He is camping on the roof of the (virtual) world. However, Sam Altman is worried. The boss of OpenAI fears that a serious hardware problem will hamper the crazy dynamics of his company: the lack of AI chips. Without them, it is impossible to create and use ever more sophisticated language models (LLM). So, the tech giants are building up arsenals of “H100”, the most popular AI chip of the moment, from Nvidia. In 2023 alone, Omdia Research estimates that the latter will have delivered around 150,000 processors of this type to Microsoft and as many to Meta. Google and Amazon reportedly received 50,000 each. And Chinese companies are not left out: Tencent has reportedly purchased 50,000, Baidu, 30,000 and Bytedance (TikTok), 20,000. according to the firm market research.
AI stars fight for the same chip
Obtaining these AI chips has become “harder than finding drugs,” Elon Musk quipped last spring. And this flea war is likely to last. Mark Zuckerberg announced in January that Meta was building a colossal infrastructure of 350,000 H100 chips. A context that explains why Sam Altman is exploring other options. According to Bloomberg, the boss of OpenAI is seeking to raise billions of dollars in order to build a network of semiconductor factories. A delicate bet. Certainly, the dizzying progress in artificial intelligence should make it possible to design increasingly efficient AI chips. Which, in turn, will help create increasingly powerful AIs. A virtuous circle that makes all players in the sector fantasize. Microsoft, Google, Meta, Amazon… Many of them are working on “in-house” chips (focusing mainly on the design phase).
Competing with specialist players like Nvidia and TSMC, however, would be a completely different matter. Opening a state-of-the-art factory requires tens of billions of dollars and the recruitment of an ultra-skilled workforce, which is hard to find. It also takes a long time. The veteran TSMC should take three years to open its first factory in Arizona. And he has already warned that the opening of a second site in this state (initially planned for 2026) would be one to two years late. In addition to their market experience, players specializing in the design or manufacturing of chips have thousands of patents on the subject and vast resources linked to the fact that they are not building an in-house chip, but products adapted to a large customer base.
The cozy margins they make also allow them to devote a lot of money to their R&D and improve their products very quickly. Successor to Nvidia’s star AI chip, the H200 should arrive in the second quarter of 2024. The shortage that is bothering Sam Altman will eventually resolve itself. THE president of TSMC promises that it is only “temporary” and that by the end of the year, supply will have matched demand. Patience.