Arm Launches First In-House Silicon to Power the Era of Agentic AI

Arm Holdings has officially entered the chip manufacturing market with the launch of the Arm AGI CPU, marking a historic shift in its 35-year business strategy. For the first time, the company will offer production silicon alongside its traditional intellectual property licenses. This new strategy addresses the rising demand for hardware that can handle autonomous AI agents at global scale. Developed in close collaboration with Meta, this new processor is engineered specifically for agentic AI infrastructure. These autonomous agents require intense reasoning and planning capabilities that traditional data center architectures struggle to support efficiently. The Arm AGI CPU aims to provide the foundation for this next era of computing. The technical specifications show a chip built for pure performance. Fabricated on the TSMC 3nm process, the CPU features up to 136 Neoverse V3 cores. It provides massive memory bandwidth of 6 GB/s per core with sub-100ns latency. By dedicating one physical core per thread, the design ensures deterministic performance without the throttling found in legacy systems.

When compared to current x86 platforms, the AGI CPU offers more than double the performance per rack. This efficiency allows data center operators to achieve unprecedented core density. For example, liquid-cooled systems can support over 45,000 cores per rack using this architecture. This is a significant jump over existing server standards.

The economic impact for hyperscalers is equally substantial. Arm estimates that its new silicon can reduce capital expenditures by up to $10 billion per gigawatt of data center capacity. This cost-to-performance ratio makes it a direct threat to established incumbents in the server market.

While Meta is the lead partner, other industry giants including OpenAI and Google Cloud have already signaled their support. Early systems from manufacturers like Supermicro and Lenovo are available now. Broader commercial availability is expected throughout the second half of 2026.