In a strategic move to diminish its dependence on Nvidia, OpenAI has teamed up with Broadcom to manufacture custom chips for its AI data centers. This collaboration is the latest among OpenAI's initiatives aimed at securing sufficient computing resources essential for applications such as ChatGPT and Sora, and to further its goal of creating superintelligent AI systems.
OpenAI believes that by creating its own chips, it can integrate insights gained from the development of advanced models directly into the hardware, which could enhance capabilities and intelligence significantly.
According to an announcement made on Monday, this partnership will allow OpenAI to create and implement “10 gigawatts of custom AI accelerators” utilizing its custom chips and systems. To provide a perspective, a standard nuclear reactor typically produces around one gigawatt of power.
Broadcom is planning to start deploying the necessary equipment by the latter half of 2026, with the partnership set to conclude by the end of 2029. Sam Altman, co-founder and CEO of OpenAI, stated that this collaboration is vital for constructing the necessary infrastructure to unleash AI’s potential and deliver tangible advantages to both individuals and businesses.
This announcement follows OpenAI’s recent agreements for a six-gigawatt supply with AMD and a ten-gigawatt agreement with Nvidia. These infrastructure deals were facilitated by OpenAI’s modification of its exclusive computing partnership with Microsoft, making them feasible.
The move to develop proprietary chips is part of a broader industry trend, with leading tech companies like Meta, Google, and Microsoft aiming to strengthen crucial supply lines and reduce their reliance on Nvidia’s AI chips. While these custom chip initiatives have not yet posed a significant challenge to Nvidia’s dominance, they have certainly been beneficial for other chipmakers, including Broadcom.