Arm Says AGI CPU Delivers Nearly Twice AMD EPYC's Modelled Performance Per Rack

Arm used the Open Compute Project (OCP) Global Summit in San Jose this month to argue that AI agents change what counts as good CPU performance. On 13 October the company published a methodology for measuring agentic workloads, developed with support from Fujitsu, Meta and OpenAI. Two days later it summarised the week's progress on the Arm AGI CPU, its first production silicon.[1][2]

The headline claim comes from Signal65, a research firm. Its report says the AGI CPU delivered an average of 1.99x the modelled performance per rack of AMD's EPYC 9755 across seven workloads. The study was commissioned by Arm, and the rack-level numbers are modelled, which limits how far they can be compared with a customer's own tests.[3][1]

Arm argues that an AI agent behaves like a virtual user with its own persistent computing environment. It keeps state, calls tools, runs code and checks results until a task is done. Multiply that across millions of agents and the CPU has to sustain many concurrent jobs, not just finish one quickly. Arm cites data highlighted by a16z showing agent token use on OpenRouter up roughly 14 times since February.[1]

Arm's methodology rests on three principles: representative agentic use cases, repeatable CPU measurements and validation against end-to-end outcomes. Fujitsu, Meta and OpenAI are named as supporters of the approach. Their supportive statements concern the need for transparent, representative measurement, and they do not endorse Arm's own results.[1][2]

Signal65 reports that the AGI CPU averaged 1.7x for tooling, 2.0x for end-to-end agentic workflows and 2.2x for reasoning and planning, measured as modelled performance per rack. Individual workloads ranged from 1.43x to 3.19x. Arm's rated-power model fits 21 to 23 AGI CPU servers into a 24 kW rack against 13 EPYC 9755 servers, which drives most of the result.[3][1]

The report also says the AGI CPU created agent sandboxes faster than a two-socket AMD EPYC 9965, taking 45% less time for cold starts. Signal65 states that the tests used early AGI CPU systems, that rack results come from Arm's rated server power and measured per-chip results, and that follow-on testing on final hardware is planned.[3]

At OCP, Arm showed the chip inside rack designs. Supermicro presented liquid-cooled AGI CPU systems, including a 21-inch configuration supporting 128 CPUs and a wider design scaling to 336. Arm says it is contributing an AGI CPU reference server to OCP, and Meta's Yosemite YV5 rack-scale design brings the processor into an OCP rack.[2]

Arm also paired the CPU with accelerators from FuriosaAI and Rebellions, with the CPU handling agent orchestration and tool execution. Arm first announced the AGI CPU on 24 March 2026. The October material shows production systems and partner designs but gives no customer volumes, prices or delivery dates.[2][4]