China Plans to Quadruple Its AI Computing by 2030

China has set out a plan to more than quadruple its artificial intelligence computing capacity by 2030, part of a wider push to reduce its dependence on foreign chips. The target appears in the country's 15th Five Year Plan for the information and communications industry, published on 7 September and covering the years 2026 to 2030.

The Ministry of Industry and Information Technology set a goal of 9,800 exaflops of intelligent computing capacity by the end of the decade, according to the South China Morning Post. Reaching it would require cumulative investment of 3.8 trillion yuan, about 532 billion US dollars, in information infrastructure over the five year period.

The starting point

China's intelligent computing capacity stood at 2,185 exaflops at the end of June, measured at the lower FP16 precision, up 177 percent from a year earlier. The country had already built 52 computing facilities each fitted with more than 10,000 accelerator cards, as reported by Unite.AI. Hitting 9,800 exaflops means building more than four times that base in under five years.

The plan calls for the orderly rollout of large clusters, some with 10,000 graphics cards and some with 100,000 or more, along with separate facilities tuned for inference, the day to day work of running models rather than training them.

The chip problem

The ambition runs straight into American export controls, which have cut China off from the most advanced Nvidia hardware. That has pushed Chinese builders toward domestic accelerators, including Huawei's Ascend line, which recently landed a 160,000 chip order from DeepSeek. Whether homegrown silicon can deliver the efficiency these targets assume is still an open question.

There is also the matter of whether the compute gets used. Analysts have noted that much of China's existing capacity sits idle, stranded by patchy demand, uneven quality and the difficulty of stitching scattered facilities into something usable. A headline number measured in exaflops is easy to announce. Turning it into working infrastructure that companies actually rent and run is the harder task, and the one the next five years will test.