Arm claims its AGI CPU delivers nearly 2x the per-rack performance of AMD's EPYC 9755 in agentic AI tests

The competition in data center CPUs is shifting from how fast a single chip runs to how much agentic work a rack can deliver. According to Tom's Hardware, Arm says its AGI CPU delivers nearly twice the performance per rack of AMD's EPYC 9755 in new tests built around agentic AI.

The framing matters: these are Arm's numbers. The report attributes the claim to Arm itself, which means the test setup, workload selection and comparison baseline all come from the chip vendor. No independent lab or cloud provider has yet published a reproduction.

The per-rack metric deserves scrutiny. How many servers fit in a rack, how many CPUs each server carries and what power envelope the rack allows all change the result directly. With power and cooling increasingly the binding constraint in data centers, vendors prefer to talk about performance per rack, or even per watt, rather than isolating a single processor's benchmark score.

Agentic AI is a deliberate choice of battleground. Agent workloads involve constant orchestration, tool calls, state maintenance and bursts of short requests, so the CPU is repeatedly moving and scheduling data between model serving, memory and the network. Such workloads tend to depend less on single-core peak throughput than on core count, memory bandwidth and I/O scheduling efficiency, which gives rival architectures room to reposition.

For AMD, being selected as the comparison point shows the fight has moved into the heart of x86 server CPUs. For buyers, peak performance is only one variable among several, alongside system cost, power budget, software ecosystem maturity and supply timelines.

Three things are worth watching: whether independent outlets or cloud providers reproduce the per-rack comparison, whether Arm publishes the full test methodology and workload details, and whether agentic workloads can actually sustain the assumed density in production. Until third-party data arrives, this reads as an architecture argument backed by specific figures rather than a settled performance verdict.