The Chip Nobody Wanted Is Having the Best Year in AI
Is the AI industry building around the wrong bottleneck?
The infrastructure conversation for the past three years has organized itself around one number: GPU count. Servers spec'd by GPU. Data center leases by GPU-to-rack ratio. Startups pricing in GPU-hours. The CPU was the forgotten neighbor in the rack — necessary, unglamorous, and priced accordingly. UBS has a contrarian view, and it arrived in a research note this week: the next wave of AI compute demand might not be built in the data centers everyone is racing to fund.
The supply side is telling a different story right now. Intel cannot make enough Xeon server processors to fill current demand — a position the company has not held for most of the past decade. AMD CEO Lisa Su told investors the server CPU market is now expected to grow more than 35 percent annually, reaching over $120 billion by 2030, a forecast she said had doubled in six months. UBS analyst Randy Abrams put the structural case together in a May note: the rise of agentic AI is driving the next wave of computing demand, shifting from simple tasks to complex, real-world work. The firm's revised forecast calls for more than 40 percent annual growth in the CPU market Investing.com / UBS Research — with the ARM architecture gaining ground fastest. UBS projects ARM-based server CPUs will reach 40 to 45 percent market share by 2030, up from roughly 15 percent now Investing.com / UBS Research. Those are not the numbers of a market everyone has been overlooking by accident.
The architecture of AI is changing. The first wave of generative AI was built around training large models — a GPU-dominant workload. Agents work differently. They do not just generate a response and stop. They spin off subagents, call tools, execute code in sandboxed environments, retrieve files, validate outputs, and manage persistent memory. Each of those operations runs on a CPU. UBS's Timothy Arcuri, citing expert interviews, found that traditional AI training typically requires 8 to 12 CPU cores per GPU. Agentic systems, he estimates, require 80 to 120 cores per GPU Investing.com / UBS Research. That is roughly a tenfold increase in the CPU attachment rate — and it changes the economics of the rack.
The CPU-to-GPU ratio is moving accordingly. Intel CEO Lip-Bu Tan said on his company's April earnings call that the ratio, which used to sit around 1 CPU for every 8 GPUs, has already moved to 1-to-4 and is trending toward parity Intel Q1 2026 Earnings Transcript. JP Morgan's analysts have said the ratio for agentic workloads could reach 7 CPUs per GPU The Inferential Investor. The OpenAI-AWS $38 billion compute partnership, announced in late 2025, explicitly included the ability to scale to tens of millions of CPUs alongside hundreds of thousands of GPUs The Inferential Investor — a provision that reads in retrospect like an early architectural confession.
High-end AI CPUs are commanding prices to match. NVIDIA's 144-core Grace CPU and AWS's 192-core Graviton 5 are each expected to command $3,000 to $4,000 per unit Investing.com / UBS Research. These are not commodity parts. They are the load-bearing orchestration layer of an agentic deployment.
The supply constraints confirm what the projections suggest. Intel cannot make enough Xeon server processors to fill current demand — a position the company has not held for most of the past decade Intel Q1 2026 Earnings Transcript. AMD CEO Lisa Su told investors the server CPU market is now expected to grow more than 35 percent annually, reaching over $120 billion by 2030 — a forecast she said had doubled in six months AMD Q1 2026 Earnings Call Transcript. Citi's estimate, published the same week as UBS's note: agentic AI CPUs could reach $59 billion by 2030, growing at a 185 percent compound annual growth rate Investing.com / UBS Research.
There is a second act UBS flagged, and it is the one that should make cloud infrastructure investors uncomfortable. Agentic AI tools are increasingly pushing workloads toward end-user devices — laptops, workstations — to take advantage of local compute, reduce cloud latency, and use what UBS calls free capacity at the edge. If agents can run well on hardware that ships in volume rather than requiring purpose-built data centers, the entire calculus of cloud AI infrastructure investment shifts. The capital expenditure programs that hyperscalers have committed to cloud data centers would need reassessment.
That is a big if. UBS's numbers are projections built on expert interviews and bottom-up models — not measured data from deployed agentic systems at scale Investing.com / UBS Research. Agentic AI is still early. The core-count requirements may compress as frameworks optimize. The PC upgrade cycle UBS mentions as a secondary tailwind is real but slow: most corporate PCs are on four-to-five-year replacement cycles, not annual ones. The structural argument is sound. The quantitative precision is not yet warranted.
But the supply constraints tell a different story than the headlines suggest. Intel cannot make enough Xeons. AMD doubled its TAM forecast in six months. Those are not projections — they are current-state observations from the people closest to the constraint.
The question for investors and builders is not whether the CPU matters in AI infrastructure. The CPU has always mattered. The question is whether the AI industry has been building around the wrong bottleneck — and whether the correction is already underway, in order books and supply chains, before the next earnings call tells the story the headlines have not written yet.