One chip, counted honestly

An H100 chip is rated at 700 watts. That number leaves a lot out.

In a common setup, each chip shares a server with seven others. The whole server is rated at 10.2 kilowatts, which works out to about 1,275 watts for each chip once everything else in the box is counted. Cooling the room and moving power through the building add another 10 to 20 percent. So each chip really costs about 1.5 kilowatts, roughly twice its label.

Multiply up and the numbers get large fast. 100,000 of these chips make a hall that draws about 150 megawatts.

A gigawatt of datacenter

The biggest new sites are planned in gigawatts. At about 1,500 watts per chip slot, one gigawatt holds roughly 667,000 chips. Epoch AI estimates that such a site costs about $38 billion to build, chips and building together, and takes about 2.1 years. That works out to around $57,000 for every chip slot.

Every price on a price sheet has to pay back some share of a building like that, plus its electricity bill.

The whole industry

On September 21, 2026, a study in the journal Communications Sustainability estimated the electricity used by AI datacenters at six large firms: Amazon, Microsoft, Google, Meta, Oracle and Apple. It put their use at about 118 terawatt-hours in 2024. A terawatt-hour is a billion kilowatt-hours.

By 2030 the authors project 239 terawatt-hours in a conservative scenario and 295 in an optimistic one, about 1% of the world’s electricity. More than 90% of the projected capacity sits in North America, Western Europe and Asia-Pacific. Oregon, Ireland and Iowa face more pressure on their power systems than larger ones such as Texas.

One question, counted small

Per question, the energy is small. Google reported in 2025 that the median text prompt to its Gemini app used 0.24 watt-hours, about nine seconds of television. Energy per median prompt fell 33 times in the year to May 2025. The figure is Google’s own and has not been independently verified, and Hello, AI suggests asking of any energy number: what did they count?

Tiny per question and large in total can both be true. Billions of daily questions and the training of new models add up to a building boom.

In one sentence: each AI chip really costs about twice its rated power, and a new study expects AI datacenters at six big firms to use about 1% of the world’s electricity by 2030.