Arm's AGI CPU Claims 1.99x AMD Rack Performance, With a Catch

What Arm actually measured

The headline number is 1.99x, but it is not a straight chip-to-chip comparison.

Signal65 ran the assessment with Arm's cooperation, measuring individual processors first and then estimating how full racks would perform inside a fixed power budget. The workloads were chosen to reflect agentic AI: software tool use, autonomous task execution, information retrieval and model serving, including Terminal-Bench 2 and SWE-bench.

Across seven tests, Arm's AGI CPU averaged 1.99x the rack-level performance of AMD's EPYC 9755. Individual results ranged from 1.43x to 3.19x. By category, Arm averaged 1.7x for tooling, 2.0x for agentic workflows and 2.2x for reasoning and planning.

The report also estimated that database and vector-search operations could need 30% to 35% fewer racks for equivalent work. That is the kind of number data center operators care about, because floor space and power contracts are often harder to secure than chips.

The power budget is doing much of the work

Arm's advantage is less about raw per-core speed than about how many servers you can squeeze into a rack.

A 24 kW rack can hold 21 to 23 Arm AGI CPU servers, according to the assessment's assumptions. The same rack fits 13 dual-socket EPYC 9755 systems. That translates to between 5,712 and 5,888 physical Arm cores against 3,328 AMD cores.

The gap comes from power draw. Arm's individual servers are rated at 950 to 1,050 watts, while AMD's dual-socket configurations are rated at 1,700 watts. More servers per rack means more aggregate throughput, even if each Arm core is not faster than an AMD core.

This is a real advantage in a world where utilities and landlords, not chip vendors, set the ceiling on AI capacity. It is also a fragile one: change the power assumptions and the ratio moves.

Agent startups are the real benchmark

The separate Sandbench test measures something AI agents do constantly: spinning up isolated execution environments.

Against AMD's EPYC 9965, Arm completed cold starts 45% faster and warm starts 27% faster. Under fourfold oversubscription, its worst case was still 41% faster. For agents that repeatedly launch sandboxes to run code, query services and process results, that setup time adds up across thousands of tasks.

Mohamed Awad, Arm's executive vice president for Cloud AI, said the industry needs to measure how well systems handle the work agents actually complete. Arm developed its evaluation framework with input from Fujitsu, Meta and OpenAI, while Signal65 independently verified the processor results.

That framing is deliberate. Arm is not claiming victory in general-purpose computing, where x86 still dominates. It is staking a claim to the agentic AI niche, where workload shape and rack density matter more than legacy software compatibility. The same logic is driving Microsoft's Maia 300, which bets on memory bandwidth rather than raw compute to cut AI costs.

What could go wrong for Arm

Plenty, starting with the fact that none of this has been validated on final production hardware.

Arm itself acknowledged that further testing on production systems will cover additional workloads and competitive comparisons. Modelled rack performance is a projection, not a shipped result. Until independent buyers run these workloads on real machines, the 1.99x figure is a hypothesis with a good slide deck.

Software compatibility is the other hurdle. Agentic AI stacks are young and messy, but many enterprises still run x86-tuned tooling, and switching architectures means recompiling, retesting and retraining teams. Deployment costs and real-world operating conditions will decide whether the modelled advantage survives contact with a data center.

AMD and Intel have heard this pitch before. What is different now is that the buyer is often an AI lab or cloud provider building greenfield capacity, where there is no legacy estate to protect and power is the binding constraint.

Availability and pricing

Arm has not announced retail pricing, general availability dates or specific cloud partners for the AGI CPU. The assessment covers engineering samples and modelled configurations, so US, UK and EU buyers cannot order systems yet.

When Arm-based servers do reach the market, expect them to arrive through OEMs and cloud providers rather than direct sales. Enterprises evaluating agentic AI infrastructure should treat the 1.99x figure as a reason to run their own proof of concept, not as a procurement decision.

Arm has made a credible case that rack-level power efficiency, not per-core speed, is the metric that matters for agentic AI. The 1.99x figure is striking, but it is a projection built on assumptions Arm helped set. The next twelve months of production benchmarks will tell us whether this is a turning point for x86 or a very good marketing document.