have going on in over 40 countries where we're engaging with them, and they want choice. Many of them are writing in their bids: 'Hey, we're going to create a sovereign AI capability, but we don't want it all one vendor. We don't want to be locked in.' They're writing in the requirement that it be multi-sourced. I think every company, every country is grappling with this, and certainly includes India. We've been very excited by the attention from the Ministry of Science and others that are giving to build up AI capability in India.
You know, talking about India, you said you're in conversation with the government in India along with conversations that you're having with governments across 40 other countries. India's rolled out its semiconductor mission. You know, 10,000 GPUs now, 30,000 is what the government is looking at. How do you see India today, Mark, from a semiconductor mission point of view? As you look at how things have panned out, are there things that you believe the government needs to put on the table to ensure that this picks up pace and we see further acceleration on the back of the measures already taken?
We're very pleased to see the initiatives of what's going on not only in the AI investment but overall making India an investment that is driving up the technology capabilities there. It's very important for us. We have a very strong presence in India, Shireen, as you know. We've been very active there. In 2023, I had the opportunity to pledge a $400 million investment that we were making in India at the semiconductor conference. We are well along on that investment from 2023 to 2028. Likewise, discussions on computing have proceeded very well. We have excellent engagements on supercomputing. Two of the top supercomputers are powered by AMD compute technology, and we have engagements with multiple universities including the Indian Institute of Science where we've stood up a test cluster where students and faculty can run AI across different types of devices and push the edges of research. As you said, there will be tenders that the government is putting out as they grow their GPU capabilities. Our ask of the Indian government in doing these tenders is: please open it up. Make sure you don't get tied in and locked in with just one vendor — the dominant competitor in our field. Please open it up for AMD and others to compete.
