OpenAI and Broadcom announce 10-gigawatt custom AI chip partnership

OpenAI and Broadcom announced a plan to co-develop and deploy 10 gigawatts of custom AI accelerators designed by OpenAI, with Broadcom leading manufacturing, systems development and rollout. The companies said deployment would begin in the second half of 2026 and continue through the end of 2029, and that they had already been working together for more than 18 months before the announcement.

The division of labour distinguished this deal from OpenAI’s other recent hardware agreements. Where the NVIDIA letter of intent and the AMD partnership committed OpenAI to buying GPUs designed by its suppliers, the Broadcom arrangement had OpenAI designing the accelerator and rack architecture itself — drawing, the companies said, on what OpenAI had learned building and operating its own models and products — while Broadcom built the chips and networked them using its Ethernet-based connectivity. It placed OpenAI alongside Google, Amazon and Meta as a hyperscaler-scale developer of custom AI silicon rather than solely a buyer of merchant chips from NVIDIA and AMD.

The announcement was the third multi-gigawatt compute commitment OpenAI had made in three weeks, following the NVIDIA and AMD deals, and it pushed the combined scale of publicly announced future capacity into the tens of gigawatts without a correspondingly public account of how the underlying capital expenditure — plausibly hundreds of billions of dollars across the three deals — would be financed. That gap, and the overlapping equity and supply arrangements binding OpenAI to its chip partners, continued to feed the broader circular-financing debate that had followed the NVIDIA and AMD announcements.