The US has banned companies like Nvidia from selling their most advanced AI chips to China since 2022. But if loopholes exist, profit-hungry corporations will find and exploit them. The Information published a bombshell report on Thursday detailing how Oracle allows TikTok owner ByteDance to rent Nvidia’s most advanced chips to train AI models on US soil.
ByteDance, which many US lawmakers believe has direct ties to the Chinese government, is reportedly renting US-based servers containing Nvidia’s coveted H100 chips from US cloud computing company Oracle to train AI models. The practice, which runs against the spirit of the US government’s chip regulations, is technically allowed because Oracle is merely renting out the chips on American soil, not selling them to companies in China.
The US government has cracked down on exporting the chips to China as an extension of the tensions between the two nations. The Biden Administration fears the nation could use advanced AI for military or surveillance purposes or to gain an economic upper hand. The US government passed bipartisan legislation in April that will force ByteDance to either sell its US operations or face a ban. But ByteDance still has until early next year to close a deal, and it’s suing the US government, which could delay enforcement.
Although ByteDance is training its models in the US, “it could be difficult to stop them from sending the models they produced back to their headquarters in China,” according to US-based cloud providers and a former Nvidia employee who spoke to The Information. Quite the loophole, indeed.
ByteDance’s Project Texas initiative, which the company claims siloes off TikTok’s US operations from its Chinese leadership to allay US fears, is at the heart of the arrangement. However, former ByteDance employees have described Project Texas as “largely cosmetic,” as they claim the company’s US wing regularly works closely with its Beijing-based leadership.
ByteDance isn’t the only Chinese company looking to game the rules. The Information says Alibaba and Tencent are discussing similar arrangements to gain access to the sought-after chips. Those deals could be harder to squash because they have their own US-based data centers and wouldn’t have to rent servers from American companies.
Not every company has been as willing as Oracle to skirt the law’s intent. “Two small American cloud providers” reportedly turned down offers to rent servers with Nvidia’s H100 chips to ByteDance and China Telecom because “they seemed to go against the spirit of U.S. chip restrictions.” However, Oracle, cofounded by American businessman Larry Ellison and run by current CEO Safra Catz, apparently found the opportunity for profit through technically legal workarounds too tempting to pass up.
The US Commerce Department, the bureau that could close the loophole, may already be aware of the practices. Earlier this year, the department proposed a rule that would require US cloud providers to verify foreign customers’ identities and notify the US if any of them were training AI models that “could be used in malicious cyber-enabled activity.” However, the Commerce Department recently said most cloud providers disapproved of the proposal, claiming “the burden of additional requirements might outweigh the intended benefit.” In the meantime, the proposed rule, which could theoretically plug the loophole, remains in limbo.
But even if the US manages to shut down that exploit, The Information says it wouldn’t cover Chinese cloud providers like Tencent and Alibaba from buying Nvidia’s chips and using them to train AI models in their own US-based data centers. The Commerce Department will have its hands full figuring this one out as business and defense interests wrestle for control.