Arm AGI CPU Takes 136 Neoverse Cores Into AI Servers as Arm Sells Its Own Silicon

As of July 8, 2026, Arm’s AGI CPU is no longer just a launch-day curiosity from March. It is one of the clearer signs that AI server design is moving beyond the usual GPU headline race. The interesting part is not simply that Arm has a processor with up to 136 cores; it is that Arm is now selling a production data-center CPU itself, rather than only licensing CPU IP and platform designs to other chipmakers. Tom’s Hardware described the March 24 announcement as Arm’s first move into selling finished production chips, while Arm’s own product brief positions AGI CPU as silicon for AI-native data centers handling inference, agent orchestration, accelerator coordination, and parallel fan-out tasks. (tomshardware.com)

The top AGI CPU configuration brings 136 Arm Neoverse V3 cores on the Armv9.2 architecture, with dual 128-bit SVE2 units per core and support for bfloat16 and INT8 AI instructions. Arm lists three main variants in its product brief: a 136-core max-core-count SKU, a 128-core TCO-optimized SKU, and a 64-core max-memory-per-core SKU. The 136-core and 128-core versions list a 3.5GHz maximum frequency, while the 64-core model reaches 3.7GHz; Arm’s broader highlights also refer to up to 3.2GHz all-core operation and boost speeds up to 3.7GHz. Key platform specs include:

• Process: TSMC 3nm, according to reports covering the launch
• Cores: up to 136 Neoverse V3 cores
• Cache: 2MB L2 per core and 128MB system-level cache
• Memory: 12-channel DDR5 RDIMM support up to 8800MT/s
• I/O: 96 lanes of PCIe Gen6 plus CXL 3.0 Type 3 support
• TDP: 300W base TDP, listed as a preset value within a configurable range
• Socket support: 2-socket configurations supported

AI clusters are often described through their accelerators, but the CPU layer still has plenty to do. In agentic AI systems, CPUs can manage scheduling, memory movement, network-facing services, storage coordination, pre- and post-processing, and the control logic that keeps accelerators fed with useful work. Arm’s investor material frames this split directly: CPUs orchestrate while accelerators generate tokens. That explains why the AGI CPU leans heavily on memory bandwidth, low-latency access, PCIe Gen6, and CXL 3.0 rather than only chasing a larger core count. Arm claims up to 6GB/s of memory bandwidth per core on the 136-core model, with sub-100ns latency targets, but readers should treat rack-level performance claims as vendor estimates until more independent testing is available. (investors.arm.com)

View NVIDIA DGX Spark Personal AI Desktop Supercomputer on partner website

Meta was the lead partner for the AGI CPU, and the processor was co-developed with Meta for data-center use around agentic AI infrastructure. Data Center Dynamics reported that Meta plans to deploy AGI CPUs alongside its own MTIA accelerator silicon, while Arm said the companies are aligned on a multi-generation roadmap. Arm also named other commercial partners around the launch, including Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom. That partner list matters because this is not being presented as a lab-only chip; it is aimed at production server platforms where software stacks, validation, rack design, and accelerator connectivity are as important as the CPU die itself. (datacenterdynamics.com)

Buy SANDISK Optimus GX PRO 8100 PCIe 5 SSD here

Arm’s product brief includes OCP-aligned reference platforms, including a 1U2N Arm AGI CPU 1OU dual-node reference server and a 2U2P reference server. Arm also describes rack-scale deployment targets, including up to 8,160 cores in a standard 36kW air-cooled rack. Data Center Dynamics reported that early AGI CPU systems were available around launch coverage, with broader availability expected in the second half of 2026. The overall direction is clear: Arm is trying to give hyperscalers and AI infrastructure builders a ready silicon path that sits beside its existing IP licensing and Neoverse CSS programs, rather than replacing them outright. For readers watching AI server hardware in 2026, the AGI CPU is important because it pulls CPUs, memory bandwidth, CXL, and rack density into the same conversation as accelerators. (arm.com)

See Crucial T705 13 GB/s NVMe SSD price