Arm chips for Codex
Arm has started shipping in‑house AI CPUs and OpenAI is listed among early customers aiming to run Codex and other workloads on lower‑latency, energy‑efficient Arm silicon. The shift signals vendor moves to cut inference latency and operating costs for code‑generation workloads. ( )
Arm formally announced the Arm AGI CPU on March 24, 2026 — the first production silicon the company has designed and shipped in its 35-year history. (newsroom.arm.com) The family runs up to 136 Arm Neoverse V3 cores per socket, is fabricated on TSMC’s 3nm node, targets up to 3.2 GHz all-core (3.7 GHz boost) and a 300 W TDP, and supports 12 DDR5 channels, 96 PCIe Gen6 lanes and native CXL 3.0. (tomshardware.com) Arm’s reference platform is a 1U, dual‑node blade that packs two AGI CPUs (272 cores) and the company’s air‑cooled 36 kW rack design holds 30 blades (8,160 cores); a Supermicro liquid‑cooled config can scale to hundreds of chips and tens of thousands of cores. (newsroom.arm.com) (tomshardware.com) Meta is the lead partner and co‑developer and Arm lists a set of early customers that includes OpenAI, Cloudflare, Cerebras, SAP and SK Telecom among committed users. (cnbc.com) (wired.com) OpenAI recently released a low‑latency Codex variant, GPT‑5.3‑Codex‑Spark, running on Cerebras accelerators in February 2026; OpenAI is also named among Arm’s AGI CPU customers, linking the company’s low‑latency code‑generation work to Arm’s energy‑efficient CPU platform. (openai.com) (tomshardware.com) Arm says the AGI CPU delivers more than two times the performance per rack compared with current x86 platforms and is working with OEMs including ASRock Rack, Lenovo, Quanta and Supermicro to make early systems available now with broader availability expected in the second half of 2026. (markets.financialcontent.com) (forbes.com)
