The battle for artificial intelligence will not be decided by chips alone. It will be won by whoever can power the most computation at the lowest cost. Jensen Huang’s warning that AI is constrained by electricity, not just silicon, inadvertently exposes Washington’s blind spot: while America fixates on GPUs and export controls, Beijing is building the wattage to run models at scale.
Framing the issue: AI demand meets the grid
Since large-language models went mainstream, the bottleneck has shifted from transistors to electricity. One widely cited study estimates GPT-4 could consume about 463 269 megawatt-hours per year—roughly 44 000 U.S. homes for a year, the load of a ~100 000-person small city. The decisive resource is power that is cheap, abundant and reliable.
Data-centre electricity use is set to more than double by 2030 and reach around 1 800 terawatt-hours by 2040—enough for some 150 million U.S. homes. Whoever secures that power at scale controls the tempo of AI progress.
Within this landscape, China is no longer catching up. In 2024 it added roughly 275–280 GW of new solar, taking total installed solar capacity to nearly 900 GW. Beijing’s ultra-high-voltage (UHV) transmission network moves cheap inland electricity to coastal AI hubs—powering the next phase of its industrial strategy.
Scale: Wind + solar capacity already measures in the multi-hundreds of gigawatts to low terawatt range, with vast projects under construction.
Integration: UHV lines connect Inner Mongolia’s solar and Sichuan’s hydro to coastal demand, cutting curtailment and balancing load.
Policy pricing: Local authorities grant preferential tariffs to AI firms such as Alibaba, Tencent and ByteDance, offsetting the efficiency gap of domestic chips.
Chip performance gains now crawl at single-digit rates, but electricity costs can fall by double digits. What matters is not the fastest transistor—it is how long you can afford to keep it running.
The U.S. strategic blind spot
U.S. policy has treated semiconductors as the core of national security. Yet electricity costs in data-centre regions have surged, while permitting friction slows large-scale renewables. Even with the best chips, expensive power inflates compute costs. Energy sovereignty has become America’s missing flank.
Historical parallels: energy, empire, hegemony
Every technological super-power rode an energy wave—Britain on coal, the U.S. on oil and hydro, China now on renewables and UHV. The axis of power is shifting from who makes the smartest transistor to who controls the cheapest kilowatt-hour.
Weaknesses, vulnerabilities, and the other side’s case
Beijing’s risks: intermittency, storage, coal reliance, and overcapacity in solar manufacturing. Washington’s counter-case: chip architecture, open innovation, and ecosystem depth still matter. Yet only one factor—power price and availability—multiplies all others.
Likely outcome and strategic outlook
Rule: lower power cost → cheaper compute. Application: China’s renewable surge + UHV grid → lower all-in training cost. Consequence: Unless the U.S. pivots to energy sovereignty, the advantage tilts east.
Conclusion
The AI race is not a contest of chips but of watts. China’s grid-industrial complex is fast becoming the engine of digital hegemony. The U.S. must now compete on electricity—generation, transmission and price—or risk discovering that in this race, the winner is whoever keeps the lights on longest.
Analytical key finding
AI progress is increasingly power-limited. China’s rapid expansion of renewables and UHV transmission gives it a structural cost advantage in compute. Unless the U.S. treats energy for AI as national-security infrastructure, its chip dominance will erode under higher cost per computation.
