Arm AGI CPU delivered an average of 1.99x modeled performance per rack against AMD EPYC 9755 across seven representative agentic AI workloads, according to independent analysis published by Signal65 and presented at the OCP Global Summit 2026 in San Jose. The figures, modeled from independently verified server-level measurements, anchored a show defined by deployment evidence rather than roadmap promises.

Arm introduced AGI CPU, its first production silicon purpose-built for agentic AI, in March at Arm Everywhere. At OCP 2026, production systems, customer traction and an open agentic rack architecture moved the processor from launch to deployable platform.

What did the keynote say?

Mohamed Awad, Executive Vice President and General Manager of Arm's Cloud AI Business Unit, framed the shift in his keynote "The Agentic Decade Starts With Open Infrastructure":

"Rather than intent initiating outside the data center at human pace, we're moving to a world where intent is now being initiated within the data center at machine pace."

That change, Awad argued, redefines three CPU roles: AI head nodes that keep accelerators productive, cloud-native CPUs running applications and services, and agentic CPUs supporting continuously executing virtual users. Arm-based DPUs complement these with networking, storage, security and isolation. Arm offers partners a range of engagement models, from processor IP and compute subsystems to Arm AGI CPU itself.

How does the silicon reach the rack?

Arm's Open Agentic Infrastructure Platform combines general-purpose compute, AI acceleration, memory, networking, I/O and system management in a modular, OCP-aligned architecture. Two rack designs illustrated the range:

• Supermicro showed Arm AGI CPU systems on OCP's Open Rack v3 (ORv3) and Open Rack Wide (ORW), including a 21-inch liquid-cooled configuration holding 128 Arm AGI CPUs and a wider design scaling to 336 CPUs.
• Arm and Meta presented AGI CPU reference server work for Meta's Yosemite YV5 rack-scale design, expanding the base of reusable infrastructure around the processor.

The configurations let providers tune compute density, power delivery, cooling and serviceability together, avoiding lock-in to a single hardware design.

Which partners build balanced systems?

Two collaborations placed AI workloads on the processor best suited to each stage:

• FuriosaAI paired Arm AGI CPU with its RNGD accelerators — the CPU runs agent orchestration, tool execution and supporting services while the accelerators handle model prefill and decode, with host memory adding context capacity.
• Rebellions showed its RebelServer architecture, combining Arm AGI CPU with Rebel100 accelerators in a co-designed inference platform.

A booth demonstration ran an agentic customer-service workflow, showing agents executing multistep tasks beyond response generation.

How should agentic CPU performance be measured?

Ravi Malhotra, Director of Solutions Product Management at Arm, argued in his session "Rethinking CPU Performance for the Agentic AI Era" that traditional benchmarks miss the profile of agentic workloads. Software development agents demand strong per-core performance for compilation and testing; enterprise automation agents and personal assistants generate large numbers of short tool calls, making concurrency and performance per rack decisive.

Arm introduced a new measurement methodology — developed with support from Fujitsu, Meta and OpenAI — that uses representative agent workloads and repeatable CPU measurements, including replaying recorded agent execution sequences. The target measures how efficiently infrastructure supports useful work when agents run concurrently at scale.

The Signal65 result supplied the independent proof point: 1.99x average modeled performance per rack over AMD EPYC 9755 across the seven workloads.

What makes the infrastructure software-ready?

Arm detailed its system architecture and enablement work: SystemReady, the Server Base System Architecture (SBSA), the Server Base Manageability Requirements (SBMR) and the Arm Architecture Compliance Suite. These establish common foundations for hardware, firmware, security and manageability, reducing the need for software vendors to requalify products for every server design.

On security, Arm and Fujitsu presented how emerging IETF attestation standards, including CoSERV, plus open-source work through Project Veraison, can make trust interoperable across confidential AI environments.

The 2026 summit closed the loop on the open converged AI data center vision Arm laid out a year earlier. The ecosystem around AGI CPU now spans silicon, racks, benchmarks and standards — measured, not promised.

By Amara Osei