China’s AI compute surge heavily reliant on US chips, offshore cloud access: report

China's ability to harness artificial intelligence computing power is set to double by next year, but US-designed chips will account for roughly 65 percent of this capacity, according to recent research from the Washington-based Institute for Progress (IFP). Total compute available to Chinese entities is anticipated to increase from around 4.6 million to 8.8 million "H100-equivalents" in 2027, a metric that gauges performance against Nvidia's data-center graphics processing units.

The research team calculated domestic chip production, legal imports, smuggled processors, and remote access via overseas cloud networks to arrive at these projections. Remote access alone is expected to grow to 2.8 million H100-equivalents by the end of 2027, up from approximately 1 million in the current year. Despite the US holding a roughly 16-fold advantage in AI training compute over China, Chinese models are only slightly lagging behind their American counterparts by 6.5 months.

The IFP attributes this gap closure to a technique known as distillation, where developers train AI systems using the outputs of superior models. However, the institute has previously dismissed allegations from the US that domestic Chinese developers engage in "aggressive, malicious" AI distillation. In response to these findings, the institute has urged US policymakers to tighten regulations on overseas cloud access, chip smuggling, and legal exports.

A separate report by the Institute for AI Policy and Strategy published on Tuesday noted that US-designed compute power remains legally accessible for rent to foreign entities barred from purchasing hardware. The report recommended that the US Bureau of Industry and Security restrict Chinese access to cloud capacity sufficient for training or running large-scale AI models, while advocating for stronger identity and ownership verification measures.

While chip exports are subject to licensing restrictions, the US eased export controls in January, allowing case-by-case reviews for Nvidia's H200, AMD's MI325X, and comparable processors under specific security conditions. Amid these developments, Chinese tech companies are striving to develop domestic alternatives. Huawei Technologies recently announced that its Ascend 960DT chip, intended for AI model training, will be available in the first quarter of 2027, three quarters ahead of its initial schedule.

The Shenzhen-based firm has already deployed over 1,000 Atlas 900 A3 SuperPoDs and has introduced the large-scale commercial use of its Atlas 950 system. However, scaling domestic hardware production poses significant challenges. A recent report by San Francisco-based research group Epoch AI identified high-bandwidth memory, a critical component that feeds data to AI chips, as the primary constraint on Huawei's output.

Epoch AI estimates that Huawei's chip production this year would generate less than 4 percent of the computing power of Nvidia's global shipments.