Compared to building a chip empire with $700 million, OpenAI has chosen a more pragmatic path. Reuters reported that OpenAI has abandoned its previous chip factory plan and instead plans to focus on internal chip design work.
On October 30th, The Information cited two informed sources as saying that OpenAI is collaborating with chip manufacturer Broadcom to develop a new type of artificial intelligence chip that will be specifically designed for the inference process of AI models. In addition, both parties are also in negotiations with TSMC, the world's largest chip contract manufacturer, to advance this project.
For Broadcom, this is a significant positive development. After Google, Meta and ByteDance, OpenAI will become the fourth largest ASIC (custom chip) customer of Broadcom. After the news came out, Broadcom's stock price rose 2.9% to $160.53. Boosted by sales of artificial intelligence devices, Broadcom's stock price has risen 44% this year.
According to reports, OpenAI has formed a chip team of about 20 people, led by top engineers who have previously built tensor processing units (TPUs) for Google. According to sources, OpenAI has booked TSMC's production capacity through Broadcom and plans to have TSMC manufacture this chip by 2026. In response, TSMC executives stated that if OpenAI can commit to a large number of new chip orders, TSMC is willing to expand its chip production capacity.
On the other hand, in order to meet the increasing demand for infrastructure, OpenAI is also preparing to adopt AMD chips in addition to currently using Nvidia chips. The market has different opinions on OpenAI's self-developed chips, and there are claims that OpenAI's move may anger its largest supplier, Nvidia, and increase the difficulty of negotiations; Some argue that OpenAI has increased its bargaining power with Nvidia.
At present, there is a greater demand for training chips in the market, but with the continuous deployment of artificial intelligence applications, analysts predict that the demand for inference chips may exceed that of training chips.