Arm's important CPU launch artifact is a benchmark method, not the 1.99x rack-performance claim attached to it. The method links representative agent work, repeatable component measurements, and completed-task outcomes so operators can challenge where a server claim stops matching a production workload.
Signal65 ran the same scripts on Arm AGI CPU and AMD EPYC systems, measured server performance, then modeled rack output under power and density constraints. Arm reports an average 1.99x more work per rack across seven workloads, 1.7x for tooling, 2.0x for orchestration, 2.2x for reasoning and planning, and up to 45% lower sandbox creation latency. The constraint being removed is benchmark ambiguity. Core counts and peak throughput hide whether CPUs keep accelerators fed, create isolated runtimes quickly, and sustain concurrent agent work inside a fixed rack budget.
Within 12 months, serious agent-infrastructure RFPs will ask for completed-task throughput and sandbox latency alongside conventional CPU metrics. Vendors that cannot expose workload definitions, scripts, power assumptions, and end-to-end validation will lose the comparison before procurement reaches price.