China's Ministry of Industry and Information Technology published the 15th Five-Year Plan for the information and communications industry, targeting 9,800 exaflops of intelligent computing capacity by 2030 and calling for CNY 3.8 trillion, about $532 billion, in cumulative information infrastructure investment across 2026 to 2030, per the South China Morning Post. National capacity reached 2,185 exaflops at the end of June, up 177% year over year, across 52 facilities each holding more than 10,000 accelerator cards. The plan directs orderly deployment of clusters at the 10,000-card and 100,000-card scale alongside inference facilities tuned to specific applications. The document was posted publicly on September 7 and is dated August 12.
The number to hold is the growth rate rather than the target. Capacity that grew 177% in a year needs roughly 35% compounded annually to reach 9,800 exaflops by 2030, which is a deceleration written into the plan itself. That is a realistic document rather than an aspirational one, and it implies Beijing expects the constraint to tighten rather than ease. The obvious candidate for that constraint is silicon supply under export controls, which is also why the plan reads as a domestic-fabrication commitment as much as a capacity commitment.
The distinction between training clusters and inference facilities matters more than it looks. Ten-thousand-card and hundred-thousand-card clusters are training infrastructure, and building them requires the highest-end accelerators, which are exactly what controls restrict. Application-tuned inference capacity can run on older or domestically produced parts, and it is where economic returns actually show up. A plan that funds both is hedging: keep frontier training viable if supply allows, and make sure the deployed economy runs regardless.
For anyone operating across both markets the practical read is that Chinese compute is being planned as public infrastructure on a fixed schedule with published targets, while American capacity is being built by private balance sheets against contracted demand. Those two systems fail differently. One is exposed to a policy change and a fabrication timeline, the other to a demand curve and GPU residual values. Neither has been tested through a downturn.
Bottom Line
The 2030 target implies about 35% compounded growth from a base that just grew 177%, which is Beijing planning around a supply constraint rather than through it. Watch the split between training clusters and inference facilities, because that ratio is the clearest public signal of how much frontier silicon China expects to get.