The Infrastructure Subsidy: Your Power Bill Is an AI Investor Now

The cloud has a street address. It is a warehouse in a rural county with a substation built to serve it, transmission lines reinforced to reach it, and generation capacity added to feed it. Someone pays for all of that. The question this desk keeps asking about the AI buildout — who carries the cost, and who books the benefit — has an unusually clean answer in the power sector, because electricity is one of the few places where the cost-shift is written down in regulatory filings.

The documented mechanics first. In March 2025, Eliza Martin and Ari Peskoe of Harvard Law School's Electricity Law Initiative published a paper whose title does the work: "Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech's Power." Their finding, drawn from rate cases and the special contracts utilities sign with data-center customers, is that the traditional utility model — spread infrastructure costs across all customers — is being applied to infrastructure that exists only because a single industrial customer demanded it. The upgrades are real: new substations, reinforced lines, new generation. The allocation is the problem. Costs incurred for one tenant are recovered from everyone.

The scale is now visible in the aggregate numbers. Utilities requested more than $29 billion in rate increases in the first half of 2025 — roughly double the same period a year earlier — with data-center-driven grid investment a repeatedly cited factor. In the PJM region, the grid operator's own capacity auction produced billions in additional cost that flows to retail bills, a story we cover line-by-line in our PJM ledger. A widely shared claim that data centers drove electricity prices up 267 percent has been challenged by fact-checkers, so we do not use it; the desk's rule is that the defensible numbers are damning enough.

And it is worth being precise about which energy actually strains the system. The industry talks about training efficiency — the one-time cost of building a model. The recurring load is inference: every prompt, every generated meme and synthetic report, drawn from the grid billions of times a day, forever. Training is the mortgage; inference is the utility bill that never ends. Our Power Draw signal tracks the trajectory — data centers a few percent of US electricity now, with credible projections toward eight to twelve percent within the decade.

Did AI do this, or did we?

The machines did not write the rate cases. The cost-shift runs on two very human decisions. The first belongs to the utilities, for whom a data center is the best customer in a generation — guaranteed industrial load and a regulated return on every dollar of infrastructure built to serve it. The incentive is to say yes to the buildout and socialize the bill, because that is what the existing rules allow. The second belongs to the hyperscalers, who negotiate their own rates down in confidential special contracts while the induced grid costs land in the general rate base. Neither party is breaking rules. The rules were written for a world where load growth was slow and shared — and nobody rewrote them before the largest private infrastructure buildout in history arrived at the interconnection queue.

What we are not claiming: that data centers pay nothing (they are large taxpayers and ratepayers), or that grid investment is waste. The claim is narrower and documented: the current allocation rules let the costs of AI-specific infrastructure be spread to households that will never use it, while the compute, the margins, and the valuations remain private. That is a subsidy in everything but name — an involuntary infrastructure investment by every ratepayer in the service territory, with no equity attached.

The fix is not mysterious; regulators can assign cost causation honestly, and some are starting to. Until they do, read your power bill the way this desk reads a filing. Somewhere in the line items, you are funding the buildout — and the return on your investment is a warehouse's glow on the horizon.