Arm launches data‑center CPU
Arm unveiled its first in‑house data‑center CPU for agentic AI workloads — the Arm AGI — and named early partners including Meta, OpenAI and Nvidia, positioning Arm to grab significant server share. The chip is being pitched as roughly double the performance of comparable x86 platforms and expected to add billions to Arm's revenue, which could shift hiring and tooling around AI hardware stacks. (reuters.com)
Arm said the AGI CPU is built on TSMC’s 3nm process, uses up to 136 Neoverse V3 cores and has a 300 W thermal design power. (newsroom.arm.com) Arm disclosed a two-die design with up to 12 DDR5 memory channels (supporting speeds up to 8,800 MT/s), 2 MB of L2 cache per core and a 128 MB shared system-level cache. (theregister.com) The company specified platform I/O including support for PCIe Gen6 and CXL 3.0 and presented a reference 1U blade with two AGI CPUs (272 cores) as part of its density math. (forbes.com) Arm and Meta described a co‑development relationship with Meta as the lead partner and first customer, and Arm listed additional committed launch partners such as OpenAI, Cloudflare and SAP. (about.fb.com) Bloomberg reported Arm’s internal sales targets for the new silicon business at roughly $15 billion in annual revenue within five years and total sales near $25 billion, with TSMC named as the chip manufacturer. (bloomberg.com) Arm positioned the AGI CPU to sit alongside accelerator GPUs/ASICs rather than replace them, arguing the design’s one‑thread‑per‑core approach and high memory bandwidth are meant to handle agentic AI orchestration and data movement. (theregister.com) Company materials and press coverage showed rack-density examples: a 36 kW air‑cooled rack with 30 1U blades totaling 8,160 cores in Arm’s reference design, and a liquid‑cooled 200 kW variant supporting many tens of thousands of cores for hyperscalers. (forbes.com) Tech outlets reported that Arm’s AGI family has already committed multiple generations and that chips are intended for hyperscale deployment later in 2026, with exact commercial terms and volumes not disclosed. (techcrunch.com)
