OpenAI is reportedly collaborating on a chip with Broadcom and TSMC

OpenAI started this year with quite ambitious plans to build a network of partner-operated chip factories, also known as 'foundries', as an alternative to acquiring chips from established vendors to satisfy its rising infrastructure demands. To fund such an ambitious initiative, OpenAI notably held meetings with potential investors in the UAE, reportedly looking to raise as much as $7 trillion. More recently, the company seems to be readjusting its goals, pausing its foundry plans to focus on partnering with Broadcom and TSMC to build its first in-house chip instead.

Like other AI companies, OpenAI plans to incorporate AMD chips (through Microsoft Azure) as an additional measure to meet its surging infrastructure demands as prices rise and chip shortages prevail. In a market still dominated by NVIDIA with an estimated 80% share, AMD is betting on its Instinct product line to capture a bigger slice of the market.

According to a Reuters report, OpenAI's first chip will be an inference chip, rather than a training one. Demand for training chips surpasses inference chip demand, but analysts expect this trend to reverse as more AI-powered applications are deployed over time. The company has assembled a 20-person team which includes engineers who previously built TPUs at Google, including Thomas Norrie and Richard Ho. Additionally, Broadcom is said to have helped OpenAI secure manufacturing capacity with TSMC, with production tentatively planned for 2026.

OpenAI CEO Sam Altman recently participated in a Reddit AMA with Kevin Weil, Srinivas Narayanan, and Mark Chen. When asked about the timeline for products such as Advanced Voice Mode with Vision or GPT-5, Altman confirmed that the company was focusing on the more recent o1 and its successors, partially due to infrastructure limitations constraining OpenAI's capabilities to release several products in parallel. As a result, Altman confirmed, there is no set date for the AVM with Vision launch.

Spokespeople for OpenAI recently confirmed the company would not launch a new model codenamed Orion this year. Altman confirmed this narrative in the Reddit AMA stating that although the company had some releases prepared for the year's end, none could be reasonably called 'GPT-5'.