China's Real AI Advantage Is Electricity

What actually caps how much AI a country can build? Electricity. AI compute is, at root, a power problem, and China has far more power than the US. China already generates more than twice as much electricity as the US, and it has grown total generation by nearly 6% a year over the past decade, with over half of that growth coming from clean sources. That's Brookings, working from Ember data. In April 2025, China's monthly solar generation hit a record 96 TWh, comparable to a full month of output from the entire US nuclear fleet, which runs at roughly 65 TWh a month. The US still leads on compute and on data-centre power use today. But energy is the long-game input, and it's the hardest advantage for the US to close. This is the under-covered structural story of the AI race.

Key Takeaways- AI compute is ultimately an energy problem. The side with cheaper, more abundant electricity can run more of it for longer.- China generates more than twice the electricity of the US, and has grown total generation by nearly 6% a year over the past decade, over half of it from clean sources.- In April 2025, China's monthly solar output hit a record 96 TWh, comparable to a full month of the entire US nuclear fleet, roughly 65 TWh.- The US leads compute today by roughly 5:1, hosting about three-quarters of tracked AI-supercomputer performance as of May 2025, so energy is the long game, not the current scoreboard.- A grid lead is far harder to copy than a model lead. It's built from years of construction, not a quarter of training.

AI is an energy problem

AI capability is bounded by electricity, not just by clever code. Data centres worldwide consumed around 415 TWh in 2024, and that figure is climbing fast as AI workloads scale. Every training run and every inference call is electricity converted into intelligence. When power is the binding constraint, whoever has more of it can simply do more AI.

It helps to be blunt about what a GPU cluster is. It's a machine for turning megawatt-hours into model weights and tokens. The chips get the headlines; they are inert without the power to run them and the power to cool them. As clusters grow, the limiting factor stops being how many accelerators you can buy and starts being how many megawatts you can deliver to a single site, reliably, at three in the morning.

Which is why the hyperscaler conversation has quietly moved off chips and onto grids. In the US, power is now the harder thing to secure. You can order accelerators. You cannot order a substation, a transmission line or a power station on anything like the same timescale.

Here's the framing most coverage misses. A model lead is a software advantage, and software advantages decay in months. An energy lead is an infrastructure advantage, and infrastructure advantages compound over decades. If AI's appetite for power keeps growing, the country pouring the most cheap electricity into compute holds the input that every other layer sits on top of.

AI compute is at root an energy problem: global data centres used roughly 415 TWh in 2024 and demand is rising sharply with AI. Each training run and inference call is electricity turned into intelligence, so the country with cheaper, more abundant power can ultimately run more AI for longer than its rival.

The pillar this piece hangs off asks, layer by layer, whether China is winning the AI race. The cost story sitting next to it is why Chinese AI is so cheap.

China's AI energy advantage in electricity

China's electricity lead over the US is large, growing, and increasingly clean. The more-than-twice gap isn't a marginal edge. It's a structural gap in the raw input AI runs on, and BloombergNEF expects China to add more than six times as much generation capacity as the US over the next five years. In 2025 alone, China added over 430 GW of wind and solar capacity, more than half of all the renewable capacity added worldwide that year.

The solar number is the one that stops people short. In April 2025, China's monthly solar generation hit a record 96 TWh, comparable to the roughly 65 TWh a month the entire US nuclear fleet produces. One slice of China's clean-energy build-out now rivals a pillar of American baseload.

None of this reads as abstract when you live here. Drive an hour west out of Hangzhou, past the tea terraces, and the factory roofs along the road are blue with panels. Take the fast train in any direction and the pylons keep pace with the carriage for most of the journey. The labs behind DeepSeek and Qwen sit in my city, and the electricity their training runs eat is being manufactured, at scale and at a falling price, by the same country. Solar and storage keep getting cheaper as China builds more of both, so each new gigawatt tends to land cheaper than the last. That's a flywheel, not a one-off.

Sources: Ember (China solar record, April 2025); EIA (US nuclear monthly output). Total Chinese generation is more than double the US figure.

China generates more than twice the electricity of the US, growing nearly 6% a year for a decade with over half of that growth from clean sources. In April 2025 its monthly solar output hit a record 96 TWh, comparable to a full month of the entire US nuclear fleet. On the raw input AI depends on, China already leads.

The compute caveat

The energy lead is not the whole picture, and I'd rather say so myself than let you find the hole. The US still holds a roughly 5:1 lead in tracked compute: as of May 2025 it hosted about three-quarters of global AI-supercomputer performance, around 74.5% to China's 14.1%. The capability gap is narrower than that and closing. Epoch's running index puts the average US lead since 2023 at about 7 months, and Stanford's 2026 AI Index had it down to under 3% by March 2026. On data centres specifically, the US consumed most of the world's data-centre power in 2024, around 45% of that 415 TWh against China's roughly 25%.

So why call energy the advantage when the US leads on compute and on current power draw? Because those are snapshots of a scoreboard, and energy is a slope. Compute leads can be bought quickly, with capital and chips. The US spent both, surged ahead, and the lead it holds right now is real.

What changes as the race runs on is the binding constraint. Grid capacity to keep feeding ever-larger clusters is the one thing you cannot buy on a short timescale. US data-centre power consumption leads today, while China's capacity is expected to nearly double to 60 GW by 2030. I'll keep that forward claim hedged, because nobody knows the exact path. The direction is what I'd bet on.

