Arm introduced its first 136-core server AGI CPU for agentic AI

Arm on March 24 presented AGI CPU — its first serial processor and first proprietary chip for data centers. The company, which for decades earned through architecture licensing, now enters the market for ready-made server microchips for agentic AI infrastructure.

A New Stage for Arm

For Arm, this is not simply another release in the Neoverse lineup. For the first time in over 35 years, the company is selling not IP blocks and not compute subsystems to partners, but a finished processor of its own design. The logic is clear: in the era of large AI clusters, CPU becomes a critical system element again, even if accelerators perform the main computational work.

The processor distributes tasks, manages memory and storage, coordinates network exchange, and ensures that thousands of parallel operations don't become a bottleneck. Arm calls the target niche "agentic AI infrastructure." This refers to an environment where multiple software agents simultaneously plan actions, invoke models, exchange data, and launch long chains of tasks without human participation.

In such a scenario, not only peak power matters, but also the ability to stably sustain load at the rack and entire data center scale. Against this backdrop, Arm directly opposes AGI CPU to traditional x86 systems, which it considers too complex and not particularly efficient for this new class of server workloads.

What the Chip Can Do

AGI CPU is manufactured at the 3-nm process and in maximum configuration received up to 136 Arm Neoverse V3 cores. The chip is designed for sustained operation under high load: Arm claims 300W TDP and emphasizes that each core runs a dedicated software thread. According to the company, this helps avoid performance drops and idle time during long parallel tasks typical of AI orchestration.

• Up to 136 Arm Neoverse V3 cores and 3-nm process
• Up to 6 TB of memory per chip and DDR5-8800 support
• Memory bandwidth of 6 GB/s per core with latency under 100 ns
• 96 PCIe Gen 6 lanes, plus CXL 3.0 and AMBA CHI support
• Up to 8,160 cores per rack with air cooling and over 45,000 cores with liquid cooling

Special emphasis is placed on deployment density. Arm's reference server configuration is a dual-node 1OU blade with 272 cores, which can be packed into standard racks with very high density. The company claims that in such a scheme, AGI CPU can deliver more than twice the per-rack performance compared to current x86 servers. An even bolder claim concerns economics: Arm estimates potential savings for AI data centers up to $10 billion per gigawatt of power.

Who is Already in the Ecosystem

Meta became the chief partner and co-developer of AGI CPU: it plans to use the new processors together with its MTIA accelerators for large-scale AI orchestration within its own platforms. Among confirmed customers and launch partners are also OpenAI, Cloudflare, Cerebras, Rebellions, SAP, and SK Telecom. Ready-made systems based on the new chip are already being made by ASRock Rack, Lenovo, Quanta, and Supermicro, and commercial servers from some vendors are already available for order.

"Arm AGI CPU will play an important role in our infrastructure as we scale, strengthening the orchestration layer that coordinates large-scale AI workloads," said

Sachin Katti, head of Industrial Compute at OpenAI.

Additionally, Arm speaks of support from more than 50 ecosystem companies, including AWS, Google, Microsoft, Nvidia, Samsung, TSMC, and Micron. This matters because a good chip alone doesn't solve the deployment problem: the market needs compatible server platforms, firmware, diagnostic tools, and a clear path to integration with Open Compute and existing cloud infrastructure. First samples are already available, and wider adoption of Arm systems is expected in the second half of 2026.

What This Means

Arm is moving up the value chain and trying to take the place not only of architecture supplier but also a full-fledged player in the AI hardware market. If claims regarding performance and economics prove true in practice, AI cluster operators will have another serious alternative to x86 for the management and orchestration layer around accelerators.