Key Takeaway
OpenAI and Broadcom are set to launch their first custom AI accelerators in late 2026, with full deployment expected by 2029. These accelerators will provide significant computing power, estimated at 10GW—enough to power around 7 million homes in the US. Designed for the mathematical operations essential to AI models, the chips will be utilized in OpenAI’s facilities and partner data centers globally. This initiative reflects a broader trend in the AI industry, where companies are increasingly developing their own chips to improve performance and reduce reliance on traditional suppliers, addressing the growing demand from over 800 million weekly users of OpenAI’s products.
The Scope of Custom AI Accelerators
The initial systems from the OpenAI and Broadcom partnership are anticipated in the latter half of 2026, with a complete rollout expected by the end of 2029.
The 10GW figure signifies a considerable amount of computing power.
To put it in perspective, 1GW can power around 700,000 homes in the US.
These accelerators, which are specialized processors tailored to perform the mathematical operations essential for AI models, will be deployed across OpenAI’s facilities and the data centers of its global partners.
This initiative comes as numerous companies in the AI sector consider designing their own chips to improve performance and lessen reliance on traditional suppliers.
For OpenAI, whose ChatGPT and enterprise products now cater to over 800 million users weekly, the demand for computing resources is immense.
Creating custom chips is seen as a crucial part of addressing this demand.
78 Comments