Cheap Power, Not Magic: How China’s Energy Advantage Is Driving Global AI Usage

A numbers-first surprise

It has been reported that OpenRouter’s February 2026 data shows Chinese models now account for roughly 61% of global token calls, with a weekly peak of 5.16 trillion tokens and three-week growth of 127% — while U.S. models have been pushed to about 39% share. Is this a sudden algorithmic masterstroke? Reportedly, no. The dominant explanation is blunt: electricity. Cheap, abundant power has turned Chinese tokens into a global hard currency for AI developers.

Energy as the industry’s hidden metric

Industry estimates cited in the report put inference-stage electricity at 60–70% of operating cost; hardware, bandwidth and labor make up the rest. Training a 100-billion-parameter model reportedly consumes more than 50 million kWh in a single run. Every million tokens, the analysis says, costs roughly 0.8–1.2 kWh to generate — token economics measured in kilowatt-hours. It has been reported that industrial power in China runs about 0.48–0.61 RMB/kWh nationwide while green-power contract prices in western compute hubs fall as low as 0.13–0.3 RMB/kWh, a gap that reportedly translates into Chinese API pricing of roughly $0.3–0.5 per million tokens in and $1–1.2 out versus U.S. rivals charging multiples of that. PUEs below 1.1 and the state-led “East Data West Compute” (东数西算) routing of western renewables to compute clusters help explain why cost, not just code, is now the moat.

Market and geopolitical consequences

Developers follow price. It has been reported that as many as 80% of some U.S. AI startups prefer lower-cost Chinese models for routine development because the savings — sometimes cited as millions of dollars annually — buy faster iteration and broader testing. Geopolitically this complicates the narrative: chips can be sanctioned or limited by trade policy, but an integrated low-cost power system is a strategic industrial advantage that can’t be exported or easily copied. Western incumbents are responding — building captive generation and signing long-term contracts — but grid constraints, long permitting cycles and volatile fuel markets make catch-up costly and slow.

China’s models are not necessarily ahead on every technical metric; the report acknowledges many flagship products still lag leading U.S. systems by months in some areas. The crucial point is practical: for broad developer adoption, “good enough” at a fraction of the price is often decisive. Energy, not a single algorithmic breakthrough, is the overlooked variable reshaping who gets to write and run large models at scale.