Broadcom stock soars 10% on OpenAI custom chip deal, adding to Nvidia and AMD agreements (www.cnbc.com)

OpenAI and Broadcom revealed an 18‑month collaboration to co‑design racks of custom AI inference chips and networking, a program that will deploy an initial 10 gigawatts of accelerators built on Broadcom’s Ethernet stack. Broadcom shares jumped more than 10% premarket on the news (the company’s stock is up ~40% this year and its market cap tops $1.5T), though financial terms weren’t disclosed. The deal joins a flurry of recent OpenAI compute commitments—roughly 33 GW total announced across Nvidia, Oracle, AMD and now Broadcom—and OpenAI plans to start rolling out the Broadcom‑co‑designed racks late next year. Technically, the systems combine custom compute, memory and high‑bandwidth networking optimized for OpenAI’s inference workloads; OpenAI even used its own models to drive chip layout and area reductions. Industry estimates peg a 1‑GW facility at roughly $50B (about $35B of that for chips), so custom silicon and integrated networking promise large efficiency and cost gains—enabling denser, cheaper, and faster models. The move highlights greater vertical integration in AI infrastructure (Broadcom’s XPUs and bespoke stacks vs. off‑the‑shelf GPUs), accelerates hyperscaler competition for frontier compute, and signals that OpenAI expects demand to scale well beyond its current ~2 GW baseline.