Chinese AI teams bypass export controls to gain access to high-end US chips through intermediaries, unearthing potential loopholes in Washington’s blockade of cutting-edge technology in the country.
US-sanctioned AI watchdog groups have found ways to obtain limited technology by using cloud providers and third-party leases, as well as buying chips through subsidiaries in China.
iFlytek, a state-backed voice recognition company blacklisted by Washington in 2019, leases access to Nvidia’s A100 chips, which are critical in the race to develop groundbreaking technologies. AI Applications and services, according to two employees with direct knowledge of the matter.
Facial recognition group SenseTime, which was sanctioned at the same time as iFlytek, used intermediaries to purchase banned components from the US, according to three senior officials familiar with the situation.
Private cloud computing companies also provide access to high performance US chips. AI-Galaxy, a Shanghai-based cloud computing company founded by former Nvidia and AliCloud employees, charges $10 for an hour’s access to eight of its A100 Nvidia chips.
The ability of Chinese AI teams to continue to gain access to critical high-performance Nvidia chips and other cutting-edge technologies highlights the challenge the US faces in enforcing its trade restrictions against Chinese companies.
In October last year, Washington introduced Unilateral export control This banned US companies from selling advanced chip manufacturing equipment and high-end semiconductors, including the A100, to Chinese groups, extending their trade restrictions only from certain blacklisted companies.
“iFlytek can’t buy Nvidia chips, but that’s not a problem because it can rent them and train our datasets on computer clusters from other companies,” said an executive familiar with the AI firm’s operations.
“It’s like a car rental system. You cannot take the chips out of the establishment. It’s a huge building with a computer cluster, and you’re buying processor time. [central processing unit] or GPUs for training models,” the person said.
While iFlytek cannot fully own the chips under US export controls, two employees said the lease system was a good, albeit more expensive, alternative. An iFlytek engineer said the company “leases the chips and equipment on a long-term basis, which is effectively the equivalent of owning them.”
iFlytek was banned from directly buying these semiconductors after it was blacklisted by Washington for its alleged role in providing technology for government surveillance of Uyghur Muslims in Xinjiang.
In some cases, SenseTime has bought advanced chips directly through its subsidiaries, which are not on Washington’s “list of entities”, according to three senior officials familiar with the situation.
SenseTime stated that it “strictly enforces various laws and regulations related to domestic and foreign trade” and that the group has developed a program to ensure “compliance with trade standards.”
iFlytek did not respond to a request for comment.
A spokesperson for the US Department of Commerce said its Bureau of Industry and Security is “actively investigating possible export control violations but does not comment on specific allegations.”
An export control expert in Washington said US export regulations do not apply to cloud providers, even if restricted chips are used. They added that the violation only occurred if the technology was used to create weapons of mass destruction.
Local governments in China rushed to help the industry stockpile high-tech chips ahead of Washington unveiling tough export controls last October.
One person close to iFlytek said US export restrictions fueled the proliferation of state-backed computer clusters that hoarded Nvidia chips and rented technology access to blacklisted companies.
In addition, a government official in Zhejiang noted that “several places [local governments and companies] have built or are building AI computing centers that provide cloud computing rental services to businesses. It’s part of a new infrastructure supported by national policy.”
“All AI companies are using and accumulating cutting-edge chips like the Nvidia A100, including us,” said an engineer from SenseTime, adding, “We are very meticulous and meticulous in our compliance and legal work. “.
“The artificial intelligence industry in China is facing huge risks due to US sanctions,” said a person close to SenseTime. “We expect the US to impose additional sanctions on Chinese AI companies, their suppliers and customers. We need to make sure our supply chain and sales remain stable,” the source added.
Private cloud providers are facing growing demand for Nvidia’s A100 chips, which are used in generative AI models such as OpenAI’s ChatGPT.
An executive at the US tech giant in China said they continue to offer local companies access to cloud computing powered by A100 chips, and the offer has attracted a large number of startups trying to clone ChatGPT. The company’s US legal team was initially wary of continuing to offer cloud services running on A100s chips, but ultimately decided it was not against export controls.
Although US export controls do not explicitly prohibit access to chips through cloud service providers, some companies still take steps to hide their identity.
An executive at a cloud services startup in Shenzhen said they’ve seen a surge in demand for the A100 chips from “weird” companies that have masked their real identities behind shell companies.
“Our colleagues have told us that we should never be serious about finding out who is behind them,” the chief said.
Nvidia stated that “While we cannot control every future use or subsequent sale of our products, we require our distributors to comply with all US export regulations and only sell to eligible commercial, consumer, and academic customers who use our products for beneficial purposes.” .
Additional reporting by Ryan McMorrow and Sun Yu in Beijing and Tim Bradshaw in London.