SANTA CLARA – For most of the past decade, AMD and Intel ran the server CPU market like a two-party system. Arm-based chips were an interesting experiment, strong in smartphones and tablets, and quietly growing inside cloud providers’ custom silicon programs. That comfortable arrangement is now facing its first serious commercial disruption, and the name at the front of it is Nvidia.
Nvidia has shipped more than 2.5 million Grace server CPU units and is already transitioning customers to its next architecture, Vera, which the company says runs 50 percent faster than x86 processors, the instruction set AMD and Intel have shared since the 1980s. Nvidia is projecting $20 billion in Vera server CPU revenue in the current fiscal year and describes a $200 billion long-term opportunity in the market it is now openly targeting.
Those numbers are not modest. Intel’s entire data center and AI segment generated $6.3 billion in its most recent quarter, a 59 percent year-over-year increase that would look like strong growth in almost any other market. AMD reported $5.8 billion in data center revenue in the same period, also up 57 percent year-over-year. Together, AMD and Intel are running at roughly a $50 billion annual data center run rate. Nvidia is targeting the market four times larger than that, and it already has shipments in the field.
The architectural shift underneath all of this is Arm. Counterpoint Research projects Arm-based server CPUs will hold nearly 90 percent of the server CPU market by 2029, up from 13.2 percent at the end of 2025. A projection of that scale is not incremental; it describes a reordering that rewrites which companies build data center infrastructure and who captures the value from it.
Meta documented a 2x improvement in performance-per-watt when it deployed Grace on its infrastructure. OpenAI, Anthropic, and SpaceX are among the early Vera customers. These are companies building AI-native workloads on new infrastructure, and the migration cost to Arm for them is meaningfully lower than it would be for a traditional enterprise IT department running legacy applications. That distinction matters for understanding how quickly this shift can realistically move.
AMD’s $5 billion partnership with Anthropic to deploy MI450 GPU clusters, announced less than two weeks ago, showed AMD cementing a position as the serious alternative to Nvidia in AI compute. That framing holds if Nvidia’s CPU ambitions stay with the AI-native customers who are already migrating. It holds less clearly if Nvidia succeeds in selling enterprise data center buyers an integrated GPU plus CPU plus networking stack that removes AMD from the equation.
Intel’s situation is harder to defend on paper. Its data center recovery centers on Gaudi AI accelerators and execution discipline after years of manufacturing missteps. On the CPU side, Intel’s answer to Arm’s rise remains x86 efficiency improvements that do not fundamentally address the performance-per-watt argument Nvidia is now making. Intel has an Arm server CPU program, but it has not been a development priority.
AMD has more options. Its GPU architecture expertise gives it real understanding of how AI workloads behave at the hardware level. Supermicro’s $60 billion order backlog for AI server systems suggests the demand for Nvidia-centered racks is genuine and sustained, but the backlog is GPU-driven. AMD does not yet have an Arm-based server CPU line to counter Vera, and its MI450 is a GPU for AI workloads, not a general server CPU replacement. That gap is what the next cycle of AMD’s product roadmap will need to answer.
What performance projections like Nvidia’s always require is context. The 50 percent advantage claim over x86 covers benchmarks optimized for Arm workloads. Enterprise IT departments with a decade of x86 software optimization baked into their middleware and runbooks need more than a benchmark to justify migration. Arm server software compatibility has improved substantially, but it is not yet seamless across every enterprise application category, and traditional procurement cycles move slowly regardless of what the benchmark shows.
The hyperscaler dimension adds another layer. Microsoft, Google, and Amazon have each built their own custom Arm data center chips, a trend that already weakened Intel’s position before Nvidia entered the space. Nvidia’s Vera competes not only with AMD and Intel’s off-the-shelf silicon but also with designs those hyperscalers are building in-house. Whether enterprises buy Vera as a Nvidia-branded alternative or continue to buy AMD and Intel merchant silicon is the commercial question on which Vera’s revenue projections actually depend.
According to Nvidia’s newsroom, the company is already using Vera internally to accelerate the design of its next-generation CPUs and GPUs, a deployment that builds operational confidence in the architecture before it reaches broad commercial scale.
Nvidia’s N1X chip already demonstrated the company’s willingness to compete in Intel’s home territory in client computing. Vera is the same move executed in the data center, one tier up in market size and strategic consequence for AMD and Intel’s core businesses. The next quarterly earnings calls for both companies will be the first occasion for investors to ask directly how their server CPU roadmaps answer what Nvidia is shipping today.
