Overview

• OpenAI will design the accelerators and processors, with Broadcom manufacturing and integrating full Ethernet-based racks tailored to OpenAI workloads.
• Engineering begins immediately, with first racks expected in the second half of 2026 and large-scale deployment projected to continue through 2029.
• Company leaders say custom silicon will embed model learnings to boost performance and lower compute costs while giving OpenAI greater control over its infrastructure stack.
• Financial terms were not disclosed, but Broadcom shares jumped about 9.5% after the announcement, regaining the No. 7 global market-value ranking as Nvidia and TSMC also advanced.
• The collaboration complements other commitments, including roughly 6 GW with AMD and major NVIDIA arrangements, as OpenAI scales toward its long-term capacity target.