Data centers could lower power bills. Energy utilities won't.
The regulatory contract pays them to grow rate base, not lower bills.
Twenty-seven gas turbines run day and night without an air permit. 2 miles away are about six public schools, including an elementary school. They are trucked in to power xAI’s Colossus 2 data centre across the state line in Mississippi. The nearby Boxtown neighbourhood already breathes a cancer risk roughly four times the national average, and the turbines could emit more than 1,700 tons of nitrogen oxides a year.
Likely the largest industrial NOx source in a metro area already failing national smog standards. Mississippi regulators have nonetheless permitted an even larger facility at the same site. The turbines run because the regional grid takes five to seven years to deliver new interconnections, and AI capital-expenditure schedules do not.
A steady, high-load-factor customer spreads the grid’s fixed costs across a broader base. In theory, it should lower rates for the households around it.
It is also, in the United States in 2026, mostly just theory. Wholesale prices near significant data-centre activity have risen as much as 267% over five years, and household prices rose 7.4% in the year ending February 2026. PJM’s capacity-market clearing prices have hit their FERC-approved cap for three auctions running. Pew estimates that data-centre capacity costs alone are adding $18/month to residential bills in western Maryland and $16/month in Ohio, before transmission and distribution buildout. The industry isn’t doing what it says it could.
Ninety years earlier, in the same country, a federal agency that had not existed six years before was stringing electric lines to Tennessee farms whose owners had never owned a refrigerator. The Tennessee Valley Authority, chartered in 1933, employed more than nine thousand workers within a year, building dams and electrifying a region whose per-capita income was half the national average. The Rural Electrification Administration, founded in 1935, brought power to most American farms within fifteen years.
What could data centres do?
Data centre actors accounted for roughly half of U.S. corporate clean-energy procurement in 2024, and the IEA’s Energy and AI report documents them as the largest single class of corporate clean-power buyers in the world. You need demand for energy if you want to build out more energy infrastructure, especially clean energy infrastructure.
Electricity systems carry large fixed costs and recovered across whatever load shows up to pay them. A new subdivision adds load that peaks at dinner and disappears overnight. A data centre runs 24/7 at 90%-plus capacity factor. The data centre is the better customer on cost recovery alone.
What hyperscalers want is a lot of cheap power on timelines that match their capex schedules. They do not care if it’s clean energy or fossil fuels. They want compute, and they want it cheap. The cheapest compute in 2026 is powered by solar plus storage plus, where available, nuclear. You think that a rational hyperscaler in a working electricity system would invest in clean generation and sign contracts that anchor utility investment. It would absorb cost overruns so its neighbours’ bills did not move. Some of them do.
You also think that ones who have a bit of political influence would encourage the government to make the investments for them, in a way that doesn’t draw ire from the public.
So why don’t they all?
What do hyperscalers do instead?
The hyperscaler arriving in a new region can spend years on interconnection negotiations and absorb cost overruns so its neighbours’ bills do not move. Or the hyperscaler can route around the institution entirely. It can build behind-the-meter gas turbines, sited where permitting frictions are lowest, and run them while the permits are pending or without seeking them at all. xAI in Boxtown is the irresponsible example of this. It sits there not because the architecture forced it to, but because its principals do not believe the surrounding energy environment is their problem. The political economy so far signalled that they are right.
This isn’t just the xAI turbines in Boxtown. We have the 4.5 GW Crusoe gas expansion in Abilene, the 400 MW Williams “Socrates” facility serving Meta in Ohio, and Microsoft’s $7 billion Chevron-built gas plant in the Permian Basin.
Cleanview tracks 56 GW of behind-the-meter capacity across 46 sites and now projects 30% of new data-centre capacity will be served on-site. Almost all are up from effectively zero a year ago. It’s become enough of a problem that a bloc of Senate Democrats opened a formal investigation in March 2026 into twelve such projects.
It isn’t just all fossil fuel expansions. Microsoft signed a $10.5 GW renewables procurement with Brookfield in 2024, the largest corporate clean-power deal on record. Google has signed PPAs anchoring new utility-scale solar buildouts and a Kairos Power agreement for advanced reactors structured as a long-horizon clean-anchor contract rather than a turbine workaround. Emissions aren’t going to be much of a problem with solar panels and wind turbines.
It is a small set of operators who have decided that absorbing the procurement risk is cheaper, than the regulatory blowback the shortcut will eventually produce.
