China's Computing Power for AI Could Double by 2027

By the end of 2027, the computing power available to China for artificial intelligence could grow from 4.6 million to 8.8 million Nvidia H100 equivalents. This forecast is contained in a report by the Institute for Progress (IFP), cited by the South China Morning Post. Researchers estimate that about 65% of the resources will be accounted for by American-developed chips, including Nvidia and AMD. The calculations take into account Chinese processors, legal imports, and the use of foreign cloud platforms. The latter channel is expected to increase from 1 million H100 equivalents by the end of 2026 to 2.8 million a year later.

Despite the persistent infrastructure gap, Chinese AI models are gradually closing the gap with American ones. According to IFP, the US has approximately 16 times more computing resources for training such systems, but the difference in capabilities of leading models is about 6.5 months. One of the factors contributing to the convergence, researchers say, is distillation, where less powerful systems are trained on the responses of more advanced models. Against this backdrop, IFP has proposed strengthening control over the export of chips, illegal supplies, and the rental of foreign cloud resources. A separate study by the Institute for AI Policy and Strategy notes that organizations prohibited from directly purchasing American equipment can gain access to computing power through cloud services. Among the proposed measures– checking clients and ownership structures, as well as restrictions on the use of resources for training and large-scale deployment of advanced models.

Chinese companies are simultaneously developing their own hardware infrastructure. In September, Huawei announced that its Ascend 960DT accelerator for AI training will be available in the first quarter of 2027, three quarters ahead of the original schedule. The company also announced the deployment of more than 1,000 Atlas 900 A3 SuperPod clusters and the start of large-scale commercial use of the Atlas 950. However, according to Epoch AI, Huawei's production expansion is constrained by a shortage of HBM memory. It is expected that the company's computing power supply in 2026 will account for less than 4% of Nvidia's global figure. Thus, the growth of resources available to China will depend not only on the development of its own chips but also on access to foreign computing systems.