China's AI-compute monitoring coverage falls from 72% to 60% as capacity reaches 2.45M PFLOPS

China's latest national data show 1.45 million PFLOPS of FP16 intelligent computing capacity under the national monitoring and dispatch platform, about 60% of the country's 2.45 million PFLOPS total at the end of July. The figure was presented as progress in building a unified national compute network. Compared with the previous official snapshot, it also reveals a harder implementation problem.

A June report on the Digital China Summit said 1.37 million PFLOPS had been incorporated into the monitoring system, then about 72% of national intelligent-compute capacity. On the face of those two official snapshots, monitored capacity increased by about 80,000 PFLOPS, or 5.8%, while its share of the national total fell by roughly 12 percentage points.

The implied earlier denominator was about 1.90 million PFLOPS if the 1.37 million figure and 72% share used the same FP16 basis. Against 2.45 million PFLOPS at end-July, that would imply total intelligent capacity grew by roughly 29% while monitored capacity grew less than 6%. The dates and reporting definitions are not fully reconciled in the public releases, so the calculation should be treated as a consistency test rather than a precise growth series. The comparison still reveals an implementation gap: calling 60% a new coverage milestone obscures the fact that the published share has fallen sharply.

The denominator is expanding faster than the control layer

The figures point to an implementation gap: infrastructure construction is moving faster than the monitoring layer that is supposed to make capacity visible and eventually dispatchable across regions. The national compute-grid project is still advancing, but the monitoring layer is behind the infrastructure build-out.

Visibility matters because a scheduler cannot route work to capacity it cannot observe. The national platform is intended to standardize data on compute resources, availability and utilization across heterogeneous facilities, then support cross-regional allocation. Eight national computing hubs and three compute-electricity coordination zones now account for more than 85% of installed intelligent capacity, giving the state a concentrated physical base on which to build that coordination layer.

The September 3 connection of Ningxia's provincial platform advances implementation. The local system can transmit monitoring data to the national platform in real time, and its scheduling module has begun testing against the national network. The official description still places parts of that function in technical validation. Monitored PFLOPS and commercially dispatched PFLOPS are different objects.

The infrastructure economics begin at that gap. A common monitoring layer can improve information about idle capacity; it creates value only when customers can move workloads across regions at usable latency, reliability and price. Otherwise, China can continue adding data centers while still duplicating capacity near demand centers and leaving western facilities underused.

The latest official record is more revealing when the 60% share is read against its prior denominator. China is adding monitored compute in absolute terms, and provincial integration is progressing. But the country's AI-capacity build-out appears to be expanding faster than the national platform's visibility. The economic case for the compute grid now rests on closing that coverage gap and turning the monitored portion into repeatable, priced cross-regional dispatch.

Sources

• National Data Administration, August 29, 2026: https://www.nda.gov.cn/sjj/swdt/mtsy/0829/20260829153449424083118_pc.html
• National Data Administration, Ningxia connection, September 3, 2026: https://www.nda.gov.cn/sjj/swdt/dfdt/0903/20260903131554193890680_pc.html
• Digital China Summit prior monitoring baseline, June 11, 2026: https://www.szzg.gov.cn/2026/english/dn/202606/t20260611_5333096.htm