If you have followed Arm for any length of time, you know the company has always been the quiet backbone of modern computing. Its technology powers everything from smartphones to servers, yet Arm has traditionally stayed one step removed from the chipmaking process itself. That long‑standing boundary shifted this week. For the first time in its history, Arm is producing its own silicon.
Arm CEO Rene Haas describes its latest announcement as a defining moment for the company. Arm is expanding its compute platform to include production silicon products, beginning with a new data center processor called the Arm AGI CPU. It is designed specifically for the emerging world of agentic AI, and Arm is positioning it as a response to a fundamental shift in how compute is being consumed.
That shift is driven by AI systems that do more than generate content. LLMs can now reason, plan, and act continuously, which dramatically increases the volume of tokens they process. According to Arm, data centers may need more than four times the CPU capacity per gigawatt to support these workloads. In that context, ARM’s move into silicon is less a departure from its licensing model and more an attempt to meet a rapidly escalating demand curve.
Introducing the Arm AGI CPU
The AGI CPU is Arm’s first Arm‑designed data center‑class processor. It is built around up to 136 Neoverse V3 cores per CPU and is engineered for high token throughput, low latency, and predictable performance under sustained load. Arm highlights several key metrics to underscore the scale of the design. Each core can access up to 6 GB per second of memory bandwidth at sub-100-nanosecond latency. The chip operates at a 300-watt TDP and assigns a dedicated core to each program thread.
Those design choices translate into dense server configurations. Arm says a 1U air‑cooled server can support up to 8,160 cores per rack, while liquid‑cooled systems can scale beyond 45,000 cores per rack. The company claims this density delivers more than twice the performance per rack compared with x86 platforms, which it estimates could reduce capital expenditures by as much as ten billion dollars per gigawatt of AI data center capacity.
Arm did not take this step alone. Meta is the lead partner and co‑developer for the AGI CPU, integrating it into its infrastructure alongside its MTIA accelerators. The collaboration reflects Meta’s need for a compute platform that can scale with its AI ambitions.
Santosh Janardhan, Meta’s head of infrastructure, frames the partnership as a way to build a more efficient data center foundation. He says Meta worked with Arm to develop a CPU that improves performance density and supports a multi‑generation roadmap for the company’s evolving AI systems. Meta’s involvement signals that hyperscalers are not just adopting Arm architectures but are willing to co‑design silicon around them.
Arm’s move into silicon is also drawing support from a wide ecosystem. The company says more than fifty organizations across cloud, hyperscale, silicon, memory, networking, and software are backing the expansion. That list includes AWS, Google, Microsoft, NVIDIA, Samsung, SK hynix, and TSMC.
Several customers beyond Meta are already planning to deploy the AGI CPU for tasks such as accelerator management, control plane processing, and cloud‑based application hosting. Those early adopters include Cloudflare, OpenAI, SAP, Cerebras, Rebellions, and SK Telecom. On the hardware side, ASRock Rack, Lenovo, Quanta, and Supermicro are preparing systems, with early units available now and broader availability expected in the second half of the year.
