As the Chinese startup DeepSeek announced that it had developed an AI model at low cost, the price of graphics processing units (GPUs) has attracted much attention. Some analysts believe that the market price of the GPUs from Nvidia, which DeepSeek uses, is about 10% to 30% cheaper than the cutting-edge products used by American companies. It does not seem to be extremely cheap.

It is reported that when DeepSeek developed an AI model comparable to the "ChatGPT" of OpenAI in the United States, it used Nvidia's "H800" chip. It is equivalent to a cheaper version of the "H100" used by American companies.

Export of AI chips

The US government restricted the export of AI chips to China in 2022. As H100 and other products became regulated, Nvidia supplied China with the reduced-performance H800 (now H800 has also become regulated). Tomofumi Nakayama, CPO of AI company KARAKURI, explained that "the computing speed of H800 is only half that of H100."

Nvidia did not disclose the price, although the actual transaction price will vary greatly depending on the purchase quantity and contract conditions. However, several people involved in semiconductor transactions revealed that H800 is about 4 million to 5 million yen (about RMB 192,300 to RMB 240,300), and H100 is about 5.5 million yen (about RMB 264,400).

About DeepSeek

It is reported that DeepSeek used 2,000 to 3,000 H800s to develop the AI model "V3". The total amount of GPUs used is simply calculated to be 8 billion to 15 billion yen (approximately RMB 385 million to RMB 721 million).

DeepSeek explained that the development cost of V3 was $5.576 million. Assuming that AI is allowed to learn for 2.788 million hours, the cost per hour is $2, which is less than one-tenth of the cost of the American AI model. In other words, the price difference of GPUs is not the difference in cost.

Professor Yutaka Matsuo of the University of Tokyo, who is the leading AI researcher in Japan, pointed out that "the development of AI models requires dozens or hundreds of repeated experiments, and it takes time before about 2.8 million hours of learning. This is logical. The time spent on this and the GPU should have been included in the cost."

There is also a view that DeepSeek has reduced the cost of GPUs by leasing.

Of course, this does not mean that cutting-edge GPUs are not needed. Akira Minamikawa, senior consulting director of British research company Omdia, said, "American companies are not lagging behind in terms of financial strength and talent, and will promote efficient development." In the future, in the competition around AI development, the question of what kind of GPU to use and how to use it may receive more attention.

Source: Content from Nikkei Chinese website

View more at EASELINK