If You Can Make A Compute Engine, You Can Sell A Compute Engine

AMD CEO Lisa Su announced on a Thursday conference call that the company now projects the total addressable market for data center CPUs will reach $120 billion by 2030, doubling its November 2025 forecast of $60 billion. The revised estimate reflects surging demand driven by agentic AI workloads, which require more CPU cores relative to GPUs, with Su noting that CPU-to-GPU ratios are shifting from 1:8 toward 1:1 or even more CPUs than GPUs in some deployments. The forecast signals a tectonic shift in the server market, where CPUs for agentic AI are expected to dominate the installed base within years, creating new competition among AMD, Intel, Nvidia, and Arm.

COMpute If You Can Make A Compute Engine, You Can Sell A Compute Engine The title says it all, really. But make no mistake: There is still plenty of competition in the markets for CPUs, GPUs, and XPUs and the DRAM, HBM, and flash memory that they all depend upon despite the demand exceeding supply. The competition is now focusing on access to devices as much as it ever did on feeds, speeds, and price. And, there is a new market emerging for CPUs, driven by the specific computing needs of agentic AI, that will, within the next few years, come to dominate the installed base of servers even if it does not represent the lion’s shares of revenues as do AI training systems by virtue of those very expensive GPUs they are laden with. That is the new gospel according to AMD, whose top brass echoed many of the comments made last week by Intel as there is not just a resurgence on spending on CPUs – which there most certainly is – but a tectonic shift underway thanks to agentic AI. This is the beginning of a trend, so its outlines are a bit vague. It was only back in early November 2025 at its Financial Analyst Day that Lisa Su, chief executive officer at AMD, put a stake in the ground and predicted that the total addressable market, or TAM, for datacenter CPUs was going to grow at a compound annual growth rate of 18 percent, reaching $60 billion by 2030. That is a lot of server CPUs in the datacenter. This compares to a roughly $13 billion to $14 billion server CPU silicon TAM back in 2015, when Intel accounted for close to 99 percent of shipments and when most of the RISC/Unix and Itanium alternatives were either dead or dying and Arm server CPUs had been at it in earnest for four years but had damned near zero traction. If $60 billion sounds like a lot, how do you like $120 billion? Because that is the new server CPU TAM that Su & Co laid down in the conference call with Wall Street analysts yesterday going over the company’s first quarter of 2026 financials. Which we will get to in a second. So now, the CAGR between last year and 2030 for server CPUs is 35 percent, and most of the incremental money that AMD, Intel, Nvidia, Arm, and the homegrown CPU crowd will chase will be allocated to agentic AI systems, which are distinct from traditional system of record and system of engagement systems that have been the foundation of the server market for decades. Su called this general purpose compute, and she referred to the relatively thin slice of GenAI training and inference where the CPU is in the head node as a second part of the server CPU TAM. The third slice of the pie – one that appears to be growing exponentially – “are CPUs just for all of the agentic AI work,” as she put it. Here is how Su explained it further: “So you should think about we need all of the accelerators to run these foundational models, and then as these agents do work, they spawn more CPU tasks. So I would say it is largely incremental. The key is to make sure – what we are seeing is in these deployments – the ratio of CPUs to GPUs are the right ratio. So if you are installing a gigawatt of compute, the percentage of CPU as part of that gigawatt will increase.” “Some of the conversation in the industry has been about CPU to GPU ratios. And it's very hard to call exactly, but we certainly see the movement towards where in the past, the CPU to GPU ratio was primarily just as a host node in like a 1:4 or 1:8 configuration node, now changing and getting closer to a 1:1 configuration – or you can even imagine if you get lots and lots of agents that you could have more CPUs than GPUs. But all in all, I think it is largely additive to the TAM. And the key is that everyone is now planning and thinking about CPUs at the same time that they are thinking about their accelerator deployments, which is a good thing.” Of course, when you are counting sockets, you have to count what you are putting into the sockets and how big those sockets are. Clearly, the 72-core “Grace” CPU was a bit much for the “Hopper” CPU that Nvidia paired it with because with the next generation “Blackwell” CPUs, there were two GPUs attached to every CPU. And in fact, there were two Blackwell GPU chiplets per socket, so the ratio was really 1:4. And as far as we know, Nvidia will have one “Vera” CPU for every pair of “Rubin” GPUs – which is actually four GPU chiplets in total for a 1:4 ratio – later this year and will double up the GPU chiplets in 2027 in the Rubin Ultra socket to effectively get a 1:8 ratio. If there are eight compute tiles on the Vera CPU, then maybe we can say the ratio is really one to one. It is tricky to count this, as Su has said. And it will get trickier when banks of CPUs start doing some inference for smaller models in a mixture of experts GenAI platform. The rule will be simple: Always use the smallest model that will get the right answer with acceptable accuracy. And that will mean that something we have contended for many years will come true: A lot of AI inference will be done by CPUs. And that means there will be special cost-optimized CPUs from Intel, AMD, and Arm that are good at this agentic workflow and are different even from the host CPU job inside of GPU and XPU nodes. That is precisely what the “Verano” variation on the “Venice” Epyc 9006 theme seems to be. Verano is apparently coming out in 2027, very likely well ahead of the “Florence” Epyc 9007 processors