China Pays $0.05/kWh for AI Power. The US Pays $0.40/kWh. The Energy War Has a Clear Leader.

Everyone is watching the chip war. TSMC export controls, Nvidia H100 restrictions, CXMT HBM3 timelines. Almost nobody is watching the power war. China pays $0.05 per kilowatt-hour for data center electricity in its western provinces. The United States pays $0.40 per kilowatt-hour in many regions. That eight-times gap in energy cost is a structural AI infrastructure advantage that no export control can address.

The $0.05 vs $0.40 Number That Defines the Race

In China's western provinces — Ningxia, Gansu, Inner Mongolia — industrial electricity for data centers costs approximately $0.05/kWh. This is not a temporary subsidy or a politically inflated figure. It reflects the combination of abundant hydroelectric capacity in Yunnan and Sichuan, massive wind and solar buildout in the northwest, and an industrial pricing framework that prioritises high-volume power consumers like data centers.

In the United States, data center electricity costs vary by region but average $0.08/kWh in the cheapest markets (parts of Texas and the Pacific Northwest) and reach $0.40/kWh in constrained markets like Virginia, where roughly 70% of US data center capacity is concentrated. The national average for commercial electricity is approximately $0.13-0.15/kWh, but hyperscale facilities negotiating directly with utilities in competitive markets typically pay $0.06-0.12/kWh.

For a hyperscale AI training cluster consuming 100 megawatts continuously, the annual electricity bill difference between China's western province pricing and a mid-tier US market is approximately $280 million per year. At that scale, energy cost is not a line item — it is the dominant infrastructure variable.

How China Built This Advantage

China's energy advantage in AI compute is not accidental. It is the result of infrastructure investment decisions made between 2010 and 2023 that the US did not match.

China generates over twice as much electricity as the United States. Total Chinese electricity generation in 2025 was approximately 9,500 TWh, compared to roughly 4,300 TWh for the US. China has been building generation capacity — coal, hydro, solar, wind, and increasingly nuclear — at a pace no other country has matched.

The data center buildout followed the power buildout. China's number of data center racks grew at 30% annually from 2016 to 2023. The country constructed data centers in provinces where power is cheapest — the northwest and southwest — and built long-distance ultra-high-voltage transmission lines to move that power where it is needed. The result is a data center infrastructure co-located with cheap generation, rather than a data center ecosystem that competes with residential and commercial demand for scarce grid capacity.

The AI boom did not create China's energy advantage. It revealed it.

China at 400 GW Spare, US at 45 GW Short

The divergence between the two countries on data center power capacity will become critical before the end of the decade.

By 2030, China is projected to have approximately 400 gigawatts of spare electricity generation capacity that can be directed toward data centers — equal to three times the entire global data center power demand projected for that year. This is not theoretical capacity. It is generation infrastructure already under construction or permitted.

The United States faces the opposite trajectory. US data centers are expected to encounter a 45 gigawatt power shortage by 2028 — a gap between projected AI infrastructure demand and available grid capacity. OpenAI has specifically warned that this power shortage could slow American AI development if it is not addressed. The constraint is not chip availability. It is power delivery.

The reasons for the US shortage are structural: permitting timelines for new generation capacity average 4-7 years, transmission line construction is similarly constrained, and data center demand has accelerated faster than infrastructure investment in a regulatory environment that was not designed for this pace of change.

What This Means for AI Training and Inference Economics

The energy cost gap translates directly into AI compute economics at every stage of the development pipeline.

Training: Large model training runs consume tens to hundreds of megawatts for weeks or months. The Llama 4 training run reportedly consumed approximately 100 MW over a sustained period. At Chinese western province pricing ($0.05/kWh), a 30-day 100 MW training run costs approximately $3.6 million in electricity. At a mid-tier US commercial rate ($0.15/kWh), the same run costs approximately $10.8 million. A US hyperscale negotiated rate ($0.08/kWh) would be $5.8 million. China's state-supported labs pay less than a third of what US hyperscalers pay even at negotiated rates.

Inference: Every query to a deployed AI model costs electricity. At scale — ChatGPT processes approximately 100 million queries per day — inference electricity costs are material. A Chinese model serving equivalent query volume pays dramatically less per query in infrastructure cost, enabling lower subscription pricing, more aggressive API pricing, or higher margins. This is not hypothetical. It is why Chinese AI API prices have consistently undercut US competitors and why that pricing gap is sustainable.

Why the US Cannot Close This Gap Quickly

The US response to the power shortage is underway but structurally slow. The CHIPS Act and related infrastructure legislation have accelerated data center construction. Several US utilities have approved large capacity additions. Nuclear power is back on the agenda, with Microsoft and Google both signing nuclear power purchase agreements.

None of these responses solve the problem before 2028-2030. Permitting a new nuclear plant takes 10-15 years in the US. Even natural gas peakers, which can be built in 2-3 years, face grid interconnection queues of several years in constrained markets. The data center demand surge triggered by AI is outpacing the physical infrastructure construction cycle.

China's advantage in this window — roughly 2025 to 2030 — is real and not addressable through export controls. You can restrict Nvidia H100 shipments to China. You cannot restrict China's access to wind capacity in Gansu or hydropower in Yunnan.

Our Analysis: The Chip War vs the Power War

The US AI policy debate in 2026 is almost entirely focused on the chip supply chain: TSMC advanced node access, Nvidia export controls, China's CXMT HBM3 timeline, and the BIS entity list. These are legitimate concerns. But they address the supply of computation hardware, not the cost of running it.

A 1% reduction in electricity cost across a large AI training cluster saves more money than a 5% improvement in chip efficiency. At the scale of national AI infrastructure — hundreds of gigawatts over years — energy economics dominate chip economics. China has understood this for longer than the US has.

The risk for US AI competitiveness is not that China gets access to NVIDIA H100s through third parties. It is that China builds out the power infrastructure to run whatever chips it has — domestic or otherwise — at a cost structure that US companies cannot match without equivalent energy investment. The chip war needs to be won, but the power war is the one that determines the long-term outcome.

See our global semiconductor market 2026 report and CXMT memory chip breakout analysis for the chip side of this picture.

Key Takeaways

• China pays $0.05/kWh for AI data center electricity in western provinces; the US pays $0.40/kWh in many markets — an 8x cost gap that translates to hundreds of millions in annual infrastructure savings at hyperscale
• China generates 2x more electricity than the US and is widening the gap through aggressive investment in hydro, solar, wind, and nuclear generation
• China has 400 GW of spare data center capacity projected by 2030 — three times what the entire world needs for AI data centers at that date
• The US faces a 45 GW data center power shortage by 2028 — OpenAI has flagged this directly as a threat to American AI leadership
• Chinese data center racks grew 30% annually from 2016 to 2023 — the infrastructure buildout preceded the AI boom and is why China was positioned to absorb it
• Export controls cannot address this gap: you can restrict chip shipments, you cannot restrict access to Ningxia wind power or Sichuan hydroelectricity
• The 2025-2030 window is decisive: US infrastructure investment is accelerating but permitting timelines mean the power shortage will persist through the period when frontier AI infrastructure decisions are being made

Sources

• Al Jazeera — China's secret weapon in the AI race: lots of cheap energy
• South China Morning Post — Energy crisis showcases strengths of China's data centre market
• Fortune — China could be the big winner in the AI race
• Tom's Hardware — AI costs spike as subscriptions hit pricing wall — firms turn to Chinese LLMs