Arm AGI CPU moves to deployment at OCP Global Summit 2026 with rack designs

From product launch to deployment

Arm introduced Arm AGI CPU, its first production silicon, in March at Arm Everywhere. The chip is purpose-built for the demands of agentic AI. According to Arm, the OCP Global Summit 2026 marked a shift from vision and launch to deployment and proof, with production systems, growing customer traction and an open agentic rack architecture on show.

In his keynote, Mohamed Awad, Executive Vice President and General Manager of Arm's Cloud AI Business Unit, described how AI agents change the way work is generated in the data center. In his words, "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."

Arm says this creates new demands across three CPU roles: AI head nodes that keep accelerators productive, cloud-native CPUs that run applications and services, and agentic CPUs that support continuously executing virtual users. DPUs add networking, storage, security and infrastructure isolation.

Open rack designs and accelerator pairings

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. Rather than prescribing one configuration, it lets infrastructure providers choose the technologies and partners that fit their workloads.

Supermicro highlighted Arm AGI CPU designs built around OCP's Open Rack v3 (ORv3) and Open Rack Wide (ORW) architectures. They include a 21-inch liquid-cooled configuration supporting 128 Arm AGI CPUs and a wider design scaling to 336 CPUs. Arm also pointed to its AGI CPU reference server work for OCP and Meta's Yosemite YV5 rack-scale design.

Two accelerator makers showed complementary approaches:

• FuriosaAI pairs Arm AGI CPU with its RNGD AI accelerators. The CPU handles agent orchestration, tool execution and supporting services, while the accelerators handle model prefill and decode, and host memory adds capacity for context.
• Rebellions showed its RebelServer architecture, which combines Arm AGI CPU with Rebel100 accelerators in a co-designed inference platform.

Arm argues that both examples underline the importance of heterogeneous computing, where performance depends on how CPUs, accelerators, memory and interconnects work together as a system.

A new way to measure agentic CPU performance

Ravi Malhotra, Director of Solutions Product Management at Arm, presented a session titled "Rethinking CPU Performance for the Agentic AI Era". He explained why traditional CPU measurement needs to evolve as AI applications move from generating responses to executing multistep tasks.

Different agents stress infrastructure differently. Software development agents need strong per-core performance for code compilation and testing. Enterprise automation and personal assistants often generate many short tool calls, so concurrency and performance per rack matter more.

Arm introduced a new methodology for measuring agentic AI CPU performance, developed with support from Fujitsu, Meta and OpenAI. It uses representative agent workloads and repeatable CPU measurements, including replaying recorded agent execution sequences. The aim, according to Arm, is to measure how efficiently infrastructure supports useful work when agents run concurrently and at scale, not only how fast a processor completes a single operation.

In analysis published by Signal65, Arm AGI CPU delivered an average of 1.99x modeled performance per rack compared with AMD EPYC 9755 across seven representative agentic AI workloads. The rack-level findings were modeled using independently verified server-level measurements.

Software readiness and what it means

Arm also stressed the software side. Developers and operators need consistent platforms that can be deployed and managed across different implementations. Arm highlighted its system architecture and software enablement work, including SystemReady, the Server Base System Architecture (SBSA), the Server Base Manageability Requirements (SBMR) and the Arm Architecture Compliance Suite.

For infrastructure buyers, the message from Arm is that agentic AI data centers can be assembled from open, repeatable designs rather than a single locked hardware configuration. Through processor IP, compute subsystems (CSS) and Arm AGI CPU, the company says it offers partners several ways to build. The performance figures come from modeled rack-level results, so real-world outcomes will depend on the workloads and system designs operators choose.