Arm presents AI-agent infrastructure with partner rack designs supporting up to 336 CPUs

Arm presented a modular platform for running AI agents at the 2026 Open Compute Project Global Summit in San Jose, with partner rack designs supporting up to 336 CPUs. Its summit presentation shifted attention from the processor alone to how computing, memory, accelerators and cooling fit together—and how that combined infrastructure should be measured.

In its October 15 account, Arm described deployment work around the AGI CPU, its first production silicon, introduced in March. The summit package included partner systems, reusable server designs and software standards intended to make different implementations easier to deploy and manage.

Different racks, rather than one fixed configuration

Arm calls the architecture its Open Agentic Infrastructure Platform. It combines general-purpose computing, AI acceleration, memory, networking, input/output connections and system management in modules aligned with Open Compute Project standards. Rather than specifying a single machine, the approach lets infrastructure providers choose technologies and partners to suit their workloads.

• Supermicro showed a 21-inch liquid-cooled configuration supporting 128 Arm AGI CPUs, built around the OCP rack-design approach.
• A wider Supermicro design scales to 336 CPUs. Its designs draw on OCP’s Open Rack v3 and Open Rack Wide architectures.
• Arm and Meta also presented work on an AGI CPU reference server for OCP and Meta’s Yosemite YV5 rack-scale design, adding reusable infrastructure around the processor.

Arm says the different configurations allow builders to optimize computing density, power delivery, cooling and servicing together. That is the practical purpose of offering multiple rack layouts: the platform does not require every data center to use the same hardware configuration to build around the CPU.

Splitting the work between processors

FuriosaAI demonstrated how the platform can divide an agent’s work. In its pairing of Arm AGI CPU and RNGD accelerators, the CPU coordinates agents, executes tools and runs supporting services. The accelerators handle model processing, while host memory provides capacity for context and other data-intensive operations.

Rebellions showed another implementation: its RebelServer architecture combines Arm AGI CPU with Rebel100 accelerators in a jointly designed system for running AI models. Both demonstrations put the CPU alongside specialized hardware, rather than treating a faster general-purpose processor as the answer to every part of an AI workload.

Software consistency is another part of Arm’s deployment pitch. It highlighted SystemReady and server architecture, management and compliance standards. Arm says these establish shared foundations for hardware, firmware, security and management, reducing the need for software vendors to qualify and customize products for each individual server design.

The rack advantage depends on density

Supporting performance evidence comes from Signal65’s October 12 analysis. Across seven CPU workloads spanning agent execution, tools, retrieval and model-serving support, it found an average 1.99× modeled performance per rack against AMD EPYC 9755. Individual workload results ranged from 1.43× to 3.19×. Signal65 says it worked with Arm on the assessment.

Those are not measurements of complete racks. Signal65 combined measured per-chip performance with Arm’s rated server power to calculate rack-level results. Its 24-kilowatt model accommodates 21–23 AGI CPU servers, compared with 13 EPYC 9755 servers. The assumed power per two-socket server is 950–1,050 watts for Arm and 1,700 watts for AMD.

Arm also introduced a measurement methodology supported by Fujitsu, Meta and OpenAI. It uses representative workloads and replays recorded agent execution sequences. Arm distinguishes code-compilation workloads, which need strong performance per core, from large volumes of short tool calls, where simultaneous activity and rack throughput matter more.

Signal65 describes its results as reflecting early AGI CPU systems. It plans follow-on testing using final production hardware, a broader workload set and additional competitive comparisons. That next round is intended to extend the evidence beyond the systems and seven workloads assessed so far.