AMD boss Lisa Su reckons agentic AI will make server CPUs matter again, which is handy for its EPYC business.
During AMD’s fourth-quarter earnings call, chief executive Dr Lisa Su discussed the company’s AI growth and competition in the AI CPU market. The call came as CPUs are being treated as useful again, thanks to agentic AI workloads.
AMD’s larger rival Intel saw its shares close 23 per cent higher last month after its own earnings report. Chipzilla posted a major profit beat, while chief executive Lip-Bu Tan talked up the importance of CPUs in today’s AI world.
Su started the call by describing the effect of AI on AMD’s cloud computing business.
“AI was the primary driver of growth in the quarter, as every major cloud provider expanded their Epic footprint to support a broad range of AI workloads, from general-purpose compute and data processing to head nodes for accelerators and emerging agentic applications,” Su said.
Head nodes are specialised computing clusters in a data centre that manage resources and direct the rest of the compute kit. In other words, they are the boring but essential grown-ups in the AI room.
Like Tan, Su said agentic AI is increasing the need for CPUs in server computing.
“Increasing and agentic AI are increasing the need for server CPU compute, as these workloads require additional CPU processing for orchestration, data movement and parallel execution, in addition to serving as the head nodes for GPUs and accelerators. As a result, we are seeing both stronger near-term demand and deeper engagement with customers on long-term capacity planning,” Su said.
Su believes the total addressable market for server CPUs will grow 35 per cent annually to hit $120bn by 2030. AMD’s previous estimate was 18 per cent annual growth, which now looks a bit shy.
Investment bank UBS was more chip-happy in its recent coverage, pencilling in a $170bn total addressable market by 2030. At its financial analyst day in November, AMD had outlined server CPU market growth of about 18 per cent annually during the next three to five years.
Asked whether extra CPU demand might cannibalise GPU demand, Su said it would be “largely additive to the [GPU] TAM”. That is exactly what any sane CPU vendor would say, but the logic is not daft.
Su said accelerators are still needed to run foundational models, while agents “spawn” CPU tasks. That means the CPU-to-GPU ratio becomes more important as data centres chase bigger AI deployments.
“So if you’re installing a gigawatt of compute, you know, the ratio, there’s a percentage of CPU as part of that gigawatt that will increase. You know, some of the conversation in the industry has been about, you know, CPU to GPU ratios,” Su said.
Su said the old CPU-to-GPU ratio was mostly about host nodes, often in one-to-four or one-to-eight configurations. Agentic AI could shift that nearer to one-to-one, which would be rather useful for AMD.
“And it’s very hard to call exactly, but, you know, we certainly see the movement towards, you know, where in the past the CPU to GPU ratio was primarily, you know, just as a host node, you know, in like a one to four or one to eight configuration, you know, now changing and getting closer to a one-to-one configuration or, you know, even, you know, you can even imagine if you get lots and lots of agents that you could have more CPUs than GPUs,” Su said.
For now, Su said the important shift is that “everyone is now planning and thinking about CPUs at the same time that they’re thinking about, you know, their accelerator deployment, which is a good thing.”
