Deepseek has had to pull pure ingenuity, breaking the “more = better” paradigm

Satya Nadella, the general director of Microsoft, It is very clear: “Deepseek’s new model is really impressive both for how they have effectively develop a model of artificial intelligence (AI) open source which performs calculations in time of inference as for its incredible computational efficiency. We must take the developments from China very, very seriously (…) As IA becomes more efficient and accessible we will see that its use triggers, becoming a merchandise from which we cannot do without. “

In this statement to FortuneNadella gives credit to the technological triumph that the Chinese company Deepseek has reached. And he honors him that he recognizes him without ambiguity, especially if we are in mind that Microsoft is one of the competitors of the AI ​​industry that has just a few hours ago witnessed how Its value in the bag has fallen in an abrupt way after the emergence of Deepseek R1. Anyway, we can be sure that to a large extent this AI model is the result of the pressure that US sanctions are exerting on Chinese companies.

Jensen Huang, the founder and general director of Nvidia, He anticipated it in one of the statements he made at the end of May 2023 in Computex: “China is dedicating mass resources to the implementation of emerging companies specialized in the development of GPU. Do not underestimate them.” This warning was aimed at the US government in a clear attempt to prevent you about the consequences that They will have the sanctions that seek to stop the technological development of China. Huang talks about GPU Chinese designers, but his statement can be extrapolated to Chinese companies that develop AI models. After all, in this area, the GPUs and the great language models go hand in hand.

USA will continue to lead in AI

A good part of the sanctions approved by the administration led by Joe Biden as of October 7, 2022 seeks to slow down the development of the Chinese semiconductor industry, and also its AI technology. In fact, as we have just seen, the integrated circuits and the AI ​​go hand in hand. These prohibitions prevent NVIDIA, AMD or Intel, among other chips manufacturers for AI applications, sell their most advanced GPU to their Chinese clients. This is presumably the germ of Deepseek’s greatest achievement.

According to Depseek the infrastructure used to train its AI model 2,048 NVIDIA H800 chips

If we stick to the information that this Chinese company has made the infrastructure used to train Depseek R1 agglutina 2,048 chips H800 of Nvidia. And training with 671,000 million parameters has cost 5.6 million dollars. This is precisely what Satya Nadella speaks in the statements that we have reviewed a few lines above. These figures are extremely restrained. Some analysts defend that, in reality, its infrastructure brings together 50,000 GPU H100 Buy through intermediaries, but for the moment it is just a conjecture.

If we give the statements made by the Deepseek spokesmen to good Financial Timesand for the moment it is reasonable to do so, the reason why their engineers have mounted their training infrastructure on NVIDIA H800 GPUs is that US sanctions have prevented them from accessing the H100 chips, which are more powerful. The prohibitions of November 16, 2023 They prevent Nvidia Delivering to their Chinese clients the H800 GPUs, but presumably at that time Depseek already had its infrastructure assembled. In any case, at this situation the meritorious is that with a relatively modest chip this Chinese company has materialized a remarkable achievement.

Depseek’s undisputed success is a victory for China, but it is a partial victory. This technological war at the moment is winning the US. Its advantage lies in an unappealable reality: the country led by Donald Trump controls so much Most GPU manufacturers Like many of the companies that are dedicated to developing AI models. And the latter have access without restrictions on the most advanced GPUs produced by NVIDIA and other companies.

China has the Huawei GPU, which They seem to be very competitive In inference processes, and also with those of companies such as Moore Threads, Metax, Biren Technology, Innosilicon, Zhaoxin, Iluvatar Corex, Denglinai or Vast Ai Tech, among others. But, for the moment, it is in a position of clear disadvantage. Even so, this confrontation goes for long, so any conclusion that we reach about which country will finally impose itself in the AI ​​domain, if any, it would be premature.

Image | Nvidia

More information | Fortune | Financial Times

In Xataka | China is closely monitoring the United States movement with Stargate. And your answer has already prepared

Leave a Comment