The computer chip that the AI boom was supposed to make irrelevant is having a comeback, and Bank of America has put a number on it. The server CPU market will grow from $35 billion in 2025 to more than $170 billion by 2030, a fivefold expansion, according to a note published June 11 by BofA Global Research analyst Vivek Arya.
The forecast, well ahead of the firm’s prior estimate of $125 billion, rests on a single driver: agentic AI. Software agents that act on a user’s behalf, calling tools, reading files and chaining together tasks, run almost entirely on CPUs. The shift in workloads is dramatic. In traditional AI training, each GPU is paired with eight to twelve CPU cores. Basic inference needs sixteen to twenty-four. Agentic AI, according to industry experts interviewed by UBS in a companion analysis, demands eighty to one hundred twenty CPU cores per GPU, five to ten times the requirement of training workloads.
The reason is arithmetic. A single agent and its derived sub-agents can occupy one to four CPU cores, and a single complex task can generate ten to a hundred sub-agents, each spinning up its own threads, memory and tool calls. Analysts who study the workloads say that in agentic AI inference, 70% to 80% of computing resources shift to CPUs, a reversal of the traditional pattern in which GPUs carry 70% to 80% of the load. The CPUs doing that work need more cores and higher frequencies than the server processors of the training era, which pushes up average selling prices at the same time volumes rise.
Arya framed the market in a single sentence: “We view the emergence of agentic AI as a powerful demand accelerant that expands the CPU opportunity and lifts both Intel and AMD and Arm-based challengers.” The beneficiaries are the usual suspects, plus a few new ones. Intel and AMD own the x86 server market, where Intel still holds roughly 85% share in traditional workloads, and both have been rerated as the inference wave built: Intel’s shares have gained about 436% over the past twelve months and AMD’s about 280%, moves that predate but anticipate the CPU story.
Arm-based players are the wild card. Arm architecture holds about 15% unit share in server CPUs today, and UBS projects that share will climb to 40% to 45% by 2030, with revenue share reaching 50% to 55% because of higher prices on AI-class chips. The designers in that camp include Ampere, the cloud-native CPU maker, and Nvidia itself, whose Grace CPU ships alongside its accelerators and whose Vera processor is set to follow. Qualcomm is entering the data center with its own server CPU. The market is becoming crowded at the top.
AMD has been the most explicit about its ambitions. At its investor day in November, the company said it expects the server CPU market to grow from $26 billion in 2025 to about $60 billion by 2030, with AI-driven CPUs accounting for roughly half of the 2030 market, and it targets capturing more than half of the overall market. Intel’s answer is its Coral Rapids product line, designed to narrow the gap with AMD and Arm in AI head nodes, and the company argues its PC spillover effect will accelerate upgrades as agents move computing to local devices.
The forecast does not mean GPUs are losing their place. Data center spending still flows overwhelmingly to accelerators, and Nvidia’s data center revenue dwarfs the entire server CPU market. But the CPU’s share of the compute bill is rising, and the direction of travel matters: AI is moving from training, a concentrated activity, to inference, a distributed one, and inference runs on whatever server is closest to the user. That server runs on a CPU.
The market for the chips themselves is also changing shape. The analysts divide the future CPU business into three segments: the traditional server market, which grows steadily; AI head nodes, the CPUs bundled into GPU racks to orchestrate tasks; and standalone AI racks, pure CPU servers dedicated to the tool invocation and sub-agent processing that agentic workloads require. The third segment, which did not exist two years ago, is where the growth is.
The forecast also helps explain the financing wave running through the industry. Hyperscalers planning $180 billion to $200 billion annual capital budgets are buying servers, not just accelerators, and the CPU content of each rack is rising as inference workloads scale. That is why the chip companies most exposed to the server socket, Intel, AMD and the Arm designers, have seen their shares reprice so sharply over the past year even as Nvidia continues to dominate headlines. The money flowing into AI data centers is beginning to flow into the parts of the machine that run the software, not just the parts that train it.
BofA’s $170 billion by 2030 is one forecast among several, and they vary: AMD’s own guidance implies a smaller market, and other firms have published similar or larger numbers. What they share is the direction. The CPU, written off as a commodity in the GPU era, is being repriced as the workhorse of the inference age. The market is no longer a one-chip show, and the companies that own the server socket are the quiet beneficiaries of the loudest boom in technology.
