Arm Launches AGI CPU With Meta for AI Data Centers

Arm launched the Arm AGI CPU on March 24, 2026, its first production silicon product, and Meta is the lead partner and co-developer for deployment in AI data centers later this year. For developers and infrastructure teams building large-scale AI inference systems, the important shift is not just a new server CPU. Arm is moving from licensor to chip vendor, with a design tuned for orchestration, data movement, and control-plane work around accelerators.

Product positioning

The Arm AGI CPU targets AI data centers, especially agentic and inference-heavy environments where CPUs spend more time coordinating work than performing the core tensor math. Arm positions the chip for reasoning support, accelerator management, task coordination, API hosting, and movement of data through heterogeneous systems.

Meta is central to that design target. Meta says it is partnering with Arm on a new class of CPUs for data centers and large-scale AI deployments, with a roadmap spanning multiple generations. In Meta’s deployment model, the CPU sits alongside MTIA to improve orchestration efficiency across training and inference clusters.

If you build AI agents or operate multi-step systems with heavy tool use, this matters because those workloads increase the amount of non-GPU work per request. Agent loops, retrieval calls, scheduling, memory access, policy checks, and service-to-service coordination all hit the CPU path. Arm is explicitly designing for that shift.

Hardware specifications

Arm disclosed a relatively detailed first-pass spec sheet for the launch.

Spec	Arm AGI CPU
Core architecture	Arm Neoverse V3
Max cores per CPU	136
TDP	300W
Memory bandwidth per core	6 GB/s
Memory latency	Sub-100 ns
Air-cooled density	Up to 8,160 cores per rack
Liquid-cooled density	45,000+ cores per rack
Manufacturing	TSMC 3nm

The density numbers are the point to watch. Arm says the chip supports high-density 1U servers and claims more than 2x performance per rack versus x86 CPUs. It also projects up to $10 billion in CAPEX savings per gigawatt of AI data center capacity.

Those claims are directional until detailed benchmark methodology appears, but the operational argument is clear. AI clusters need more CPU capacity around the accelerators, and rack-level density is becoming a first-order design constraint.

Meta’s role

Meta is not just an early customer. It is the lead partner and co-developer, and the two companies say they will work on multiple generations of these CPUs for AI data centers.

This matters because Meta already runs one of the largest internal AI infrastructure programs in the industry, including MTIA. A CPU designed with Meta’s workload mix in mind is likely optimized for the exact bottlenecks hyperscalers care about, namely orchestration overhead, memory access latency, and efficient feeding of accelerator fleets.

For teams thinking about multi-agent systems or large retrieval and tool-execution graphs, Meta’s involvement is a strong signal about where server CPU demand is moving. More agentic behavior means more scheduling and coordination work per user-visible output.

Competitive context

Arm’s move changes the structure of the server CPU market more than the benchmark landscape, at least for now. The company already powers hyperscaler CPUs such as AWS Graviton, Microsoft Azure Cobalt, Google Axion, and NVIDIA Vera through the Neoverse platform. The difference here is that Arm is now selling silicon itself.

Vendor approach	Example
Arm IP licensed to customer-designed CPU	AWS Graviton, Azure Cobalt, Google Axion, NVIDIA Vera
Arm-designed production silicon sold by Arm	Arm AGI CPU

That creates a new option for companies that want Arm-based AI infrastructure without building a custom CPU program. It also introduces a delicate channel balance, because some of Arm’s biggest ecosystem partners are also building their own Arm-based processors.

For buyers, the appeal is speed. You get a CPU product aimed at AI data center orchestration without funding a full custom silicon effort. For competitors, the pressure is on rack efficiency and CPU-to-accelerator balance, not just raw core counts.

Availability and ecosystem

Early systems are available now through partners, with broader availability in the second half of 2026. Arm named ASRock Rack, Lenovo, Quanta Computer, and Supermicro as system partners.

Arm also named additional customers with commercial intent to deploy the chip for agentic CPU use cases, including Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom. That list matters because it shows the product is not framed purely as a Meta-specific part.

If you work on production inference or on GPU-cluster deployment, this is the operational takeaway: the bottleneck around AI systems is moving outward from the accelerator. CPU orchestration, memory behavior, and rack density are becoming procurement decisions that directly affect latency, utilization, and total cluster cost.

Treat this launch as a signal to re-check your own infrastructure assumptions. If your roadmap includes tool-heavy agents, retrieval pipelines, or accelerator-dense inference clusters, model the CPU side of the system with the same rigor you apply to GPUs.