Arm Holdings has launched its first physical processor in the company’s 35-year history: the AGI CPU, a 136-core data center chip designed specifically for agentic AI workloads. Meta co-developed the chip and is the anchor customer, according to Arm’s official announcement.

The move breaks Arm’s founding model. Since 1990, the company has collected royalties by licensing chip designs to manufacturers including Apple, Nvidia, Amazon, and Google. It has never sold its own silicon until now.

Technical Specifications

The AGI CPU packs up to 136 Arm Neoverse V3 cores per processor, manufactured on TSMC’s 3-nanometer process. Key specifications from the Arm press release:

  • 300-watt TDP with a dedicated core per program thread, eliminating throttling and idle threads
  • 6GB/s memory bandwidth per core at sub-100 nanosecond latency
  • Supports air-cooled 1U server chassis with up to 8,160 cores per rack
  • Liquid-cooled configurations deliver over 45,000 cores per rack

Arm claims more than 2x performance per rack compared with x86 CPUs from Intel and AMD, and projects up to $10 billion in CAPEX savings per gigawatt of AI data center capacity, according to the same press release.

The chip was developed over approximately 18 months at a purpose-built $71 million laboratory in Austin, Texas, Proactive Investors reported.

Customer Commitments

Meta’s head of infrastructure, Santosh Janardhan, said the company “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,” per Arm’s announcement. Meta will deploy it alongside its in-house MTIA AI accelerators.

Beyond Meta, Arm confirmed commercial commitments from Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom for use cases including accelerator management, control plane processing, and API and application hosting, per the same announcement.

On the hardware side, OEMs and ODMs including ASRock Rack, Lenovo, Quanta Computer, and Supermicro are building systems. Early units are available now, with broader availability expected in the second half of 2026, TrendForce reported.

Over 50 companies across hyperscale, cloud, silicon, memory, networking, and software are supporting the platform, including AWS, Broadcom, Google, Marvell, Micron, Microsoft, Nvidia, Samsung, SK hynix, and TSMC, per Arm’s press release.

Revenue Ambitions

Bloomberg projects $15 billion in annual revenue from the chip business within five years, as cited by TrendForce. Arm shares rallied on the announcement.

Bloomberg also noted that Arm’s entry into chip manufacturing could complicate its relationships with licensees, many of whom run in-house chip programs of their own, per the same TrendForce report.

Why CPUs Are the Bottleneck Now

The timing ties to a structural shift in AI infrastructure. As AI moves from training large models to running continuously active agents that reason, plan, and execute tasks, the workload profile shifts from GPU-heavy training to CPU-heavy inference, coordination, and data movement.

Arm estimates that CPU compute demand per gigawatt of data center capacity will increase more than fourfold as agent-driven applications scale, per its press release. Nvidia recently told CNBC that CPUs are “becoming the bottleneck” as agentic AI reshapes compute requirements, as cited by TrendForce.

For the agentic ecosystem, this is infrastructure-level validation: a company that has spent 35 years collecting royalties decided the agent workload opportunity was large enough to risk its core business model and build silicon for it.