AMD launches 6th Gen EPYC chips built for AI agents

AMD has made its 6th Gen EPYC server processors, including the new SP7 family built on the company’s “Zen 6” architecture, generally available, with systems already shipping from major server OEMs and ODMs; cloud instances built on the new chips are expected soon from the major cloud providers, including Microsoft Azure and Oracle Cloud Infrastructure. AMD says the lineup spans workload-optimized designs for agentic AI, enterprise applications, cloud computing, high-performance computing and general-purpose work, extending a run in which the company says its 5th Gen EPYC chips became AMD’s fastest-ramping, highest-volume EPYC generation yet.

AMD’s central claim, by its own measure, is that 6th Gen EPYC is the best server CPU portfolio on the market; it says the flagship EPYC 9996 chip outperforms an Nvidia Vera CPU-based system by up to 3.7x, and tops an Intel Xeon 6980P system by up to 2.2x on performance and 2x on performance per dollar. “AI is significantly expanding the role of the CPU across every layer of the data center,” said Dan McNamara, AMD’s senior vice president and general manager for compute and enterprise AI. “With 6th Gen AMD EPYC, customers can standardize on the proven x86 foundation while choosing a workload-optimized solution to run more agents, keep accelerated infrastructure fed, and power the business-critical applications they depend on.”

AMD is framing agentic AI infrastructure as needing three distinct CPU roles — general-purpose services, GPU host nodes, and agent sandboxes — with a different chip tuned for each: a high-frequency variant running up to 5.3 GHz that AMD says delivers up to 1.8x the large-language-model throughput of the previous EPYC generation, and a high-density variant AMD says packs roughly 3.3x more agents per watt than Nvidia Vera and 2.5x more than Intel’s Xeon 6980P in sandbox deployments.

On top of individual chips, AMD is offering three rack-level configurations under its Helion brand — Helion Liquid, paired with AMD’s Helios rack system for maximum agent density; Helion Air, for adding agent capacity inside existing air-cooled data centers; and Helion Air Inference, which combines CPU execution with local GPU inference in one rack — with some configurations reaching, AMD says, up to 73,728 threads per rack. Meta is already validating 6th Gen EPYC platforms at scale as a lead customer, while Azure and OCI are rolling the chips into new cloud instances aimed at AI data processing, technical computing and production-scale inference.

Why it matters

AMD is betting that the next phase of the AI buildout runs through ordinary server CPUs as much as GPUs — hosting the agents, sandboxes and surrounding applications that sit around every AI model — and it’s using its own benchmarks to argue x86 server chips, not just accelerators, are where a meaningful share of that spending should go.