East Data West Computing

China's national AI-infrastructure placement strategy, formalized in 2022 and accelerated through 2025–2026, that **separates compute placement from data origin** along the country's energy gradient. Coastal hubs (Beijing, Shanghai, Shenzhen, Guangzhou) generate most of the data and host hot inference; western provinces (Guizhou, Inner Mongolia, Gansu, Ningxia) hold the GPU/Ascend training clusters that consume that data, drawing on **~400 GW of projected spare grid capacity by 2030** at electricity rates as low as **3¢/kWh**. S3-compatible object storage (Aliyun OSS, Tencent COS, Huawei OBS) is the data plane that connects the two halves: write hot in the East, replicate to West, train on cheap power, replicate inference artifacts back East.

Three forces converge on this design. (1) US semiconductor export controls cap Chinese access to top-tier silicon, forcing Chinese labs to compensate with **raw scale and cheap electricity** — which only the western provinces have. (2) Coastal grids cannot absorb hyperscale training loads without displacing residential and industrial demand. (3) National data-localization law forbids cross-border egress, so the entire two-region pattern must close inside China. The result is a deliberately asymmetric architecture: data-gravity in the East, compute-gravity in the West, with object storage as the high-bandwidth conduit. Compare to the US, where a projected **~44 GW power shortfall by 2030** and 8+-year interconnection queues actively block hyperscale build-out at any latitude.

Training storage and replication for foundation models on Ascend or domestic-supplemented NVIDIA hardware (DeepSeek-V3, GLM-5, Kimi K2, Qwen 3.5 corpora), petabyte-scale cross-region replication driven by power-cost arbitrage rather than disaster recovery, multi-region cold-tier consolidation in low-cost-power regions while serving inference from coastal hot tiers.