OpenAI, the maker of ChatGPT, is set to finalize its first in-house AI chip design in the coming months, taking a key step toward reducing its reliance on NVIDIA. According to Reuters, the company is expected to send the chips to TSMC for fabrication soon as it aims for mass production in 2026.
Notably, the report suggests that TSMC is producing OpenAI’s AI chip on its 3nm process, featuring a systolic array, HBM, and strong networking—similar to NVIDIA’s design. OpenAI and TSMC declined to comment, according to Reuters.
The chip will be used for both training and inference. It has been designed by an in-house team of 40 people led by Richard Ho. The team is working in collaboration with Broadcom.
Ho got his BSc and M.Eng at Manchester University and a PhD at Stanford. He was key to the design of Google’s TPUs and served as SVP of engineering at Lightmatter before joining OpenAI.
The objective of the chip initiative is to give OpenAI a degree of independence from total reliance on NVIDIA’s processors, which cost between $25,000 and $40,000 and are sold out till late in the year.
However, sources cited by Reuters note that OpenAI’s AI chip, though capable of training and running models, will initially have a limited role. Meanwhile, OpenAI sees its AI chip as a strategic move to gain leverage over suppliers, as highlighted in the report.
OpenAI’s in-house team, led by ex-Google chip expert Richard Ho, has doubled to 40 members and is co-developing the AI chip with Broadcom, the report adds.
OpenAI is certainly not the only tech giant working on in-house chips. Microsoft, Amazon, Meta, and Google are also trying to develop their own chips to gain negotiating leverage with NVIDIA. Though DeepSeek’s breakthrough raises doubts about future chip demand, soaring costs and NVIDIA reliance are pushing firms, including OpenAI, to seek alternatives, the report says.
