The Invisible Trillion · The Energy Question, Honestly

Where the Real Problems Actually Are

So far this chapter might read as if I’m saying “the energy cost of AI is small, stop worrying.” That is not what I’m saying. There are real problems, and I want to name them clearly.

Local grid strain. Discussed at length in the previous section. The short version is that the global “1–1.5% of electricity” figure hides regional concentrations that strain specific grids, and the cost of expanding those grids falls disproportionately on residents and small businesses who do not use the data centers driving the expansion. This is not a future problem. It is a present problem, visible in PJM’s capacity auction results and in residential electricity bills across thirteen states.

Water. Chip fabrication uses enormous amounts of ultrapure water — a big fab can use 10 million gallons a day. Data center cooling also uses water, though less than fabs, and mostly in regions where water is actually scarce. Some new data centers being built in Arizona and Nevada are using techniques that use almost no water (air cooling, closed loops), but older ones in drought-affected areas are a genuine problem. You are not wrong to worry about this; you should just know that the industry is aware and the best new designs are much better than the old ones.

Power source matters more than power amount. A data center running on solar-plus-battery in Arizona has a very different environmental impact than a data center running on coal in West Virginia. Two data centers doing the exact same work, using the exact same amount of electricity, can have carbon footprints that differ by a factor of twenty depending on where their power comes from. When people argue about the environmental cost of AI, they are often unknowingly arguing about different data centers in different places, using different power, and talking past each other.

Embodied energy in the chips themselves. Every chip in a data center took energy to make — mining the materials, refining the silicon, running the fab, shipping the chip. For a server that will run for five years, the embodied energy is typically 10–20% of its lifetime total. This is not huge, but it is not nothing, and it is often left out of the accounting.

What we don’t know, which matters. The big AI companies do not publish detailed energy figures for their training runs or their inference infrastructure. Estimates come from leaks, indirect calculations, and informed guesswork. The numbers in this chapter are the best public estimates, but they could be off by a factor of two in either direction, and the lack of transparency is itself a reason to be skeptical of any confident claim — including, to be fair, the confident claims in this chapter.

All 19 chapters of The Invisible Trillion · Library · All courses