OpenAI close to finalising its own chip design, send it to TSMC for fabrication by 2025-end

OpenAI is close to completing the design of its first in-house AI chip. If successfully fabricated and produced in large scale, this will allow them to reduce their dependency and on NVIDIA. Sources indicate that the design will be sent to Taiwan Semiconductor Manufacturing Company (TSMC) for fabrication by late 2025, with mass production expected in 2026. 

This custom chip is part of OpenAI’s strategy to strengthen its infrastructure and negotiating position with chip suppliers, as rising costs and demand for AI hardware push companies to seek alternatives.

The process of "taping out," where a chip design is submitted for production, is both costly and risky. Each tape-out can cost tens of millions of dollars and take six months to produce a prototype. 

More from Tech

OpenAI is designing its own in-house chip, will tie up with Broadcom, TSMC as it moves away from NVIDIA

NVIDIA, AMD squabbling over whose GPUs are faster for DeepSeek's AI benchmarks

Any errors could require redesigns, leading to additional delays and expenses. However, OpenAI appears confident, making significant progress on a task that has taken other tech giants like Microsoft and Meta much longer.

A strategic shift towards custom silicon

The chip is being developed by an in-house team led by Richard Ho, a former Google engineer experienced in AI chip design. Collaborating with Broadcom, the team is creating a chip based on TSMC’s advanced 3-nanometer process. 

The design includes high-bandwidth memory (HBM) and a systolic array architecture optimised for AI tasks, similar to the technology used in NVIDIA’s chips.

Initially, OpenAI plans to deploy the chip on a limited scale, focusing on inference tasks—running AI models rather than training them. Expanding the program to the scale of competitors like Google and Amazon would require a significant increase in resources, with costs potentially exceeding $500 million per chip design.

This effort aligns with a broader industry trend. Companies like Meta and Microsoft have also been exploring custom chips as demand for AI infrastructure grows. Microsoft plans to spend $80 billion on AI infrastructure in 2025, while Meta has earmarked $60 billion. Despite these investments, NVIDIA remains the dominant player, holding roughly 80% of the market.

TSMC to handle fabrication

TSMC, the world's largest chip manufacturer, will fabricate OpenAI’s chip using its latest 3-nanometer technology. If the initial tape-out goes smoothly, OpenAI could begin testing the chip by late 2025. The success of this initiative would reduce OpenAI’s dependency on NVIDIA and provide more control over its hardware infrastructure, which is critical to supporting the continued growth of AI models like ChatGPT.

As competition intensifies in the AI sector, OpenAI’s move to develop custom silicon reflects the high stakes and massive investments required to stay ahead in the race for AI dominance.