The Invisible Trillion · The Energy Question, Honestly

How Much Of The Growth Is Real?

There is one more thing about the energy projections that the rest of the book will spend more time on, and that I want to flag here because it changes how you should read the “rising fast” language.

The energy growth projections for data centers — the ones that have utility commissions and grid operators worried, the ones that justify the capacity buildouts described above, the ones that show data center electricity going from 1.5% to 4% or more of total U.S. electricity by 2030 — all rest on a specific assumption. They assume that the current pattern of AI usage continues and grows. Specifically, they assume that the substantial majority of AI inference (the moment-by-moment use of trained AI models to answer queries, generate text, produce images, and so on) continues to flow through hyperscale data centers, billed through frontier-model APIs, at something like the current pricing.

This assumption is not guaranteed to hold, and a later chapter takes apart the reasons it may not. The short version: AI capability is migrating to edge devices faster than the projections assume. Phones and laptops in 2026 run models that three years ago would have needed hyperscale clusters, and a 4-billion-parameter model on a phone handles the substantial majority of what ordinary people actually use AI for.

The cost gap is about three orders of magnitude — local inference runs at roughly one-thousandth the cost of a frontier API, for tasks the local model handles adequately. As users and enterprises wake up to that gap, much of the inference now running in hyperscale data centers migrates to the edge, where the energy is amortized over devices people already own and drawn from a wall outlet at retail rates.

This matters because much of the buildout is justified by demand projections that the people funding it have a financial interest in keeping high. The hyperscalers want high projections to justify their capacity and their valuations; the utilities want them to justify rate-case requests and recoverable infrastructure spending; the financial system around the buildout wants them to support its equity valuations and bond issuances. None of these parties has any incentive to forecast that demand will soften, and the projections that reach news coverage almost all come from them or their consultants. The result is a systematic upward bias in the numbers that drive energy planning.

This does not make the projections worthless. Some growth is real, especially for training and specialized applications. But the honest version of “AI energy use is rising fast” is narrower: it is rising fast at the current architecture, and a substantial share of the rise comes from inference patterns that may not persist as small models improve and enterprises rationalize their cloud spending. That is the sentence the projections should be using and mostly are not — and the residential ratepayers funding the resulting grid expansions are being asked to bet that the softer reading is wrong.

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