Arm developing custom CPU for OpenAI's in-house accelerator (www.tomshardware.com)

OpenAI is reportedly partnering with SoftBank-owned Arm to design a server-class CPU that will anchor racks built around the custom inference accelerator it is co-developing with Broadcom. The Broadcom SoC, fabricated by TSMC and in development for ~18 months, is aimed at inference workloads and is expected to enter production in late 2026 with plans to scale to roughly 10 GW of compute between 2026–2029. Arm’s role goes beyond licensing: it’s taking on full CPU design and possibly manufacturing, marking a major push into the data-center market and giving OpenAI a vertically integrated stack (CPU + accelerator) that could also be paired with Nvidia and AMD systems. The move is significant because it could materially change economics and supply dynamics in AI infrastructure. OpenAI’s chip programs now total up to 26 GW of planned capacity and, if fully deployed alongside Nvidia/AMD purchases, could imply more than $1 trillion in build and equipment costs. For Arm and SoftBank, the contract could generate billions in revenue and accelerate Arm’s server ambitions. For the AI ecosystem, a successful Broadcom/TSMC accelerator plus Arm CPUs would offer OpenAI leverage against Nvidia’s H100/Blackwell GPU dominance, provide a hedge against GPU shortages, and push more system-level co-design (chips, CPUs, racks) into mainstream AI deployments.