OpenAI to Soon Complete its Own Chip Design to Lessen Reliance on Nvidia

OpenAI is moving forward with its plan to reduce its reliance on Nvidia for chip supply by developing its first generation of in-house artificial intelligence silicon.

The ChatGPT maker will finalise the design for its first in-house chip in the coming months and will send it to Taiwan Semiconductor Manufacturing Co for fabrication. The process of sending a first design to a chip factory is known as "taping out." 

The update demonstrates that OpenAI is on track to meet its ambitious goal of mass production at TSMC by 2026. A typical tape-out costs tens of millions of dollars and takes about six months to produce a finished chip, unless OpenAI pays significantly more for faster manufacturing. There is no guarantee that the silicon will work on the first tape out, and if it fails, the company will have to diagnose the problem and repeat the tape-out process.

The training-focused chip is viewed within OpenAI as a strategic tool for increasing the company's negotiating leverage with other chip suppliers. Following the initial chip, OpenAI engineers intend to create increasingly advanced processors with broader capabilities with each iteration.

If the initial tape out goes smoothly, ChatGPT will be able to mass-produce its first in-house AI chip and potentially test an alternative to Nvidia's chips later this year. OpenAI's plan to send its design to TSMC this year demonstrates that the startup has made rapid progress on its first design, which can take years for other chip designers.

The chip is being developed by OpenAI's in-house team, led by Richard Ho, which has grown to 40 people in recent months, in collaboration with Broadcom. Ho joined OpenAI more than a year ago from Alphabet's Google, where he oversaw the search giant's custom AI chip program. Last year, Reuters reported on OpenAI's plans to collaborate with Broadcom.
 
Ho's team is smaller than the large-scale efforts at tech behemoths like Google and Amazon. A new chip design for an ambitious, large-scale program could cost $500 million for a single chip version, according to industry sources familiar with chip design budgets. The costs of developing the necessary software and peripherals could double

Generative AI model makers such as OpenAI, Google, and Meta have demonstrated that increasing the number of chips strung together in data centers makes models smarter, and as a result, there is an insatiable demand for the chips.

Meta has announced that it will spend $60 billion on AI infrastructure in the coming year, while Microsoft has stated that it will spend $ 80 billion by 2025. Nvidia's chips are currently the most popular, with a market share of approximately 80%. OpenAI is participating in the $500 billion Stargate infrastructure program, which was announced by US President Donald Trump last month.

However, rising costs and reliance on a single supplier have prompted major customers such as Microsoft, Meta, and now OpenAI to consider in-house or external alternatives to Nvidia's chips.

While OpenAI's in-house AI chip can train and run AI models, it will initially be deployed on a limited scale, primarily for running AI models, according to the sources.

The chip will play a limited role in the company's infrastructure.

To build an effort as large as Google or Amazon's AI chip program, OpenAI would need to hire hundreds of engineers.

TSMC is producing OpenAI's AI chip with its advanced 3-nanometer process technology. According to sources, the chip has a common systolic array architecture, high-bandwidth memory (HBM), which Nvidia also uses in its chips, and extensive networking capabilities.