China's grid is starting to pick where AI lives

China's AI buildout is reorganizing the grid around itself. Power-hungry training and batch workloads are migrating toward the renewable-rich western provinces: Inner Mongolia, Ningxia, Gansu, Guizhou, while latency-sensitive inference and financial services stay anchored in the eastern coastal clusters. The shift reframes the grid from passive infrastructure into an active siting constraint: a new data center picks where the cheapest, cleanest electrons can be delivered at scale, not where talent or fiber already cluster.

Storage and intelligent workload scheduling turn that constraint into a tool, steering non-urgent compute toward hours of surplus wind and solar. Latency-tolerant work flows to cheap, clean power. Latency-bound work stays near users. The grid stops serving the data center. The data center starts serving the grid.

Wood Mackenzie's August projection puts a number on the load: 774 terawatt-hours by 2030, roughly the annual electricity use of a mid-sized industrial economy, lifting data centers to about 6% of national demand, with a possible 17% by 2060. The same research expects the eight national hubs to stay dominant through 2060, with renewable-rich provinces gradually absorbing more power-intensive work. The map redraws, but the eastern clusters hold.

The mechanism travels. Wherever clean power is cheap and grid-connected, latency-tolerant compute will follow. Whoever builds the scheduling layer that pairs workloads to electrons owns the next siting decision.