The Invisible Trillion · The Energy Question, Honestly

What This Actually Means

The honest version of the energy story, after everything in this chapter, is more complicated than the version most people will tell you.

The global aggregate is small. Data centers worldwide use 1 to 1.5% of global electricity. AI is a fraction of that. Your individual AI queries cost a few watt-hours apiece, less than almost anything else you do. The “AI will boil the oceans” framing is not supported by any number anyone can produce. The people telling you that AI energy use is catastrophic at a global level are mostly wrong, in the literal sense that their numbers do not add up.

The regional reality is large. In the specific regions hosting the buildout — Northern Virginia, parts of Ohio, the Memphis metropolitan area, parts of Arizona, Texas, Oregon, Iowa, and several other concentrated regions — data centers are a substantial share of the local electricity load, and the cost of expanding the grid to serve them is falling, through capacity-market mechanisms and rate-case proceedings, on residential and small-business customers across thirteen states in PJM alone. The 94% residential electricity price increase in Washington, D.C. is not abstract. It is paid, every month, by ordinary households who do not use the data centers driving the cost. The cost-shifting is real, large, and unfair.

The architecture is wrong. Much of the inference currently flowing through hyperscale data centers does not need to be there. The same tasks could be done on edge devices, at one-hundredth the energy cost, with no marginal grid expansion required. The reason most inference runs in the cloud is not technical necessity; it is business-model preference on the part of the providers. The energy implications of those preferences are being paid by people who do not benefit from them.

The growth projections are biased upward. The “AI energy demand will quadruple by 2030” projections all assume that the current routing of inference through hyperscale data centers continues. The chapters that follow this one will explain in detail why that assumption is likely to be partly wrong. To the extent it is wrong, the grid expansions being built right now are being sized for a demand pattern that may not hold, and the residential ratepayers being asked to fund those expansions are being asked to fund infrastructure that may end up underutilized.

The grid decarbonization argument is still real. Whatever the AI energy load turns out to be, its carbon footprint depends on what powers it. A buildout that runs on solar, wind, and nuclear has a small carbon footprint. A buildout that runs on natural gas — including the on-site gas turbines now being installed by AI infrastructure operators in residential neighborhoods, often without permits — has a large one. The grid leverage point remains the most important environmental leverage point, and people working on grid decarbonization are working on the right thing.

None of this is a reason to abandon AI as a category. The technology is real, the productivity gains are real, and some of the applications (self-driving, scientific computing, drug discovery, genuine education tools) are worth the energy they consume. The argument is about how to build the infrastructure, where, and who pays for it, not about whether AI should exist.

What follows is that “the energy cost of AI” is not a single number to be for or against. It is a set of specific costs, falling on specific people, produced by specific architectural and financial choices that could mostly be made differently.

Which is what separates the vague version from the actionable one. “AI uses too much electricity” is not a policy. “Data centers should pay for the grid capacity they require through their own rate class, rather than spreading the cost across residential ratepayers” is a policy. So is “AI inference should run on edge devices by default, with cloud routing reserved for tasks that genuinely require frontier-scale capability.” So is “on-site combustion power in residential areas should require full permitting and full emissions accounting.” Each of those is specific enough to argue about, fund, or fight. The vague feeling is not.

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