Arm ships AI data‑center CPU
Arm unveiled its first production AI data‑center CPU built for agentic workloads—high core counts on a 3nm process—marking a serious push beyond mobile into heterogeneous, energy‑efficient server computing. Analysts say this expands the hardware ecosystem away from x86/GPU dominance and will influence software patterns for AI at scale. (indexbox.io) (counterpointresearch.com)
Arm announced the Arm AGI CPU on March 24, 2026, marking its first production silicon release after more than 35 years as primarily an IP licensor. (newsroom.arm.com) Meta is the lead partner and first customer on the AGI CPU program, with Arm saying Meta helped co-develop the design. (cnbc.com) Arm also named commercial commitments from companies including OpenAI, Cloudflare, Cerebras, SAP and SK Telecom, and said multiple OEM/ODM partners will ship systems. (techpowerup.com) The chip packs up to 136 Arm Neoverse V3 cores, runs at up to 3.7 GHz (3.2 GHz base) and is specified with a 300 W TDP. (theregister.com) Arm publishes the AGI CPU’s I/O and memory specs as 12 DDR5 channels supporting up to DDR5‑8800 (about 6 GB/s per core), 2 MB L2 per core plus a 128 MB system-level cache, and 96 PCIe Gen6 lanes with CXL support. (eetimes.com) Design choices include a one‑thread‑per‑core model (no SMT) and a conscious decision to avoid on‑die accelerators in order to maximize usable area and deterministic sustained throughput for agent orchestration workloads. (theregister.com) Arm’s reference 1OU, 2‑node server puts two AGI CPUs and 272 cores per blade, with a standard 36 kW air‑cooled rack holding 30 blades (8,160 cores) and a Supermicro liquid‑cooled 200 kW design housing 336 chips (>45,000 cores); Arm says early system units are available now with broader availability expected in the second half of 2026. (newsroom.arm.com)
