OpenAI is advancing its efforts to develop its first-generation artificial intelligence (AI) chips in a bid to decrease reliance on Nvidia (NVDA, Financial) chip supplies. Sources indicate that the design of OpenAI's initial in-house chip is expected to be completed in the coming months and will be sent to Taiwan Semiconductor Manufacturing Company (TSM) for production. This initial production process is known as "taping out." Both OpenAI and TSMC have declined to comment on this development.
The expectation is to achieve mass production by 2026 at TSMC. Each taping out process is traditionally costly, potentially requiring millions of dollars and around six months to finalize chip production. Hence, OpenAI might incur significant expenses.
Internally, OpenAI views this chip, primarily focused on AI model training, as a strategic tool to strengthen its negotiation leverage with other chip suppliers. Should the trial production succeed, OpenAI plans to mass-produce its first in-house AI chip and test alternatives to Nvidia's chips later this year. This represents rapid progress compared to the typical years-long duration for other chip designers.
OpenAI's chip development team, led by Richard Ho, who joined from Google's AI chip project, has doubled to 40 members recently. The chip features a common pulse array architecture, high-bandwidth memory (HBM), and extensive network capabilities, similar to Nvidia's production methods. However, OpenAI's chip will initially deploy on a limited scale, focusing on AI model execution within the company's infrastructure.
