For 35 years, Arm’s business was architecture. They designed instruction sets and microarchitectures, charged royalties, and let everyone else make the chips. Apple, Qualcomm, AWS, Google, Microsoft – they all licensed Arm IP and competed on their own silicon. Arm stayed upstream, took a cut, and never competed with its customers.
That model just changed.
Arm has announced the Arm AGI CPU – its first production data center CPU, designed by Arm and available to order directly from Lenovo, Supermicro, and ASRockRack. Not licensed to a partner. Not the architecture underlying someone else’s chip. Arm’s own silicon, with Arm’s name on the purchase order.
The technical foundation is Neoverse V3 – the same core generation underpinning AWS Graviton4, Google Axion, and Azure Cobalt. The reference configuration runs 272 cores per 1OU, 2-node blade. The high-density Supermicro liquid-cooled variant fits 336 Arm AGI CPUs into a 200kW rack, which works out to over 45,000 cores. Arm claims 2x the performance per rack versus current x86 systems, with memory bandwidth as the primary differentiator – Neoverse V3 cores hold up better than legacy architectures under the kind of sustained parallel load that agentic AI workloads generate.
The lead co-developer and launch customer is Meta. That detail matters. Meta is not a company that buys commodity silicon – they build their own. MTIA, their custom AI accelerator, is exactly the kind of in-house silicon investment that traditionally would have included a CPU layer built from scratch too. Instead, Meta collaborated with Arm on the AGI CPU and is deploying it at what they describe as gigawatt scale alongside MTIA.
Santosh Janardhan, Head of Infrastructure at Meta, described the collaboration: “We worked alongside Arm to develop the Arm AGI CPU to deploy an efficient compute platform that significantly improves our data center performance density and supports a multi-generation roadmap for our evolving AI systems.”
The fact that Meta chose to co-develop rather than build from scratch is significant. It tells you something about where the effort is worth spending. Custom accelerators for matrix multiplications at scale – that’s differentiated work. CPU architecture for orchestration at scale – apparently not differentiated enough to justify rebuilding from zero when Neoverse V3 already exists and Arm is willing to collaborate.
Arm CEO Rene Haas framed the strategic intent directly: “Our collaboration with Meta to co-develop the Arm AGI CPU reflects the next phase of the Arm compute platform – expanding into delivering production silicon CPUs optimized for large-scale agentic AI deployments.”
The thesis underneath the launch is worth engaging with. The GPU-dominated infrastructure narrative of the last few years has treated the CPU as plumbing – present, necessary, but not the constraint. Arm is arguing that agentic AI inverts this. When your workloads are agents fan-out orchestrating thousands of parallel subtasks – managing tool calls, coordinating subagents, handling inference routing – the CPU is back at the bottleneck. It’s handling orchestration logic, memory operations, context management. GPU utilisation depends on how efficiently the CPU feeds it. Under that model, the CPU is not plumbing. It’s the pacing element.
That argument is defensible. It’s also a business argument. Arm moving from IP licensor to silicon supplier puts them directly in the value chain at a point where the margins are better than royalties.
Other launch partners include Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom. The hardware is orderable today.
If you’re planning AI infrastructure capex and your server CPU shortlist starts and ends with x86, there is now a production-ready alternative with Meta’s fingerprints on it and Arm’s warranty behind it. The question of whether it fits your workload is worth asking before the next procurement cycle, not after.
Arm just decided 35 years of staying upstream was enough.