AI is an electricity business now

The arithmetic is not complicated, which is why it is so easy to check and so often skipped. Take a hundred thousand accelerators at a kilowatt each. Add servers, networking and storage, roughly another four hundred watts per chip. That is 140 megawatts of computing load. Multiply by a power usage effectiveness of 1.2 for cooling and conversion, and the site draws about 168 megawatts continuously, or around 1.5 terawatt hours a year. Our data center power calculator does this with your own numbers.

That is comparable to a city of several hundred thousand people, for one building.

Why the grid is the bottleneck

You can buy chips in months. You cannot buy a grid connection in months. A new high-voltage interconnection requires substation capacity, transmission headroom, a utility study and often new lines, and in mature markets the queue runs for years. Transformers themselves have multi-year lead times.

The result is that data center siting has become an exercise in finding existing electrical capacity: retired coal plants with intact switchyards, aluminium smelters, industrial parks, places with hydro. The sites chosen say more about where the power is than where the customers are, because a millisecond of extra latency matters far less for training than a year of waiting.

Cooling, and why it changed

Air cooling works up to roughly 30 to 50 kilowatts per rack. Modern accelerator racks pass 100 kilowatts, and the next generation goes further. Air physically cannot remove that much heat from that volume, so the industry moved to liquid: cold plates on the chips, and in some designs full immersion.

Liquid is more efficient and brings its own problems. It needs plumbing in the white space, a leak discipline the industry did not have, and a retrofit path for buildings designed for air. It also changes the water question: closed-loop liquid cooling uses very little water, while evaporative cooling uses a great deal. Reporting that quotes a data center's water consumption without saying which system is in use is not telling you anything.

What is actually being consumed

Training a frontier model is a large, visible, bounded expense. Inference is the larger one, because it runs forever and scales with users. A model queried a billion times a day consumes more energy in a month than its training run did, and that ratio keeps moving towards inference as adoption grows.

This matters for forecasting. Training demand is lumpy and can pause. Inference demand is a load that only goes away if people stop using the products.

The honest version of the emissions question

Large operators buy renewable energy and many match their annual consumption with purchases. Matching annually is not the same as running on clean power hourly, because a site draws at three in the morning when the solar farm produces nothing. The meaningful metric is hourly matched carbon-free energy, and very few sites report it.

The second-order effect is larger than the direct one anyway: a data center that connects to a constrained grid can keep a fossil plant running that would otherwise have closed, and that shows up in nobody's corporate report.

Nuclear, and the promises attached to it

Several operators have signed agreements for nuclear power, including restarting a shut reactor and contracting for small modular reactors that do not exist yet. The first category is real and delivers power on a known schedule. The second is a purchase agreement for a technology still in licensing, and should be read as a funding commitment rather than as capacity.

What to watch

Watch interconnection queue statistics in the large data center regions, which are public and boring and predict construction better than any announcement. Watch whether operators start reporting hourly carbon-free matching. And watch electricity prices for households near large new loads, because that is where this becomes a political story rather than a technical one.