Table of Contents
The architect behind the world’s most popular processors is now building its own silicon
In sum – what we know:
- A business shift – Arm is moving beyond its traditional IP licensing model to manufacture and sell its own production silicon, the Arm AGI CPU, for the first time.
- Built for orchestration – The 136-core chip skips multithreading to prioritize deterministic performance and high memory bandwidth, targeting the CPU-side management of complex AI workloads.
- Aggressive claims – Arm projects the AGI CPU can deliver twice the performance per rack compared to x86 systems, potentially saving operators $10 billion per gigawatt of data center capacity.
For 35 years, Arm has essentially stuck to deigning the processors, and letting others build them. That, however, could be changing — the Cambridge-based chip architect has taken the wraps off the Arm AGI CPU. It’s the first production silicon the company has ever manufactured and sold under its own name.
The change is a big one for Arm. Arm built one of the most successful businesses in semiconductor history on pure IP licensing. Its processor designs already run in everything from smartphones to the biggest cloud instances. But designing cores that other companies turn into products is a fundamentally different game from shipping finished chips yourself.
CEO Rene Haas was clearly aware of the gravity. “AI has fundamentally redefined how computing is built and deployed. Agentic computing is accelerating that change,” he said, calling the launch “a defining moment for our company” and the “next evolution of the Arm compute platform.” Meta signed on as both the lead development partner and the first customer.
The tech
Underneath the branding, the Arm AGI CPU is a genuinely ambitious piece of silicon. It offers up to 136 Neoverse V3 CPU cores in a two-die design, fabricated on TSMC’s 3-nanometer process. Clock speeds push as high as 3.7 GHz with a 3.2 GHz base, within a 300-watt thermal design power envelope.
Arm opted for a single thread per core, skipping simultaneous multithreading entirely. Mohamed Awad, EVP of Arm’s Cloud AI Business Unit, described the chip as “a clean sheet design” built around the demands of agentic workloads, emphasizing the determinism advantages this approach provides over competing designs. The logic makes sense — single-threaded cores deliver more predictable performance scaling, which matters a lot more when you’re coordinating thousands of tasks across a data center than when you’re trying to wring peak throughput from individual cores.
The memory subsystem is just as aggressive. Every core gets 2 MB of L2 cache, with 128 MB of shared system-level cache on top of that. The chip supports 12 channels of DDR5 at up to 8800 MT/s, pushing 825 GB/s of aggregate bandwidth (roughly 6 GB/s per core) with sub-100 nanosecond latency. That level of memory bandwidth is essential for the workloads Arm has in its crosshairs, where data movement bottlenecks will absolutely crush performance no matter how fast the compute cores themselves happen to be.
Built for AI
To be clear, the Arm AGI CPU isn’t designed to train or run AI models — GPUs and specialized accelerators still own that job. What this chip targets is the CPU-side orchestration layer that holds large-scale AI deployments together. These include things like the data movement, accelerator coordination, and scheduling work that keeps everything running in concert. As agentic AI systems get more complex, with multiple models and tools operating simultaneously, the CPU overhead of managing all those interactions scales dramatically. Arm’s bet is that today’s x86 processors aren’t cut out for this coordination role, and that purpose-built silicon can unlock meaningful efficiency gains.
The competitive landscape, though, is both crowded and fast-moving. Just a week before Arm’s launch, Nvidia unveiled its Vera CPUs at GTC, going after the exact same data center CPU market. Intel and AMD are continuing to push their x86 server lines forward, and both have deep, entrenched relationships with hyperscale buyers. Arm’s pitch centers on fundamental architectural advantages in power efficiency and deterministic scaling — but Nvidia has its own ecosystem gravity through CUDA and its dominant GPU position, and the x86 incumbents benefit from decades of software compatibility and sheer customer inertia.
Meta’s role as lead development partner does give Arm’s ambitions a little more weight, though it’s worth keeping in mind that Meta has a long track record of working with multiple chip vendors at once. Being Meta’s first customer is significant, but it doesn’t necessarily mean being Meta’s only option.
Infrastructure and performance
Arm isn’t pitching this as a chip alone. The standard deployment is a 1U dual-node server running two AGI CPUs per blade, good for 272 cores per blade. In an air-cooled environment, a 36-kilowatt rack fits 30 blades and delivers 8,160 cores total.
The performance claims are, predictably, aggressive. Arm says the AGI CPU delivers over 2x the performance per rack versus the latest x86 systems, driven primarily by its class-leading memory bandwidth. The core argument is that x86 processors degrade under sustained, heavily threaded workloads as cores fight over shared resources, while the AGI CPU’s architecture holds steadier as utilization climbs. Arm projects potential CAPEX savings of up to $10 billion per gigawatt of AI data center capacity.
Chips are available to order now, with Meta expected to begin deploying at scale later in 2026. Other customers and leading original design manufacturers are committed too, including the likes of OpenAI, SK Telecom, and more. The breadth of ecosystem backing, combined with an already-committed product roadmap, signals that Arm sees this not as a one-off experiment but as a permanent new business line. Whether the market agrees is going to come down to how the silicon actually performs in production, and whether Arm can navigate the tricky politics of competing directly with its own licensees.
