OpenAI is collaborating with Broadcom Inc. to develop a specialized AI chip designed specifically for running AI models post-training, according to sources familiar with the project.
According to sources, OpenAI and Broadcom are consulting with Taiwan Semiconductor Manufacturing Co., the largest chip contract manufacturer globally. OpenAI has been planning a custom chip and exploring its applications for about a year, although discussions remain in the early stages.
OpenAI did not provide a comment, while Broadcom’s representative did not respond to requests. A TSMC spokesperson stated they do not comment on rumors or speculation. On Tuesday, Reuters reported on OpenAI’s ongoing discussions with Broadcom and TSMC, following a June report by The Information that Broadcom had discussed creating an AI chip for OpenAI.
Designing and producing a chip is a lengthy and costly process. OpenAI is focusing less on GPUs, which are primarily used to train generative AI models—a market dominated by Nvidia—and is instead seeking a specialized chip designed for inference, enabling the software to respond to user requests. Analysts and investors anticipate a growing demand for inference-supporting chips as more tech companies leverage AI models for increasingly complex tasks.
One source mentioned that OpenAI might still explore creating its own network of foundries or chip factories in the future, but for now, the company recognizes that collaborating with partners on custom chips is a faster, more feasible option. Reuters previously reported that OpenAI was scaling back its efforts to establish its own chip manufacturing capabilities.
On Tuesday, Broadcom’s shares rose by 4.2%, closing at $179.24 in New York trading, adding to their 54% increase this year. TSMC’s U.S.-listed shares also closed over 1% higher.
Broadcom is the leading designer of application-specific integrated circuits, which are chips tailored to a specific customer-defined purpose. Its largest client in this sector is Alphabet Inc.’s Google, with additional partnerships including Meta Platforms Inc. and ByteDance Ltd., the owner of TikTok.
Last month, when asked about new customers in light of the high demand for AI training, Broadcom CEO Hock Tan stated he would only expand his limited customer list once projects reach volume shipment stages.
“It’s a challenging product for any customer to deploy, so we don’t count proof-of-concept projects as production volume,” he said during an earnings call.
OpenAI’s services demand substantial computing power for development and operation, primarily relying on Nvidia chips. To address this demand, the industry has been actively seeking alternatives to Nvidia, including adopting processors from Advanced Micro Devices Inc. and creating in-house versions.
OpenAI is proactively pursuing investments and partnerships in data centers, which will serve as the future hub for AI chips. The startup’s leadership has presented to the U.S. government the necessity for larger data centers, while CEO Sam Altman has engaged with global investors, including those from the Middle East, to secure funding for this initiative.
“It’s certainly a challenge,” OpenAI Chief Financial Officer Sarah Friar stated on Bloomberg Television on Monday. “It’s a stretch not only in terms of capital but also in my own learning. To be honest, we are all navigating this area: Infrastructure is key to our success.”