China's AI Compute Power to Double by 2027, Still Relies on US Chips

The Washington-based Institute for Progress (IFP) projects that AI computing capacity available to China's entities will double next year. IFP expects available capacity to increase from about 4.6 million H100-equivalents to 8.8 million in 2027. H100-equivalents measure computing capacity against Nvidia's data-center graphics processing units. IFP estimates that US-designed chips will supply about 65% of that capacity, indicating continued dependence on foreign hardware.

IFP said the projections combine domestically produced chips, legally imported chips, smuggled processors and computing capacity accessed through overseas cloud networks. The report, published this week, forecasts remotely accessed capacity of 2.8 million H100-equivalents by the end of 2027, versus roughly 1 million this year. IFP said the United States has roughly 16 times China's AI training computing capacity. However, IFP said leading models from China trail leading models from the United States by only 6.5 months.

IFP identified distillation as an important reason China has narrowed the gap. Distillation involves training AI systems using outputs generated by more capable models. China has previously dismissed as groundless US allegations that developers in China conducted "aggressive, malicious" AI distillation. IFP urged US policymakers to tighten controls on overseas cloud access, smuggled chips and legal exports.

The Institute for AI Policy and Strategy said in a separate report published on Tuesday that foreign entities prohibited from purchasing US-designed hardware can still legally rent US-designed computing capacity. The Washington-based group urged the US Bureau of Industry and Security to limit China's access to cloud capacity sufficient to train frontier AI models or operate those models at scale. The group also recommended stronger verification of customers' identities and ownership.

Chip exports remain subject to licensing requirements. Washington moved in January to permit individual export reviews for Nvidia's H200, Advanced Micro Devices' MI325X and similar processors, subject to security conditions. US exporters must demonstrate that sales would not reduce supplies available to US customers. Buyers in China must establish formal export-compliance procedures. Those buyers must also submit chips to independent testing in the United States.

Major technology companies in China are working to strengthen domestic alternatives. Shenzhen-based Huawei Technologies said last month that the Ascend 960DT, a chip intended for AI model training, would be available in the first quarter of 2027. Huawei's announced availability date is three quarters earlier than originally scheduled. Huawei said deployments of the Atlas 900 A3 SuperPoD had exceeded 1,000 units. Huawei also said the Atlas 950 system had reached large-scale commercial use. These cluster architectures connect thousands of processors to perform demanding AI tasks.

However, expanding domestic hardware production remains difficult. San Francisco-based research group Epoch AI identified high-bandwidth memory as the principal limit on Huawei's production in an updated report last month. High-bandwidth memory supplies data to AI chips. Epoch AI estimated that Huawei's chip production this year would provide less than 4% of the computing capacity represented by Nvidia's worldwide shipments.