CPUs Are Back: The Datacenter CPU Landscape in 2026
CPUs are making a comeback in datacenters by 2026, driven by AI inference demand. Intel, AMD, and new entrants are competing on architecture, power efficiency, and total cost of ownership to reshape the server market.
Background
- After years of GPUs dominating the AI narrative, this article argues CPUs are making a comeback in the datacenter. The key insight: AI inference (running trained models) is growing faster than training, and CPUs are surprisingly competitive there — cheaper, easier to deploy, and more power-efficient than Nvidia's GPUs.
- Major chip battles covered: AMD (Bergamo/Turin), Intel (Granite Rapids/Sierra Forest), and a surge of ARM-based server CPUs from Ampere and Amazon (Graviton) challenging the x86 monopoly. Even Nvidia is building its own ARM CPU ("Grace"), signaling a shift.
- The datacenter is moving toward "disaggregated" architecture: pools of compute and memory connected by high-speed links, instead of fixed CPU-RAM per server. New CPU designs prioritize density and efficiency over raw speed.
- Why now: Datacenters are power-constrained. Hyperscalers (AWS, Google, Microsoft) need alternatives to Nvidia's expensive, power-hungry GPUs. CPUs offer a path there.