OpenAI, Broadcom to mass produce first custom AI chips in 2025

OpenAI plans to begin mass production of its first custom artificial intelligence chips in 2025 through a partnership with Broadcom, a move aimed at easing pressure on the company’s reliance on Nvidia and securing capacity to run and train large AI models, people familiar with the matter said. Broadcom chief executive Hock Tan on a recent earnings call referenced an undisclosed new customer committing roughly $10 billion in orders, comments that industry sources and multiple reports have since linked to OpenAI.

The chips, which industry participants say were codesigned with Broadcom, are expected to be used internally by OpenAI rather than sold to outside customers. The collaboration follows an initial phase of cooperation that began last year, according to people close to the project, and accelerates a trend among hyperscalers and AI companies to develop bespoke silicon optimized for their workloads. Broadcom and OpenAI declined to comment.

OpenAI’s move reflects mounting demand for specialized hardware to handle the intensive compute needs of next-generation generative AI models. For several years Nvidia has been the dominant supplier of such accelerators, and the scramble for...