Jensen Huang's claim that China will win the AI race—despite being starved of advanced chips by US export controls—hinges on a single metric: energy. The NVIDIA CEO argues that as chip improvements plateau, the constraint shifting from hardware to power supply. And by that measure, the gap is already enormous.
The numbers tell the story. In 2024, the US had roughly 750 TWh of annual renewable capacity (solar and wind combined), while China approached 2,000 TWh. More striking: the capacity China added in 2024 alone exceeded America's entire renewable infrastructure. Between 2000 and 2015, the two countries tracked broadly in parallel. Then China executed an exponential curve while the US plodded linearly. The divergence reflects not just investment, but coordination. China deploys Power Purchase Agreements binding renewable output directly to AI labs, securing cheap, dedicated power. Meanwhile, the Trump administration is cutting state subsidies for renewable developers, and US electricity prices near data centers have soared 270% in five years—a direct tax on compute that Chinese operators simply don't face.
The implication ripples through the entire economics of AI training and inference. OpenAI's revenue projections assume explosive scaling, but operating costs are collapsing into electricity bills. A text query costs the equivalent energy of an eight-second microwave run; generating a video consumes nearly an hour. Heavy users likely spend more on electricity than their $20 monthly subscription. The arbitrage is disappearing. Meanwhile, data centers in America are already operating below capacity—not from lack of demand, but from lack of power. Those running at partial load burn dirtier fuel. The US grid powering AI is 48% more carbon-intensive than the national average, and it's getting worse, not better.
China's bet is structural and patient. Alibaba and Tencent may lag on chip architecture today, but energy constraints will flatten that advantage as models compete on inference scale, not just parameter count. The US response—regulatory restrictions on chip sales—buys time only if you believe the bottleneck stays in silicon. If Huang is right, that clock is running out.