OpenAI is finalizing the design for its first custom AI training chip and is currently in the tape out phase, the final process before a semiconductor is manufactured.

According to a report from Reuters, the chip will be fabricated by TSMC and made using the chipmaker’s 3nm technology.

Citing unnamed sources, the news outlet said the silicon will feature the same systolic array architecture with high-bandwidth memory as Nvidia’s chips.

Mass production of the chip is slated for 2026, but, despite OpenAI’s significant investment in building out its hardware team – including hiring former Lightmatter chip engineering lead and Google TPU head Richard Ho to head up its hardware team – there’s no guarantee this first tape out will result in the chip being fully functional.

The company currently has approximately 40 people working in its custom silicon team, and is collaborating with chipmaker Broadcom, which was heavily involved in the development of Google's TPU AI chip.

The Reuters report added that while the custom chip will be capable of both training and running AI models, its deployment will initially be modest and the chip will have a limited role within the company's infrastructure.

Neither OpenAI nor TSMC responded to Reuters’ request for comment.

OpenAI's custom chip plans have long been rumored

OpenAI was first reported to be in talks with chip designers about developing a new AI server chip in July 2024, although rumors that it was considering making its own AI hardware first emerged in late 2023.

Before he was suddenly fired and then rehired by OpenAI in November 2023, CEO Sam Altman had apparently been seeking investment to build an AI chip company codenamed Tigris.

Altman had also reportedly held discussions with Middle Eastern investors and chip manufacturing companies in an effort to raise between $5 and $7 trillion to shore up the AI chip supply chain with a venture that would include the development of a global network of chip fabs. It appears that plan has since been abandoned.

OpenAI allegedly opted to make its own custom chips to reduce its reliance on Nvidia GPUs, which are both expensive and difficult to source, with many customers facing long lead times due to high demand for the chips.

In January, OpenAI launched the Stargate data center project with backing from SoftBank, Oracle, and investment firm MGX. The $500bn venture aims to develop gigawatts of capacity to support AI lab's needs.