AMD's Matt Ramsay Reveals Server CPU Revenue Set to Exceed Entire 2025 Market by 2027 as Agentic AI Reshapes Demand
KeyBanc Technology Leadership Forum, August 11, 2026
Advanced Micro Devices offered investors one of its most detailed pictures yet of how quickly its server and AI businesses are inflecting, with Corporate Vice President of Financial Strategy and Investor Relations Matt Ramsay telling KeyBanc analyst John Vinh that the company's early view for 2027 server CPU revenue alone will be roughly 20% larger than the size of the entire global server market in 2025. The disclosure, made at KeyBanc's Technology Leadership Forum, underscores how dramatically the demand backdrop has shifted for AMD's Epyc lineup as AI spending pivots from training toward inference, and now toward what Ramsay described as a further shift within inference itself.
The Real Story Is Agentic Inference, Not Just Inference
Ramsay used much of his time to explain a nuance that he argued the market has not fully priced in: it is not simply that AI spending is rotating from training to inference, a shift AMD flagged at its Advancing AI event weeks earlier, but that within inference, chatbot-style workloads are rapidly giving way to agentic inference. That distinction matters enormously for AMD's CPU business specifically. "Agentic AI from a CPU perspective is not a monolithic workload," Ramsay said. "It's a very diverse set of workloads for which there are very many optimization points."
In practice, agents hand tasks to large GPU or XPU clusters to run the core intelligence, then automatically decide what to do next, pull data from the cloud, enterprise systems or the web, reorganize it, and hand it back for the next inference step. That orchestration work, Ramsay explained, requires high-thread-count, high-performance CPUs capable of juggling many simultaneous tasks, including running code that was just generated by a prior inference step. AMD is positioning its upcoming Venice CPU family, built on 2-nanometer process technology and now sampling with customers, directly at this workload, with core counts scaling up to 256 cores and 512 threads.
Server Growth Numbers Are Getting Bigger, Not Smaller
The magnitude of the ramp is striking. AMD's server business grew more than 50% in the first quarter, accelerated to more than 70% in the second quarter, with both cloud and enterprise segments growing above 70%, and management has guided to greater than 80% growth in both the third and fourth quarters. Layered on top of that much larger base, AMD's early view for 2027 calls for at least 70% growth again. Ramsay noted the context for how unusual this is: "There was a couple of years where the server market maybe grew a little bit less than it had historically as CapEx quickly shifted towards AI systems," and now growing 15% to 20% a year, once considered a strong result for the segment, looks quaint next to current trends.
The binding constraints on faster growth are not demand but supply. Ramsay cited TSMC wafer capacity, advanced packaging, given that Venice is the first server product in the industry to use it, and matched-set DRAM availability across the OEM and ODM ecosystem as the three gating factors. AMD has been working directly with TSMC, where it is typically the third- or fourth-largest customer, and recently committed $10 billion to Taiwanese ecosystem investment, weighted toward back-end packaging capacity, following a visit by CEO Lisa Su. Ramsay's framing of the industry's responsiveness was blunt: "If you give accurate forecasts and sufficient lead time, the industry is quite good at getting you supply. It's when you come and ask for stuff way underneath lead time where things get a little more complicated."
Helios Ramp: No Smoking Guns, But a Long Blocking-and-Tackling Exercise Ahead
On the Helios rack-scale AI platform, shipments of MI450 accelerators, CPUs and Pensando networking gear to ODM partners begin in September, with a large revenue ramp in the fourth quarter followed by another step-up in the first quarter of 2027. Ramsay quantified the scale of what is coming: data center revenue including AI is expected to more than double next year, growing "much more than 100%." Customer feedback on sampled systems running full model code has been positive, with Ramsay relaying the sentiment AMD wants to hear: "Wow, this thing really works the way that you told us it was going to work."
Addressing investor concern that AMD's first rack-scale launch could suffer the growing pains a competitor experienced in its own initial rack-scale transition, Ramsay pointed to the ZT Systems acquisition as having brought in system-level engineering talent that has "proven invaluable to hardening the design and derisking the different design points." AMD is deliberately concentrating its initial ramp with a small number of ODM partners before broadening distribution once the approach is proven, rather than scaling the full ecosystem simultaneously. "There's no smoking guns," Ramsay said. "We've shipped sampled racks, people are running code, they're very, very happy with it. We've proven that we can build racks. Now the question is the vast amount of blocking and tackling that we need to do to build the racks at the scale that we're talking about."
Gigawatt-Scale Commitments From OpenAI, Meta and Anthropic
Ramsay provided specifics on customer commitments underpinning the Helios ramp, which he characterized as a roughly six-to-seven-quarter product cycle running from September through the first quarter of 2028. AMD has secured 1-gigawatt commitments each from OpenAI and Meta, along with a 1-gigawatt commitment and a further 2-gigawatt ambition from Anthropic. Beyond those three strategic partners, Oracle Cloud Infrastructure customers, Microsoft Azure customers and Microsoft's own internal AI workloads, along with a number of neocloud operators, are also expected to run on Helios. Ramsay said AMD would "love to be able to do the first full gigawatt with Anthropic in 2027," while cautioning that execution on land, power, shell and capital commitments remains a variable.
ROCm Software Gap Narrowing, With a Telling Anthropic Anecdote
Perhaps the most concrete evidence offered on software progress was a story about Anthropic renting a cluster of MI355 accelerators earlier this year and getting its flagship inference model tuned and running within a weekend, without AMD's involvement or even its knowledge at the time. Ramsay called ROCm's progress over the past 18 months "phenomenal," crediting recent acceleration to AI tools now being used inside AMD's own software development process. Under the multi-gigawatt Anthropic partnership, AMD is also working to ensure code generated by Claude can automatically land on ROCm and AMD's Instinct platforms for any customer using Claude for model work. Ramsay argued the gap with Nvidia's CUDA has closed to the point where, for AMD's largest and most sophisticated customers, "it's not really a conversation" anymore, since those customers already know what they want their applications to do and simply route the work through ROCm.
A $220 Billion Server CPU TAM by 2030, Two-Thirds Agentic
AMD reiterated the updated total addressable market figure from its Advancing AI event: $220 billion for server CPUs by 2030, with agentic workloads representing roughly two-thirds of that pool. Management maintained its target of capturing more than 50% revenue share of that expanded TAM. Ramsay broke the market into three distinct buckets with different optimization requirements: medium-core-count, high-frequency chips running up to roughly 5 gigahertz as AI head nodes attached to GPU or XPU clusters; the highest core-count and thread-count parts for CPU-only agentic racks; and traditional enterprise and cloud workloads sitting in between, where deployment location is a business decision rather than a workload-driven one.
x86 Versus ARM: AMD Frames It as a Capabilities Question, Not an Instruction-Set Debate
Pressed on ARM's share gains in server, albeit from a smaller base, and how AMD stacks up against both ARM and its principal x86 rival, Ramsay pushed back on framing the competition as an instruction-set contest. He pointed to AMD's chiplet architecture, which allows a relatively small number of taped-out chiplets to be configured into a large number of SKUs, as having driven AMD's x86 server share from 0.4% years ago into the high 40s today. On agentic workloads specifically, Ramsay said the demand signal AMD is tracking is agents or threads per megawatt, and that customer demand is pulling toward AMD's highest core-count SKUs rather than any particular architecture. He also argued that x86's accumulated reliability, availability and serviceability features, hardened over five or six generations of enterprise and hyperscaler deployments, are difficult to replicate quickly, a point he tied to the security requirements of agents operating as autonomous workers with access to enterprise data.