Nearly $ 600 billion. It is the amount of market capitalization which evaporated after the publication in open source, by the Chinese Deepseek, of a model of AI capable of reasoning. A fall explained in a hasty way. This company created in May 2023 would have managed to train its model with limited resources. More specifically, with 2.8 million hours of use of graphics cards where GPT-4 would have required 50 million hours, on more powerful models. This situation has panicked investors on real computing needs, linked to the AI revolution, sending the all the semiconductor industry to the carpet.
The reaction may seem disproportionate. On the one hand, this Chinese performance was expected. In addition, the figures are to be taken with tweezers. Tests upstream of training are actually the most gourmet resource phase. Finally, the absolute number of graphics cards is not so important. Deepseek would have used 2,048 cards and Openai ten thousand. A drop of water when compared these figures with the 6.5 million graphics cards that Nvidia should deliver in 2025.
Deepseek himself coded the chips
It is important to understand that once a model trained, it enters a so -called inference mode. This is the step where the model draws conclusions from new data – the requests that users make, the documents or internet pages that it is authorized to consult. The more efficient the models become, the more use cases diffuse, the more the inference market grows compared to that of training. For example, Openai would have spent, in 2024, $ 3 billion to rent hours of operation of graphics cards for training, and 4 billion for inference.
https://www.youtube.com/watch?
Nvidia chips are extremely efficient for training artificial intelligence models because of their design integrating a lot of memory and a specialized software environment. Called Cuda, it makes it possible to operate multiple chips together and therefore to carry out many small calculations in parallel. This architecture is particularly suitable for neural networks and the digital library which allows communication between each of the models of the model.
Deepseek stood out for his mastery of an approach called “a mixture of experts”. It consists in assembling separate subnets, each specializing in a subset of input data, in order to jointly carry out a task. Above all, Deepseek did not just take the Nvidia software environment: he himself coded the way of circulating information between the chips. What challenge Nvidia’s domination and pave the way for competitors. Indeed, from its publication, Deepseek V3 was able to operate in inference on fleas ascend 910c from Huawei. This chip would already reach 60 % of the effectiveness of a NVIDIA H100 chip.
China goes up the entire sector
This new competition has impacts on a highly integrated value chain. Companies like Openai rent servers to data centers – Amazon or Microsoft – which buy chips designed by Nvidia. These chips contain memoir cards provided by South Korean Samsung Electronics and SK Hynix, which represent the biggest NVIDIA shopping item. They are manufactured by Taiwanese TSMC, using lithography machines, the main supplier of which is Dutch ASML. Each of these actors is in a monopoly or duopoly situation. However, under the constraint of import restrictions imposed by the United States since 2019, China goes up the entire sector.
On memory, Changxin Memory Technologies (CXMT), created in 2016, is about to fill the delay it accuses about leaders. She delivered her first servers for Data Centers at the end of last year. The Huawei ascend chip is manufactured by the Chinese SMIC semiconductor manufacturer using a 7 nanometers process, even though the country no longer has access to extreme ultraviolet lithography machines. The Chinese Shanghai Micro Electronics Equipment and Shenzhen Xinkilai Technology (Sicarrier), a subsidiary of Huawei, develop their own lithographic equipment, supported on a laboratory network like the Harbin Institute of Technology, and target Challenger ASML. This is what is really hidden behind the Deepseek earthquake.
Robin Rivaton is CEO of Stonal and member of the Scientific Council of the Foundation for Political Innovation (Fondapol)
.