Not Just GPUs: Agentic AI Brings CPUs Back

The AI boom has been told, for three straight years, as a story about one kind of chip. Graphics processors trained the models, served the models, and reordered the global semiconductor industry around themselves. Data centers were designed as GPU farms with CPUs attached, and the ratio told the story: roughly one processor for every four, or even eight, accelerator chips.

The newest phase of the build-out is quietly rewriting that ratio. Agentic AI — systems that plan tasks, call tools, make decisions and act on a user's behalf — is dragging the center of gravity back toward the CPU. The reason is structural: agents do not just generate text. They browse, run code, query databases, coordinate sub-agents and evaluate whether a request has actually been fulfilled. The orchestration layer that schedules sub-tasks, routes tool calls and moves data between models is classic CPU territory, and it scales with every agent a company deploys.

Research is beginning to quantify the shift. A November 2025 paper published on arXiv, "A CPU-Centric Perspective on Agentic AI," found that tool processing on CPUs can account for up to 90.6 percent of total latency in agent workloads, and that CPU dynamic energy consumption can reach 44 percent of the total at large batch sizes. Translated from the engineering: in many agent systems, the model is not the bottleneck. The bookkeeping is.

## The Math That Reshapes the Data Center

Arm, whose designs sit at the center of this shift, estimates that a traditional AI data center needs about 30 million CPU cores per gigawatt of capacity, but that demand will surge to roughly 120 million cores per gigawatt in the agent era — a fourfold increase. The expected CPU-to-GPU ratio in agent-heavy facilities moves from today's one-to-four through one-to-eight, toward one-to-one or one-to-two.

If those numbers are even half right, they rewrite procurement plans across the industry. The first confirmation is already in the pricing: Intel and AMD raised prices across select server CPU lines toward the end of the first quarter of 2026, a sign that demand has outrun supply even before most enterprises have deployed agents at scale.

## Intel's Fight, AMD's Momentum, and a Wave of New Entrants

The server CPU market this demand lands on barely resembles the one from five years ago. Intel's Xeon once commanded over 95 percent of the data center market. Then yield problems on the Intel 7 process delayed the Sapphire Rapids launch by nearly two years, and AMD's EPYC Milan walked through the opening. AMD has never given that ground back, and 2026 stacks the fight: Intel launched Xeon 6+, known as Clearwater Forest, on its new Intel 18A process with 288 cores and Foveros Direct hybrid bonding, with the next-generation Diamond Rapids waiting behind it. AMD's EPYC Venice arrives on TSMC's N2 process with 256 cores and 512 threads, and industry analysts expect AMD to keep taking share through the year.

The more striking change is who else now sells server CPUs. In March 2026, Nvidia began selling its Vera CPU as a standalone product — 88 custom Olympus cores on TSMC N3, tied to its GPUs by an NVLink interconnect rated at 1.8 terabytes per second — with early partners including Alibaba, ByteDance, Cloudflare, CoreWeave and Oracle. Weeks later, Arm did something it had not done in 35 years of licensing blueprints instead of building chips: it launched the AGI CPU, a 136-core Neoverse V3 design, with launch partners including Meta, OpenAI and Cerebras. The largest cloud providers, meanwhile, have been designing their own Arm-based chips for years.

The strategic logic is what matters. Nvidia selling CPUs is not diversification for its own sake; it is rack-level lock-in, because a CPU and GPU stitched together with a proprietary interconnect are harder to unbundle than chips bought separately. Arm becoming a platform company puts it in subtle competition with its own licensees, betting that in the agentic era, selling complete systems beats collecting royalties. Intel is fighting for relevance on the strength of its manufacturing comeback, and AMD is arguably best positioned on raw performance per dollar.

And the winner will be decided less by core counts than by glue. Agent orchestration is latency-sensitive: what limits throughput is often memory bandwidth and the interconnects between CPU, memory and accelerator, not the peak speed of any single chip. That is a systems problem, and it favors companies that sell whole architectures rather than parts.

## The Inference Economy's Second Act

Training a frontier model is a sustained, parallel grind — the workload GPUs were born for. Serving an agent is bursty, branch-heavy and full of waiting: on web crawls, database queries, tool responses and human approvals. The AI industry's center of gravity is moving from training runs to inference and action, and the hardware mix is following it. Every agent a company ships quietly increases demand for exactly the chips the industry spent three years treating as accessories.

## What This Means For You

**If you buy or plan infrastructure:** recalculate. Agent workloads mean counting orchestration cores and memory bandwidth, not just accelerators, and the two big x86 vendors have already moved prices up. Facilities designed purely as GPU farms a year ago may be badly under-provisioned for the agent workloads they will actually run.

**If you invest in the AI theme:** the trade is broadening. Exposure to the build-out no longer means GPU exposure alone: Intel's execution on 18A, AMD's share gains, Arm's platform pivot, and the interconnect and memory suppliers around them are all separately priced bets on the same shift. That cuts both ways — more ways to participate, and more ways to be wrong about which layer captures the value.

**If you build software for a living:** the orchestration layer is becoming a career path. Scheduling, tool routing, memory tuning and latency profiling around agents are growing as fast as the agents themselves. The GPU race priced the model. The CPU race prices everything the model does next.