expected perhaps in late 2027 or early 2028 and very likely paired with the next generation MI500 series GPUs and the kicker to the Helios racks that will sport Venice CPUs and MI400 GPUs. The Venice CPUs are slated for later this year, of course. “We are clearly feeling like we are seeing significant share gain as we are going into our Turin portfolio that has ramped very nicely. Venice is extremely well positioned, and we are working with customers right now on – beyond Venice and what we are doing in those architectures. So we feel really good about the market as well as our opportunity to grow to greater than 50 percent share of that market.” So there it is, gauntlet laid down at the feet of Lip-Bu Tan, chief executive officer of Intel. In any event, this agentic AI server CPU boom is just starting to materialize and it is not what drove the impressive Epyc CPU revenues in Q1 2026. What drove that AMD boom is a shortage of supply and an excess of demand. As we have contended for some time now, companies need to upgrade their general purpose machines and consolidate down generations of older gear to save power, space, and money so they can afford to invest in AI. Which companies are doing, too. There is more demand than meets supply, but both Intel and AMD have been increasing supply and therefore revenues are rising. It doesn’t hurt pricing that demand is still bigger than supply, and neither company has any incentive to make as many chips as they can sell. With substrates and other components also being in short supply, it is better to hang back, charge a premium, and bring more dough to the bottom line. Which is exactly what CPU and GPU makers are doing. They are selling lots of current generation products, and because of the high demand, they are also able to sell back generation products that are cheaper to make and more profitable at this point, too. In the March quarter, AMD brought in a total of $10.25 billion, up 37.8 percent year on year and down a tad sequentially. Operating income, thanks to the pricing power that AMD is enjoying, rose by 83.1 percent to $1.48 billion, and net income rose by 95.1 percent to $1.38 billion. That’s only 13.5 percent of revenue, mind you, at the bottom line, so AMD is not gouging anyone so much as being able to command the some of the premium it has long deserved. The company is also making massive investments in future server and PC architectures and that is also weighing on profits. But better that than get into a technical debt hole as both AMD and Intel have done in the past. AMD exited the quarter with $12.35 billion in cash and equivalents, up 68.9 percent from a year ago. That is not quite enough money to buy Cerebras Systems before it goes public next week on May 14. . . . AMD’s Data Center group, which sells server CPUs, DPUs, high-end FPGAs, and soon rack-scale components that its partners will use to create Helios double-wide racks for AI systems and potentially for HPC applications , saw revenues grow by 57.2 percent to $5.78 billion. Operating income rose by 71.6 percent to $1.6 billion, and that income represented 27.7 percent of sales. That’s about average for AMD these days. The Client group, which sells chips for PCs, had $2.9 billion in sales, up 25.8 percent, and the Gaming group, which sells high-end graphics cards for gamers and sometimes for those trying to do HPC and AI on the cheap, had $720 million in sales. AMD has combined operating income for the Client and Gaming groups because it is trying to mask something, and that something is that the PC chip business is not that profitable, with what we estimate is only $403 million, representing 14 percent of sales. The Gaming group does much better, at $173 million in our estimates, representing 24 percent of revenues for its operating income. That leaves the Embedded group, which sells FPGAs and which had $873 million in sales, up 6.1 percent, with $338 million in operating profit, which is 38.7 percent of sales. Yup. FPGAs continue to be the most profitable part of AMD – and partially thanks to the bad management of FPGA rival Altera by Intel. Altera has been spun out and we need to circle back and take a look at how it is doing with its newfound freedom. The Data Center group continues to grow in importance for AMD, and represented 56.3 percent of revenues in Q1 2026. Only three years ago, Data Center group only accounted for a quarter of AMD’s overall revenues, and a decade ago the datacenter represented a paltry 2.5 percent of sales. Lisa Su said she would fix AMD and give it credibility in the datacenter, and she has absolutely done that. Now comes the spreadsheet witchcraft that we so enjoyed. We were so pleased to report last quarter that the Instinct GPU business was larger than the Epyc CPU business for the first time in AMD’s history. Well, thanks to the CPU boom and the dearth of HBM memory out there in the world, and a product transition from the MI300 series to the MI400 series, the Instinct line was not able to keep pace with the Epyc line in the first quarter. Our model shows the Instinct GPU business growing by 64 percent to a smidgen above $1.9 billion in the March quarter, which is great, but it also is a 28.2 percent sequential decline from the record $2.65 billion AMD racked up in datacenter GPU sales in Q4 2025. The Epyc CPUs, by contrast, grew by 53 percent to $3.65 billion in our model, and based on hints that Su & Co gave out on the call, we think sales to hyperscalers and clouds represented 76.2 percent of that CPU pie, which is $2.78 billion, while enterprises, telcos, service providers, academia, and governments accounted for $867 million in Epyc sales, up 51.4 percent. The rest of the Data Center group dough, which is barely material, went to datacenter network interface cards, DPUs, and datacenter-class Xilinx FPGAs. A final note: When AI settles down – that was meant to be funny – all of this agentic management CPU function could end up being hard-coded in an agentic router. This is, in fact, what happened to network routers. They were originally just software running on proprietary minicomputers.