Who Pays for the AI Grid? Data Centers Meet Decarbonization

This blog has run two long series that never met each other. One examined the economics of artificial intelligence — investment, productivity, labor displacement, policy. The other has spent the past year on the economics of climate and energy. The place where they collide is the electricity system, and the collision is happening faster than either literature anticipated. After nearly two decades of essentially flat electricity demand in the United States and much of Europe, load growth has returned, and a substantial share of it comes from data centers.

The first thing to say about the forecasts is that they disagree wildly. Projections of data center electricity consumption by 2030 span a range wide enough to imply completely different infrastructure programs, because they rest on assumptions about model efficiency, chip improvement, utilization, and demand elasticity that no one can currently pin down. Utilities, however, cannot wait for the uncertainty to resolve. They plan on interconnection requests, and those requests are inflated by speculative queue-stuffing — developers filing in multiple territories for projects they will build in one. The result is a planning process biased toward the high end of the range, which is exactly the condition under which a system overbuilds. Readers of the earlier post on retiring fossil infrastructure early will recognize the shape of the risk: assets built for demand that does not arrive become stranded, and someone absorbs the loss.

That someone is usually the ratepayer, and this is where the interesting economics lie. Grid costs are largely recovered through volumetric and demand charges spread across a utility's customer base, an arrangement that made sense when load growth was slow and broadly distributed. It makes considerably less sense when a single facility requests several hundred megawatts and triggers hundreds of millions of dollars in transmission and generation investment. Absent carefully designed large-load tariffs, minimum take obligations, and exit fees, the cost of building capacity for hyperscalers is socialized across households and small businesses who receive no corresponding benefit. Several jurisdictions have begun writing special contract classes to address this; many have not, and the regulatory proceedings determining the answer are technical enough to attract almost no public attention despite being straightforwardly distributional.

The corporate clean energy claims layered on top deserve more scrutiny than they typically receive. The dominant convention has been annual matching: a company purchases enough renewable energy certificates over a year to equal its consumption, and reports itself as running on clean power. But electricity is consumed hour by hour, and a data center drawing steady load overnight in a region where the marginal generator is gas is running on gas, whatever the annual ledger says. Hourly, or 24/7, matching is a genuinely more demanding standard, and the shift toward it — along with the recent wave of nuclear power purchase agreements and restart deals — reflects real pressure on the older accounting. The deeper question is additionality: buying output from clean generation that already exists reallocates it rather than adding any, and if that generation would otherwise have served other customers, the net emissions effect approaches zero.

The framing that dominates public argument — whether AI is "worth" its energy consumption — is not one economics can settle, and it obscures the question that is actually being decided. Load growth is not intrinsically a problem; electrification of heating and transport will drive far more of it over the coming decades, and a growing grid is a decarbonizing grid's normal condition. What matters is who bears the cost of the buildout, who bears the risk if projected demand fails to materialize, and whether the new load is met with additional clean generation or by keeping fossil capacity online past its retirement date. Those are allocation questions, and this blog's earlier examination of carbon pricing's distributional politics applies directly: an energy transition whose costs fall on people who did not choose them and cannot afford them tends not to survive contact with the electorate.