Arm has proposed a new way to measure how well data center CPUs handle AI agents, winning support from Fujitsu, Meta, and OpenAI, and released an analysis claiming its own AGI CPU delivers nearly twice the rack-level performance of AMD's EPYC 9755.
The British chip designer argues that as AI agents take on long-running jobs, the question in the data center is no longer just how fast GPUs can compute. CPUs now have to keep huge numbers of agents running at once, schedule their work, and keep expensive accelerators fed with data, and Arm says existing benchmarks fail to capture that.
AI Agents Are Piling New Work Onto Data Center CPUs
A chatbot answers a prompt once. An AI agent keeps going: it holds on to the state of a task, retrieves context, calls tools and services, checks the results, and decides what to do next. When thousands of agents do this simultaneously, compute, memory, storage, and input/output all have to support heavy concurrency and sustained operation, and the CPU absorbs much of the tool calling, data access, and task orchestration.
Inside AI head nodes, the CPU also handles job scheduling, data preparation, and data movement so that accelerators are not left idle waiting for inputs. Arm says agentic AI will not replace existing cloud-native workloads but will add a new layer of demand on top of them, from search and service calls to code execution and data requests.
The shift is happening fast. Citing data from OpenRouter, a platform that routes requests across AI models, Arm's announcement said token usage by AI agents has grown about 14 times since February 2026.
Arm's Method Tests Real Agent Tasks End To End
Arm contends that traditional CPU metrics do not fully reflect what agents need. Its proposed methodology rests on three principles: use representative agentic AI scenarios, make CPU measurements repeatable, and confirm through end-to-end results whether faster hardware actually translates into better task completion.
Component-level tests strip out noise from model inference and networking to isolate how the CPU itself performs. End-to-end tests then check whether the whole system finishes the work more efficiently. Arm said the method will keep evolving alongside agent architectures and software.
All three backers said the industry needs more representative and comparable benchmarks. Toshio Yoshida, senior vice president at Fujitsu and head of advanced computing development at Fujitsu Research, said building FUJITSU-MONAKA, the company's next-generation high-performance, energy-efficient Arm-based CPU, showed Fujitsu how much transparent, repeatable, and representative benchmarks matter, which is why it backs Arm's push for industry collaboration.
Meta software engineer Abhishek Dhanotia said that as agentic AI moves into production, the industry needs a clearer picture of how different workloads use the underlying infrastructure, and consistent comparisons can guide how those systems evolve. Colin Jarvis, OpenAI's head of global forward deployed engineering, said AI systems are moving beyond generating responses to completing complex tasks through reasoning, tool use, and interaction with software environments, making it increasingly important to understand infrastructure performance in areas such as orchestration and tool execution.
Signal65 Model Puts Arm AGI CPU At 1.99 Times AMD
According to figures released by Arm, research firm Signal65 ran the same benchmark scripts on the Arm AGI CPU and AMD's EPYC Turin processors, then used the server-level measurements to build a rack-level performance model.
Across seven workloads, the model put the Arm AGI CPU at an average of 1.99 times the performance of the AMD EPYC 9755. The gap was about 1.7 times for tool calling, about 2 times for agent workflows and task orchestration, and about 2.2 times for inference and planning.
In a separate comparison against the AMD EPYC 9965, Arm said its chip cut sandbox creation latency by up to 45%. Sandboxes are the isolated environments agents use to carry out tasks, so creating them faster lets agents start working sooner, although the real-world benefit depends on system configuration and workload.
Arm designed the AGI CPU around high core density, dedicated memory per core, and high-bandwidth I/O, aiming squarely at the always-on nature of agentic AI. At the Open Compute Project (OCP) Global Summit this week, Arm is also contributing an AGI CPU reference server design, while Meta's Yosemite YV5 brings the Arm AGI CPU into an OCP rack-scale design.
Rack Results Still Need Real World Proof
Data centers are bound by power, cooling, and floor space, so a processor's theoretical peak says little about what operators care about most: how much useful work a rack can sustain within those limits. Because Signal65's figures come from a rack-level model built on server-level tests, the 1.99 times result reflects specific test and modeling conditions. It should not be read as a gap every real-world deployment will see.
As agents move from answering prompts to finishing tasks, CPU evaluation is likely to expand from single peak scores to concurrency, sustained throughput, and task outcomes. Whether Arm's approach becomes a widely accepted industry yardstick will depend on validation across more platforms, workloads, and live deployments.
Taiwan has a stake in how that plays out. Arm Chief Executive Rene Haas told audiences at Computex in Taipei earlier this year that the company's server CPU ecosystem is built entirely in Taiwan.