On Monday, OpenAI unveiled a strategic partnership with Broadcom aimed at procuring custom-designed chips and networking components, with an ambitious plan to expand its data center capacity by 10 gigawatts. The joint venture for developing tailored hardware with Broadcom is projected to incur costs in the tens of billions of dollars. While this collaboration offers potential technological and financial benefits, it also carries substantial financial risks. OpenAI has outlined plans to invest over US$1 trillion in infrastructure over the next few years, fueling concerns about a potential AI bubble.

According to sources with knowledge of the matter, the custom chips co-developed with Broadcom are anticipated to be 20% to 30% cheaper than purchasing NVIDIA chips. However, the project is still in its nascent stages, and the chips have not yet reached mass production. This strategic move not only promises cost savings but also empowers OpenAI to gain greater control over its supply chain, mitigate supply shortages, and achieve precise customization of chips along with optimized hardware design.

Initially, OpenAI and Broadcom had only planned to develop a single chip. However, they soon recognized the necessity of creating an entire custom system. OpenAI aims to commence the deployment of these chips by the end of 2026.