A few weeks ago we discussed how Agentic AI is fundamentally reshaping CPU demand in data centers. The ratio of CPU to GPU is moving from 1:4 toward 1:1, and in some configurations even 4:1. Orchestration workloads are turning CPU from a peripheral scheduler into a first-class compute bottleneck. Today UBS published a full report that puts hard numbers behind exactly this thesis.
Their expert calls quantify what I've been describing from hands-on experience. In traditional AI inference, 70-80% of compute sits on the GPU. In agentic workflows, that ratio inverts. 70-80% of the work lands on the CPU, across orchestration loops, tool calling, sandbox execution, and context handling. Each agent and its subagents consume 1-4 cores, with a single task spawning 10-100 parallel subagents. UBS estimates this drives a 3-8x increase in CPU demand per user.
The TAM math is striking. UBS projects the server CPU market growing roughly 5x from ~$30B today to ~$170B by C2030, validated through both bottom-up accelerator attach modeling and top-down analysis using NVIDIA's $3-4T AI TAM target. By 2030, they estimate ~40MM XPUs will require ~33MM AI CPUs, with the CPU-per-GPU attach ratio approaching 0.8x from 0.3x today.
The competitive split is where it gets most interesting. ARM takes ~42% unit share but ~52% revenue share by 2030E, driven by head node dominance (78% of AI head node CPUs) and meaningfully higher ASPs. NVIDIA Grace at 144 cores prices at $3,000-4,000, AWS Graviton 5 at 192 cores trends higher. AMD ranks second on multithreading and core density. Intel benefits least but still rides the rising tide on the traditional server and PC side.
One thing UBS flags but doesn't fully model: agentic workloads are increasingly shifting execution to edge devices and PCs, which could compress the cloud CPU multiplier from 5-8x down to ~4x. The eventual cloud-to-edge balance is the biggest open variable in this entire framework. Worth watching closely.