Psst! Where might a Chinese purchaser at any point buy top-end Nvidia simulated intelligence contributes the wake of U.S. sanctions?
A good bet is to go to the famous Huaqiangbei electronics district in the southern Chinese city of Shenzhen. Pay particular attention to the SEG Plaza skyscraper, which has shops selling everything from drones to camera parts on its first ten floors. Although the chips are not advertised, discreetly asking works.
They are not cheap. On the condition of anonymity, two vendors who spoke with Reuters in person stated that they could provide a small quantity of the American chip designer's A100 artificial intelligence chips at a price of $20,000 per unit, which is double the usual cost.
Although it is not against the law in China to buy or sell high-end American chips, export restrictions imposed by the United States have established a de facto underground market with vendors eager to avoid scrutiny from either Chinese or American authorities.
President Joe Biden's organization in September requested Nvidia to quit sending out its two most exceptional chips - the A100 and the as of late evolved H100 - to central area China and Hong Kong, part of endeavors to hinder Chinese artificial intelligence and supercomputing advancement in the midst of heightening political and exchange strains. Following that, a variety of export controls pertaining to semiconductors were implemented.
However, in the wake of OpenAI's ChatGPT's meteoric rise to global prominence, demand for high-end chips has skyrocketed, particularly for Nvidia's microprocessors, widely regarded as the best for machine-learning tasks.
Ivan Lau, Pantheon Lab's co-founder, stated, "We are talking with two vendors now to get some." Lau is trying to acquire 2-4 brand-new A100 cards in order to run Pantheon Lab's most recent AI models.
He added, "Those vendors who bought the chips outside the United States were quoting HK$150,000 ($19,150) per card." They told us right away that there would be no support or warranty.
Ten vendors in Hong Kong and mainland China told Reuters that it was easy to get a small number of A100s. Their data featured both serious interest in China for the chips and the overall simplicity with which Washington's approvals can be bypassed for little cluster exchanges.
Reuters was unable to determine the extent to which the transactions taking place contribute to satisfying demand or estimate the overall volumes of Nvidia A100 and H100 chips entering China.
The vendors, who requested anonymity due to the fact that the imports violate U.S. trade restrictions, stated that the majority of buyers are app developers, startups, researchers, or gamers. According to one vendor, customers also included Chinese government officials.
In a statement provided to Reuters, Nvidia stated that it did not permit the export of the A100 or H100 to China and instead provided substitutes with lower capabilities that complied with U.S. law.
The statement read, "We would take immediate and appropriate action if we receive information that a customer is violating their agreement with us and exporting restricted products in violation of the law."
China's industry ministry, the State Council Information Office, and the U.S. Department of Commerce did not respond to requests for comment.
In September, Nvidia stated that Chinese companies may decide not to purchase alternative Nvidia products, which could result in a loss of $400 million in third-quarter sales.
Large Chinese technology companies like Tencent Holdings and Alibaba, which have deep pockets to purchase huge quantities, are currently purchasing its new slower variants, the A800 and H800, which were developed to cushion that impact and are tailored for China.
Online offers The Chinese vendors claimed that they primarily obtained the chips in two ways: grabbing any extra stock that ends up on the market after Nvidia ships a lot of it to big U.S. companies, or importing it through local businesses that are incorporated in places like India, Taiwan, and Singapore.
As a result, the quantities they are able to obtain are minimal, well below the requirements for developing a sophisticated AI large language model from scratch.
According to TrendForce, a research company, a model like OpenAI's GPT would require more than 30,000 Nvidia A100 cards. However, only a few are able to carry out intricate machine-learning tasks and enhance existing AI models.
The majority of the forty A100 sellers listed on a website for buying electronics were situated in the electronics district of Huaqiangbei. However, listings for A100s could also be found on Xiaohongshu, which is similar to Instagram, Douyin, which is the Chinese version of TikTok, and Alibaba's Taobao e-commerce site.
ByteDance, the owner of Douyin, Alibaba, and Xiaohongshu did not respond to requests for comment.
Refurbished chips that were sold as A100s had become a common method of fraud, according to some of the vendors.
Since their introduction in March, Nvidia's more advanced H100 chips appear to be much more difficult to come by.
Vinci Chow, an economics lecturer at the Chinese University of Hong Kong, stated that he had been informed that some packs of eight H100 chips were available for purchase. His department has purchased four A100 cards from local vendors for research purposes.
However, only one of the ten vendors contacted by Reuters stated that they could acquire H100s.
According to Charlie Chai, an analyst at 86Research based in Shanghai, the U.S. probably doesn't care much about small transactions of the chips.
“We will see more strict enforcement only if/when China poses a greater threat following significant catch-ups,” he stated.
He added that because many of the Chinese AI startups that were driving purchases would eventually leave the market, the premiums that Chinese vendors are charging for the A100 and H100 chips could fall in the future. $1 equals 7.8307 Hong Kong dollars)