A new CPU enters AI data centers, built for agent-based workloads. It aims to handle rising compute demand while fitting within power limits.
Arm has introduced the Arm AGI CPU, its first move into producing silicon, targeting AI data centers that run agent-based workloads. The processor is designed to handle rising compute demand as AI systems shift from model training to reasoning, planning, and execution.
The CPU delivers up to 136 Arm Neoverse V3 cores, 6GB/s memory bandwidth per core, and sub-100ns latency. It operates at 300W TDP and assigns one core per thread to maintain performance. The design supports dense deployments, with air-cooled racks scaling to 8,160 cores and liquid-cooled systems exceeding 45,000 cores per rack. This increases workload density and improves use of accelerators within the same power limits. Arm estimates more than 2x performance per rack compared to x86 systems and potential savings of up to $10 billion per gigawatt of AI data center capacity.
The launch follows a shift in computing needs. Agent-driven AI systems generate a high volume of tokens and require more CPUs for coordination and data movement. As these workloads scale, data centers are expected to need over four times the current CPU capacity per gigawatt, increasing the need to deliver more compute within fixed power limits.
To support this, Arm is expanding beyond its IP and Compute Subsystems business into chip production. Partners can choose between licensing Arm designs, using subsystems, or deploying Arm-built silicon.
Meta Platforms is the lead partner and co-developer, using the CPU with its infrastructure, including its training and inference accelerator. Other companies such as Cerebras Systems, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom plan to use the processor for control plane tasks, accelerator management, and enterprise workloads.
Hardware partners including ASRock Rack, Lenovo, Quanta Computer, and Supermicro are building systems based on the chip. Early systems are available, with wider deployment expected later this year. Support across the ecosystem includes Amazon Web Services, Broadcom, Google, Marvell Technology, Micron Technology, Microsoft, NVIDIA, Samsung Electronics, SK hynix, and TSMC.
This marks a shift in Arm’s role in the data center market, moving from enabling designs to delivering processors for AI infrastructure.
Click here for original announcement.
