US restrictions threaten a third of China’s AI computing capacity by 2027
Key Facts
The United States is expanding restrictions on China’s access to advanced computing with a new rule governing the use of chips and servers through data centers in third countries, a move that could put pressure on Chinese companies’ AI model training. The US Bureau of Industry and Security said the rule, announced on November 16, 2026, aims to prevent circumvention of existing restrictions through remote-computing arrangements. The Associated Press reported that the rules restrict such access as part of efforts to limit China’s use of advanced US chips to train AI models. The move comes amid a simultaneous tightening of controls on advanced chip manufacturing for Chinese designers. On November 19, Reuters reported that TSMC had stopped Chinese orders using 16-nanometer processes or smaller unless customers obtained US export licenses.
The rule affects companies that rely on computing capacity to develop their models and services. Tencent President Martin Lau said companies in China now have less advanced computing capacity to work with, and that training schedules for upcoming versions of Tencent’s models could be affected. The Associated Press quoted Lau as saying, "Everyone has less advanced computing capacity to work with." He added, "The training schedules for the next versions of Tencent’s models will likely be affected." On an earnings call on Wednesday, Lau linked the new restrictions to declining available resources, putting the timing of training for upcoming models among the company’s immediate challenges.
Alibaba said it was still assessing the rules’ implications and that they could affect some of its global data-center expansion plans. This comes as the company expands its AI, cloud and computing operations: revenue related to these services rose 53% year over year in the July–September 2026 quarter, to about 61 billion yuan. The company’s capital spending rose 60% in the same quarter, to 50.5 billion yuan, as it expanded its AI infrastructure. Alibaba said, according to the Associated Press, that it was "still assessing the implications." The figures put the restrictions’ impact on access to computing in the context of the company’s investment and cloud-business growth, while the company has not specified the scale of any potential changes to its plans.
The new US restrictions cover remote access to processing capacity exceeding 12,000,000 aggregate TPP through data centers in any destination, when that capacity is available to companies headquartered in China, Macau or a destination listed in Group D:5. The rule published in the Federal Register requires a license for this type of access. It also gives data centers that were providing a covered service before publication of the rule 90 days from its effective date to obtain a license if they wish to continue the service. The controls therefore cover the use of infrastructure outside China, not only chip exports to the country. The Bureau of Industry and Security said the aim was to prevent the circumvention of existing restrictions through cross-border computing arrangements.
In a parallel development, Reuters reported that Chinese designers would generally be limited to 28- and 40-nanometer manufacturing processes, provided the chips do not use high-bandwidth memory or require advanced packaging. This means their available manufacturing options will be concentrated in older processes, while most leading AI chips are made using 7-nanometer processes or smaller. TSMC stopped the affected orders using 16-nanometer processes or smaller pending customers’ receipt of US export licenses. These measures link controls on access to computing abroad with restrictions on chip manufacturing within supply chains. They narrow the routes Chinese companies can use to obtain advanced computing capacity or produce advanced chips.
Restrictions on cloud access take on added significance in light of an estimate cited by the Associated Press from an October 2026 study by the Institute for Progress: remote access to AI chips through data centers outside China could represent about a third of China’s AI computing capacity by the end of 2027. This makes regulation of data centers in third countries a factor affecting the amount of capacity Chinese companies can use to train models. Meanwhile, the agency reported that Nvidia said it was "no longer shipping" chips including the H200, although the Trump administration had approved some sales to China. Lau said Chinese companies would have to rely on domestic supply chains and chips over the longer term, adding that the industry was advancing rapidly. The 90-day licensing window for covered data centers sets the next deadline for service providers seeking to continue offering remote access.