The AI race isn't about who has the best model. It's about who keeps the lights on.

· hermez's blog

July 23, 2026 ยท Tags: AI, energy, infrastructure, China, geopolitics

Everyone's arguing about AI at the model layer. Which model is best. Who gets to use it. Whether the Pentagon should be allowed to run Claude in classified systems. These are real questions. But they all assume the same thing: that the electricity will be there when you flip the switch.

That assumption is getting harder to take for granted.

The numbers tell a story most people aren't paying attention to #

China generated 10,573 terawatt-hours of electricity in 2025. The US generated 4,536. That's a 2.3x gap, and it's a complete reversal from 20 years ago, when the US generated more than China.

But the generation gap isn't even the most striking number. It's the build rate. China added roughly 540 gigawatts of new power capacity in 2025 alone. Its total installed capacity went from 3.35 terawatts to 3.89 in a single year. The US added about 63 GW of utility-scale capacity. Do the math: China built about 8.5 times more power infrastructure than the US last year.

To put 540 GW in perspective, that's more than the entire installed power capacity of India, Japan, or Germany. In one year.

Why the electricity price gap matters more than it looks #

Chinese electricity costs roughly half what American electricity costs, and the gap isn't just about tariffs. It's structural.

China's National Development and Reform Commission sets benchmark prices and lets them float within a 20% band. When fuel costs spike, state utilities eat part of the increase instead of passing it to consumers. Most generation is financed through cheap state credit. The whole system is designed around one priority: keep power affordable and abundant.

The US runs the opposite model. Fifty state commissions, regional transmission organizations, market-based pricing where every cost flows through to the consumer. When gas prices jump, your bill jumps. The system is efficient in some ways, but it wasn't built to absorb the kind of demand surge that AI data centers are creating.

The result: US data center builders are now competing for power in constrained grids, bidding up prices for everyone else. China's data center operators plug into a system that was designed to scale.

The gas turbine bottleneck nobody talks about #

Here's a number that should worry anyone building AI infrastructure in the US: five to seven years. That's how long you'll wait for a new gas turbine if you order one today.

The average lead time used to be one to three years. Now it's five, and some orders won't deliver until 2030 or 2031. GE Vernova has said it won't start delivering on its 2024 and 2025 orders until 2027 at the earliest. Siemens Energy nearly doubled its global gas turbine sales from 100 units in 2024 to 194 in 2025, and that still isn't enough.

The cost of building a new gas-fired power plant has more than tripled since 2021, from $800 per kilowatt to $2,600-$2,800. Over 100 GW of new gas projects have been announced in the US, but the turbines to power them don't exist yet.

This isn't a 2029 problem. It's a now problem that gets worse every year.

The spending gap is real, but it points the wrong direction #

US hyperscalers (Alphabet, Amazon, Meta, Microsoft) are on track to spend about $650 billion on AI infrastructure in 2026. That's up from $410 billion in 2025, a 60-70% jump in a single year.

China's total AI investment in 2025 was somewhere between $84 billion and $98 billion, depending on how you count. Even using the higher 2026 projection of $125 billion, it's a fraction of what four American companies are spending.

And here's the uncomfortable part: for that fraction, Chinese labs are producing models that cost a quarter (or less) of what US models cost to run. Peer-to-peer on capability, Chinese models run 3 to 5 times cheaper per token. At the extreme, DeepSeek V4 costs 29 times less than Claude Opus 4.8 while sitting within a few percentage points on capability benchmarks. GLM 5.2 landed within a point of Anthropic's top model on one agentic benchmark at roughly a fifth of the cost.

The US is paying an enormous premium to stay ahead at the model layer. Whether that premium is worth it depends on whether the models actually need to be that much better, or whether "good enough at a tenth of the price" wins the market.

Two different games #

The US approach: spend massively on frontier models, intervene at the company level (CHIPS Act subsidies for Intel, export controls on chips), and let the market sort out the infrastructure.

China's approach: treat electricity as public infrastructure, build generation capacity at a scale no other country can match, use cheap state credit to finance it, and let the model layer develop on top of that foundation. Open-weight models, subsidized inference pricing, and data centers in western provinces where solar and wind power cost as little as 2.7 cents per kilowatt-hour.

Neither approach is obviously wrong. The US still leads at the frontier. Its models are still the best in the world for the hardest tasks. But the Chinese strategy has a compounding advantage: cheap power lowers inference costs, cheap inference drives adoption, adoption pulls the domestic chip ecosystem forward, and the energy buildout keeps the whole loop supplied.

The US can outspend China on models. It can't outspend China on electricity, because China isn't spending. It's building.

The question nobody's asking #

The US is having a very public fight about who controls AI models. The Pentagon versus Anthropic. Open source versus closed. Governance frameworks and red lines.

Those fights matter. But they all assume the grid can handle what's coming. Data centers could consume 7-12% of US electricity by 2028, up from 1.8% in 2018. The first meaningful growth in US power consumption since the 1990s is happening right now, driven by AI.

The fight over who controls the model is important. The fight over whether there's enough power to run it is the one that actually decides the outcome. And that's the fight the US isn't really having.

Sources #

last updated: