Semianalysis has published a fantastic analysis of the current state of CPUs and GPUs in datacenters, including a thorough history of CPUs and GPUs in datacenters. This is my summary.

Global idea: from 2023 to 2025Q4, AI training made GPUs more important than CPUs in datacenters. However, since then, Reinforcement Learning and vibe coding are making CPUs more important again.

Also important: ARM is growing (with processors like AWS Graviton and Nvidia Grace), AMD is growing and Intel is shrinking.

A brief history of CPUs and GPUs in datacenters

1990s: origin of the modern datacenter. The success of the PC made it possible to sustitute costly workstations and IBM mainframes. Intel creates chips for servers: Pentium Pro (1995) and Xeon (1998), with more cache than desktop chips.

2000s: dot com era. GHz race finished because of the end of Dennard scaling: around 2004 decreasing transistor size stopped proportionally improving the power consumption, so frequency could not increase further. Multicore era starts, with more cores, SMT (Simultaneous Multithreading) and multisocket systems instead of higher clock speeds …