The good news is that it appears it can only buy them from unidentified buyers in extremely modest amounts.
The United States government has placed significant limitations on the GPUs that Nvidia is permitted to export to China. The company’s most recent H100 chip has been placed on the naughty list, and despite its advanced age, the company’s older A100 chip is also on the list. Nvidia produced versions of both cards that were explicitly designed to be neutered, and now those versions have also been restricted. In spite of this circumstance, it would appear that state actors, colleges, and scientific bodies in China are nevertheless able to get the prohibited chips, albeit in extremely small amounts.
Reuters has recently published a story that sheds light on how challenging it might be to shut off a country as vast as China from something as small as a graphics processing unit (GPU). Its analysis reveals that a number of Chinese organizations have lately acquired Nvidia GPUs that are prohibited, and it is based on publicly available tenders, which look to be similar to purchasing contracts. According to the findings of an investigation conducted by Reuters into public records, the Harbin Institute of Technology and the University of Electronic Science and Technology of China have lately acquired limited graphics processing units (GPUs). For the purpose of training a big language model, the former purchased six A100 chips in May of the previous year. The latter made a purchase of a single GPU for an as-yet-undetermined reason. Reuters reports that both organizations have been accused of being “involved in military matters” and are prohibited from obtaining Nvidia gear as a result of this allegation.
It should come as no surprise that the report does not specify to whom the graphics processing units (GPUs) were sold or how they got to possess them in the first place. However, the study states that purchasing and selling Nvidia graphics processing units (GPUs) within China is not against the law; the only restriction is that they cannot be shipped to China; once they are within the country’s borders, however, they can be sold without incurring any penalties. According to reports from Reuters, it is conceivable that the chips were stolen from major orders placed by businesses based in countries that are geographically close to the United States, such as India, Taiwan, and Singapore. According to the findings of the study conducted by Reuters, the suppliers were not included on the list of companies that Nvidia has certified. Nvidia has stated in a statement that it complies with export restrictions and that it requires its partners to do the same.
According to Reuters, the tenders that it analyzed suggested that all of the chips were being utilized “for AI,” which is not surprising that this is the case. Nevertheless, taking into consideration the relatively low numbers involved, it would appear that the United States’ chip ban is working rather well. Because it would take thousands of these graphics processing units (GPUs) for an organization to put together something even remotely viable, we are skeptical that the government (or Nvidia) would be overly concerned about a few of them falling through the cracks. Nevertheless, it demonstrates that there is a working underground for these graphics processing units (GPUs), which was always an inevitable conclusion regardless of the chip limits.
All things considered, the research revealed that more than one hundred transactions had taken place since the H100 and A100 from Nvidia were prohibited in October of 2022, as well as when the H800 and A800 were prohibited in October of last year. There was at least one group associated with the military among the purchasers, in addition to state entities and prestigious educational institutions. Due to the fact that the tenders indicate that the purchases contain a single GPU or two or three GPUs at the same time, it appears that the projects are on a relatively small size or are being constructed gradually over time. In spite of this, it would take decades to create something like to ChatGPT at that rate.