OpenAI is reportedly accelerating its efforts to reduce dependence on Nvidia by developing its own artificial intelligence (AI) silicon, with plans to finalise the design of its first in-house chip in the coming months. The ChatGPT maker is preparing to send its design for fabrication to Taiwan Semiconductor Manufacturing Company (TSMC), according to sources familiar with the matter, reported Reuters.
Reportedly, the process of submitting a chip design for production, known as ‘taping out’, marks a significant milestone in OpenAI’s ambition to mass-produce its AI chips by 2026. Given that an initial tape-out can cost tens of millions of dollars and take approximately six months to complete, there is no certainty that the first iteration will function as intended. Any issues could necessitate another cycle of troubleshooting and re-submission, potentially delaying progress.
The initiative is reportedly seen as a strategic move to enhance OpenAI’s negotiating leverage with leading chip suppliers. Following this initial design, OpenAI plans to develop increasingly sophisticated processors to expand its capabilities over successive iterations.
If successful, the new AI chip could provide an alternative to Nvidia’s dominant products, which currently hold an estimated 80 per cent market share. The rapid development of OpenAI’s design, which has progressed faster than many other chip designers, suggests the company is making significant strides in hardware innovation. Tech giants such as Microsoft and Meta have spent years attempting to develop their own AI chips with mixed results.
TSMC is reportedly set to manufacture OpenAI’s AI chip using its cutting-edge 3-nanometre process technology. The design incorporates a systolic array architecture, high-bandwidth memory (HBM), and extensive networking capabilities—similar to Nvidia’s existing chips.
While OpenAI’s first-generation chip will likely be capable of both training and running AI models, it is expected to be deployed initially on a limited scale, focusing primarily on model inference rather than full-scale training.
(With inputs from Reuters)
