Arm brings its first data-center CPU to market

A data-center rack validated by Arm can now hold 8,160 cores in an air-cooled design, or more than 45,000 cores in a liquid-cooled system. The processor at the center of those designs is Arm's AGI CPU, the company's first move from licensing chip blueprints to producing silicon of its own, developed with Meta for data centers.

Announced in March, the AGI CPU contains up to 136 Arm Neoverse V3 cores running at up to 3.7GHz across two dies. It is built on TSMC's 3nm process. Arm says the chip can deliver more than double the performance per rack of platforms based on Intel's x86 architecture in certain configurations, but that figure is based on internal benchmarks rather than performance in real-world deployments.

Arm has validated two Open Compute Project rack designs. One is a 36kW air-cooled system with 30 compute blades and 8,160 cores per rack. The other is a 200kW liquid-cooled server that can house 336 AGI CPUs for more than 45,000 cores. The designs are intended to make an Arm-based system available beyond the hyperscalers that can afford to build their own processors.

That is the commercial opening. Arm says a chip of this kind requires access to more than $500 million in funding, putting a custom design beyond much of the market. Meta will be the first company to deploy the AGI CPU, while AI cloud provider Verda plans to use it in its data halls in Finland and Iceland, although it has not set a timetable. The chip was designed around Meta's needs but not as a Meta-only processor.

Practically, this gives data-center operators another route to Arm-based computing without assembling a full silicon team, while keeping room for Nvidia hardware: Arm says the AGI CPU, Nvidia's Grace processor and GB300 are all compliant with its System Ready standards. Arm's rack figures are validated designs, and the headline performance comparison remains a company benchmark.