The honest way to hold both truths is to separate the lap from the race. A 5:1 compute lead is a commanding position in the lap being run now. It tells you very little about who can keep the lights on for a much bigger fleet in 2030. Energy isn't the scoreboard. It's the input that decides how long any compute lead stays affordable to hold.

The US leads tracked AI compute today by roughly 5:1, and consumed around 45% of the world's data-centre power in 2024 versus China's 25%. So energy is not today's scoreboard. It's the long-game input that determines how sustainable a compute lead is as AI's power demand keeps climbing.

Why this advantage is hard to copy

An energy lead sticks in a way a model lead never does, because it's made of concrete and copper rather than code. Close a capability gap and you need compute, talent and a few months. Nobody conjures a doubled national grid or a solar manufacturing base in that window. China's edge here was poured, quite literally, across years of construction and industrial policy, as the Brookings exchange on powering the AI race lays out.

Start with the grid. Connecting new generation means transmission lines, substations and permitting, and all of that moves at the speed of construction and regulation, not software. The US has genuine bottlenecks: interconnection queues alone can outlast a model generation, even where the will and the capital are sitting ready. China has been adding capacity at pace for a decade, and a decade of head start doesn't erase.

Underneath the grid sits the manufacturing layer, which is the part people skip. China makes the solar panels and the batteries, the components that make clean power cheap to add in the first place. Each new gigawatt of Chinese clean capacity tends to cost less than the one before it, while the US often pays more to build the same thing. Whoever makes the hardware owns the cost curve.

This is the most underrated story in the AI race, and I don't say that lightly. A model lead is a sprint advantage. A grid-and-manufacturing lead reinforces itself: cheaper panels mean more capacity, more capacity means lower power costs, lower power costs mean cheaper compute, and round it goes again. The US can out-spend China on chips whenever it chooses to. Out-building it on the physics of cheap, abundant electricity is a decade-scale project.

An energy lead is harder to copy than a model lead because it's physical infrastructure, not software. China's grid build-out and its dominance in solar and battery manufacturing let it add clean capacity faster and cheaper than the US. Closing that gap is a decade-scale project, not a quarter of training.

What it means for the AI race

Energy turns "who's winning" from a snapshot into a slope. The US leads the frontier today and holds that 5:1 compute lead. But if AI's power demand keeps rising while China keeps adding cheap, clean generation faster than the US, the structural advantage tilts towards whoever can keep feeding the machine.

If you're a Western strategist, the practical instruction is to watch a different indicator. Benchmark scores and release notes dominate the headlines, and they measure the lap. Grid build-out, clean-energy share and the cost of new capacity tell you who can sustain the race. They move slowly and rarely make the news, which is precisely why they're underpriced.

When I sit with UK firms and they ask how long a Chinese model advantage is likely to last, this is the point I keep returning to. Any capability edge you build on can evaporate the week the next release lands. The forces underneath it (cost, openness, and the power that makes inference cheap) run on a much slower clock. China's energy position is a large part of why I expect cheap, open Chinese models to keep getting cheaper rather than dearer, and why I'm comfortable telling a client to plan on it.

Energy turns the AI race from a snapshot into a slope. The US leads compute today by about 5:1, but China's faster, cheaper, cleaner grid build-out positions it for the long game. Watch grid capacity and clean-energy share, not just benchmarks, to judge who can sustain the race.

FAQ

Does China have more electricity than the US? Yes, and by a wide margin. China generates more than twice the electricity the US does, and has grown total generation by nearly 6% a year over the past decade, with over half of that growth from clean sources. In April 2025 its monthly solar output hit a record 96 TWh, comparable to a full month of the entire US nuclear fleet.

Why does AI need so much energy? AI compute is electricity converted into intelligence. Training and running large models means powering and cooling huge GPU clusters around the clock. Global data centres used roughly 415 TWh in 2024, and demand is rising fast as AI scales. More AI means more power, directly.

Is energy a bottleneck for AI? Increasingly, yes. As clusters grow, the limiting factor shifts from buying chips to delivering reliable power to one site. The US leads data-centre power use today, around 45% of global consumption in 2024, but securing new grid capacity is becoming the harder problem.

Does cheap energy give China an AI advantage? It gives China a durable long-game advantage. The US still leads tracked compute today by about 5:1. But cheap, abundant, clean power lowers the cost floor for running compute, and that grid lead is far harder for the US to close than a model gap.

The bottom line

Most coverage of the AI race fixates on models and benchmarks. The story underneath all of it is electricity, and it goes largely uncovered. AI compute is, in the end, an energy problem, and China generates more than twice the power the US does, has grown total generation by nearly 6% a year over the past decade with over half of that growth from clean sources, and its monthly solar now rivals the entire US nuclear fleet. The US leads compute today by roughly 5:1, and that lead is real. But energy is the long game, and a grid built over a decade is far harder to copy than a model trained in a quarter.

So watch the right numbers. Grid capacity, clean-energy share and the cost of new power tell you who can sustain the race, rather than who happens to be ahead this lap.

If you want the full verdict, start with is China winning the AI race, then read why Chinese AI is so cheap.