What are the key dynamics and concerns surrounding China's data centre expansion, particularly regarding localised AI infrastructure development, regional competition, and comparisons to the United States?

Overview

China is in the midst of a state-driven, massive-scale data centre buildout to close the AI infrastructure gap with the United States. While the US still leads in raw facility count and total capital spending, China is deploying faster construction timelines, cheaper energy, and a centrally planned "Eastern Data, Western Computing" strategy. However, the rapid expansion has generated serious concerns around regional overbuilding, grid strain, speculative waste, and underutilisation of computing power.

---

Key dynamics of China's data centre expansion

1. Massive state-backed investment

• In June 2026, China announced plans to spend ~2 trillion yuan ($295 billion) over five years to build a nationwide network of data centres, aimed at surpassing the US in AI .
• China's data centre market had ~37 GW of installed capacity as of 2025, roughly 27% of global capacity .
• The country's AI computing market (data centre operations, AI computing services, and domestic AI chips) reached roughly ¥500 billion (RMB) in 2025 .
• Cloud infrastructure spending in mainland China rose 26% year-on-year in Q4 2025, reaching $14.7 billion .

2. "Eastern Data, Western Computing" (EDWC) strategy

• This flagship policy, launched in 2022 and reaffirmed in the 15th Five-Year Plan (March 2026), directs compute-intensive data centres to western China where land, renewable energy, and cooling are cheaper .
• Eight national computing hubs and 10 data centre clusters have been established .
• The goal is to balance regional development, reduce energy costs in the east, and absorb surplus renewables capacity in the west .
• Concern: Despite the policy, actual AI chip deployment remains heavily concentrated in eastern provinces (Beijing, Shanghai, Guangdong), leading some analysts to call EDWC "fake" in practice .

3. Regional competition and local government overbuilding

• Provincial and local governments have aggressively competed for data centre projects, offering subsidies and fast approvals, which has led to speculative overbuilding .
• By mid-2025, over 100 planned data centre projects had been withdrawn due to grid instability and oversupply .
• China is now creating a national marketplace to sell surplus computing power, acknowledging the glut .
• Underutilisation of AI computing capacity is a growing problem: abundant cheap power alone has not created a unified, operable computing network .

4. Energy advantage: the "electron gap"

• China has a significant energy advantage over the US for powering AI data centres, a phenomenon called the "electron gap" .
• China builds power infrastructure much faster than the US and has excess renewable capacity (solar, wind, hydro) in western regions .
• The US faces an energy bottleneck — Wood Mackenzie reported a 50% decline in new US data centre projects at end-2025 due to grid constraints .
• Caveat: China's power system has its own institutional bottlenecks — rigid pricing, grid inefficiencies, and inconsistent peak loads that make data centres less compatible with green energy suppliers .

5. Localised AI infrastructure and chip constraints

• US export controls on advanced AI semiconductors (Nvidia H100/H200, etc.) have forced China to accelerate domestic chip development .
• China's data centre market now includes roughly ¥100 billion (RMB) in domestic AI chips, with companies like Alibaba developing their own AI processors .
• However, performance gaps remain: Chinese AI chips lag behind cutting-edge US equivalents, and chip compatibility issues complicate the national computing network .
• Intelligent computing capacity within the eight national hubs has grown rapidly, but much of it relies on less advanced hardware .

---

Comparison: China vs. United States

Dimension	United States	China
Number of data centres (2025-2026)	~4,280–5,427 (largest globally)	~369–449 (3rd–4th globally)
AI capex (2025 est.)	At least $350B (Microsoft, Amazon, Meta, Google)	Less than $40B (USCC estimate)
Planned investment (2026-2031)	~$50B/month in construction spending (April 2026)	~$295B over 5 years ($59B/year)
AI model performance gap	Leads in frontier capability	Has "effectively closed" the gap on benchmark models
Energy advantage	Grid bottlenecks slowing new builds	Cheaper, faster-to-deploy energy, but grid inefficiencies remain
Construction speed	~3 years per facility	Weeks to months (fast approval, centralised planning)
Cost efficiency	More cost-efficient data centre construction	Higher construction costs but cheaper energy and labour
Chip access	Cutting-edge (Nvidia, AMD,自家)	Export-restricted; domestic chips improving but trailing

---

Main concerns and risks

• Overcapacity and wasted investment: Local government competition has created a data centre glut, with many facilities underused and over 100 projects cancelled .
• Grid and energy integration challenges: China's push to connect data centres with renewable energy faces technical hurdles — inconsistent peak loads and forecasting difficulties .
• Regional imbalance: EDWC has not fully shifted compute to the west; most AI computing power remains concentrated in wealthy eastern provinces .
• Chip dependency: US export controls constrain China's access to the most advanced AI GPUs, limiting raw compute performance even as model efficiency improves .
• Environmental cost: Rapid data centre expansion increases power demand and carbon footprint, though China is experimenting with using data centres to absorb renewables oversupply .

---

Bottom line

China's data centre expansion is enormous in scale and speed, driven by central government planning and a genuine energy-cost advantage. The US still leads in facility count, total AI capex, and chip access. But China's structural advantages — faster construction, cheaper power, and a unified national strategy — are narrowing the gap. The biggest risks are on the downside: overbuilding, regional policy misalignment, chip bottlenecks, and grid integration problems that could turn the world's largest planned AI infrastructure buildout into a costly case of speculative excess.