Red STOP AI protest flyer with meeting details taped to a light pole on a sunny city street, San Francisco, California, May 20, 2025. (Photo by Smith Collection/Gado/Getty Images)
Gado via Getty Images
The most consequential debate in American utility regulation right now is not about renewables or transmission or nuclear restarts. It’s about cost allocation — specifically, who pays for the enormous grid investment required to serve data centers, and whether it’s the companies building them or the households who happen to live in the same service territory.
The scale of the question caught most people off guard. One analysis puts the amount already added to consumer electricity bills as a result of data center demand at roughly $23 billion. Wholesale power prices in capacity-constrained regions have risen sharply, transmission upgrade costs are being spread across ratepayer bases, and in several markets the retail price increases have been large enough to become a live political issue.
What makes this genuinely complicated is that the obvious story isn’t entirely right. Data centers are not straightforwardly parasitic on the grid. They are large, creditworthy, extremely predictable customers with flat load profiles, which is close to the ideal customer from a utility’s perspective.
In some regions the arrival of data centers spread fixed costs across more kilowatt-hours and actually pushed rates down. Recent academic research found that, nationally, data center growth between 2015 and 2024 modestly reduced average retail electricity prices by allowing utilities to spread fixed generation, transmission and distribution costs across a larger sales base. That dynamic reverses once demand growth outruns generation and transmission capacity, at which point utilities must build new infrastructure, wholesale prices rise, and everyone on the system pays the higher clearing price.
Congressional analysis of the sector notes that data center consumption has grown into a material share of national electricity use and is projected to keep climbing.
So the outcome depends almost entirely on regulatory design, which is why the rules determining who pays have become the central battleground. Several states have already begun moving toward special large-load tariff structures designed to ensure AI data centers bear more of the infrastructure costs they create. Virginia, where data centers account for the nation’s largest concentration of electricity demand, recently strengthened rules requiring many new facilities to pay for dedicated upstream transmission infrastructure. Georgia has adopted tariff structures requiring very large loads to make long-term financial commitments and cover certain grid upgrade costs, while Colorado is considering similar large-load tariffs through Xcel Energy. These approaches generally require hyperscale customers to cover interconnection costs, commit to minimum usage levels, or post collateral if projects are canceled. Other states have not adopted similar protections, leaving a greater share of the costs to flow into the general rate base.
For investors, this creates a set of exposures that are not being priced consistently.
Regulated utilities in high-growth data center corridors have been treated as a straightforward growth story, and the logic is sound as far as it goes. Utility earnings are a function of the rate base — the capital deployed and approved by regulators — and a decade of flat demand meant a decade of limited rate base growth. Suddenly there is enormous demand for new generation, transmission and distribution, and utilities get to earn a regulated return on all of it. That’s a real and durable earnings tailwind for the first time in a generation.
The risk sitting underneath it is political rather than operational. Regulated returns exist at the pleasure of state commissions, and commissions respond to voters. A utility that earns a healthy return on capital deployed to serve data centers while residential bills climb 20% is a utility with a rate case problem, and rate case problems become allowed-ROE problems, which is where the earnings actually live. The states with the fastest data center growth are, not coincidentally, the states where this fight is most advanced.
Independent power producers and merchant generators are the cleaner expression of the same theme. They sell into wholesale markets, capture the price increases directly, and have no regulated return cap and no residential customers to anger. They also have no rate base protection when prices fall, which they will if the buildout slows or if new generation arrives faster than demand.
The nuclear angle has been the loudest trade and deserves the most skepticism. Long-term power purchase agreements between hyperscalers and existing nuclear plants are real, economically significant and have transformed the outlook for operators who spent years facing closure. But restarts and new construction operate on timelines measured in the better part of a decade, with cost overruns as the historical norm rather than the exception. The contracts signed today are being valued as though the megawatts arrive next year.
Then there’s the demand-side question that most of these theses skip. Every projection of data center power consumption extrapolates current buildout plans. Those plans assume AI revenue that has not yet materialized at the necessary scale, and they are being funded increasingly with debt. If capital expenditure growth moderates — not stops, just moderates — the load forecasts that justify a decade of utility capital spending get revised, and utilities that built generation against contracts with counterparties who no longer need it will have stranded assets in a rate base that regulators are reluctant to let them recover.
Natural gas deserves a mention because it is the actual answer to most of the near-term demand, whatever the press releases say about clean firm power. Gas turbines can be built in two to three years rather than ten, the fuel is domestically abundant, and the manufacturers with turbine order books have visibility measured in years. That’s a less inspiring story than small modular reactors, and it’s considerably more likely to be what gets built. The order backlogs at the handful of companies that make heavy-duty turbines are a better real-time read on what utilities actually expect than any published load forecast.
The most defensible position here is probably the least exciting one. Own the parts of the energy infrastructure chain that benefit from grid investment regardless of which specific technology or customer drives it: transmission equipment, electrical components, grid-scale storage, engineering and construction. Those businesses get paid for building the system out and don’t depend on any single demand forecast being right.
That’s a duller trade than buying a nuclear operator on a hyperscaler headline. It also doesn’t require the AI revenue to show up on schedule, which at this point is the assumption doing the most work in the most portfolios.