After U.S. Commerce Secretary Raimondo blocked Nvidia's chips for China, at least two Chinese Internet companies tried to replace them with self-developed AI chips, which had an impact on Nvidia, and Nvidia founder and CEO Jensen Huang said he cooperated with U.S. policies, but wanted to continue to sell compliant chips to China.
Chinese companies have made many efforts to solve the problem of AI chips, and the best alternative chips in this regard are the chips of domestic AI chip companies, which claim that their AI chips are comparable to NVIDIA's A100 chips, and these chips can meet most of the needs of domestic enterprises.
In order to speed up the establishment of an ecosystem, some domestic AI chip companies have taken the initiative to provide AI chips to a number of universities for free, hoping to make professors and students in related industries familiar with these domestic AI chips and establish an ecology for them.
Domestic Internet companies have two self-developed AI chips, their AI chips in the low-end AI chips have a certain substitution ability, these Internet companies self-developed AI chips can be specially customized, in terms of efficiency is even higher than NVIDIA, Google is an example, Google has replaced general chips with self-developed AI chips, Alpha Dog is achieved with Google's self-developed AI chips, proving that the specially customized chips of Internet companies are not worse than NVIDIA.
After years of hard work, Nvidia's AI chip performance is four times that of China's best AI chips, highlighting that it does have a unique advantage in AI chip technology.
Another killer feature of NVIDIA is the CUDA ecosystem it has built for many years, as early as around 2005, it launched the CUDA platform, and after many years of building, most of the world's AI software is based on the CUDA platform, thus forming the CUDA ecosystem, which is also the biggest obstacle for AI companies to adopt other AI chips.
AMD's newly launched AI chip performance reaches more than 2400 TFLOPS, while NVIDIA's top AI chip H100 has a performance of more than 1970 TFLOPS.
If the United States further prevents NVIDIA from China's high-end chips, then Chinese companies will not be able to use domestic or self-developed AI chips to replace NVIDIA chips, so that another ecology will gradually form for a long time, and CUDA's ecological advantages are no longer there.
To that end, Huang told Raimondo, "We don't want to break the rules." Tell us that we will cooperate", meaning that Huang hopes that the United States will give the rules that should not ban all AI chips from China**;;In interviews on other occasions, Raimondo said that she would allow Nvidia to sell most AI chips to China, except for the most complex and powerful AI chips.
For this reason, NVIDIA originally planned to customize the high-end A100 and H100 again, in fact, it castrated up to eighty percent of the performance, in this case, Nvidia's custom AI chips have been nearly eighty percent lower than China's AI chips, even if there are ecological advantages, it can not attract Chinese companies, maybe NVIDIA also realized that this kind of custom chip really has no market demand, so far it has not been launched, only the low-end AI chip is customized for the Chinese market.
NVIDIA is so reluctant to give up the Chinese market, because the Chinese market can bring it a lot of revenue - accounting for a quarter of its revenue, it is estimated that the Chinese market can bring NVIDIA $9 billion in revenue every year, and because of the high sales in the Chinese market, more profitable, NVIDIA is worried that the loss of the Chinese market will lead to a sharp decline in performance like Micron and other American chip companies, affecting future development.