The queue is what makes the shortcut attractive. By the end of 2025, over 2,060 GW of generation and storage were seeking grid connection across the U.S., according to Lawrence Berkeley National Laboratory, more than twice the country’s existing fleet. The median wait from request to commercial operation has doubled, from under two years for 2000–2007 projects to over four years for 2018–2024 projects.
CenterPoint Energy in Texas reported a 700% jump in large-load requests in a single year. Other Texan utilities like ComEd and PPL (along with Oncor) now report more gigawatts of data-centre applications than their entire historical peak demand.
Lack of hardware makes it really hard to clear out the queue. By Wood Mackenzie’s Q2 2025 survey, lead times for standard power transformers have reached 128 weeks and 144 for generator step-up transformers. Roughly 80% of large U.S. power transformers are imported. Cleveland-Cliffs is the only domestic American producer of the grain-oriented electrical steel transformers require.
There are nearly $2 billion in announced manufacturing expansions. Which will not come online until 2028 or later. You need the grain-oriented-steel workforce and the high-voltage line crews. Along with substation engineers who connect new generation. We can’t scale up because the apprenticeship pipelines for these trades have been thinning for two decades while nobody was paying attention.
What do utilities want?
The other part of the problem is who the hyperscalers need to work with. American investor-owned utilities operate under a regulatory contract essentially unchanged since the early twentieth century. The state grants the utility a regional monopoly and an obligation to serve every customer in it.
In exchange, a commission sets rates that recover the utility’s operating costs plus a regulated return on its rate base (the depreciated capital sunk into plant and lines). This was a sensible design for extending wires to new neighbourhoods at predictable cost. It is also a disastrous design for coordinating a multi-trillion-dollar buildout against a state competitor that does have the capacity to do so.
This rewards rate-base growth above almost everything else. The Averch-Johnson effect (named after the 1962 paper that formalised it) is the standard result on this kind of regulation. A utility whose authorised return exceeds its cost of capital has a structural incentive to “over-invest”.
The American Economic Liberties Project’s January 2025 paper by former Sempra chief economist Mark Ellis estimated that authorised returns on equity exceed the actual cost of capital by enough to cost American utility customers roughly $50 billion a year, about $300 per household. The stock market capitalises that stream of excess returns at about $1 trillion. For 2024, the average authorised return on equity for U.S. regulated utilities was 9.7%, against a broad-market forecast of 6.7%. Over three decades, Ellis finds, IOU capital investment per kilowatt-hour delivered grew about 3.8% per year faster than demand. The legal threshold articulated by Justice Brandeis in 1923 and adopted by the Supreme Court in 1944 holds that returns above the cost of capital are “unjust and unreasonable“. Nobody enforces it, because nobody’s job description is to.
You can do this experiment yourself (well, almost). Two California utilities (San Diego Gas & Electric and the Sacramento Municipal Utility District) serving major metro areas under the same state energy policy. Same wildfire risk and demand growth. One answers to Sempra’s shareholders, the other to a board elected by the City of Sacramento. Between 2020 and 2023, SDG&E’s residential rate rose 65% and now exceeds 45¢/kWh. SMUD’s rose less than inflation and stayed near 17¢/kWh. The gap is more than two and a half times.
Financing explains some of it. Public utilities raise capital through municipal bonds at 4 to 5% interest and are exempt from corporate income tax. IOUs raise capital at nearly double the cost. But these account for perhaps a third of the gap. The rest is the management philosophy that follows from ownership. SMUD’s managers are not maximising earnings per share and its investors are not expecting double-digit returns. SDG&E’s are. Sempra’s CEO took home roughly $14 million in total compensation in 2023, while SMUD’s general manager made under $700,000. Sempra’s stock price tracks SDG&E’s rate base because that, by regulatory design, is the source of its profit. A larger rate base produces higher earnings, and the resulting share price feeds executive pay. Everyone in the chain responds rationally to the incentive structure. Nobody is behaving badly. The chain is the problem.
The pattern is industry-wide. The American Economic Liberties Project found that between 2020 and 2023, IOU residential rates rose 49% more than inflation while publicly-owned utilities’ rates rose 44% less. Investor-owned utilities used 2025 earnings calls to disclose $1.4 trillion in planned capital investment through 2030, a 21% upward revision in a single year, the rate base growing in real time with the profit stream attached.
The mechanisms range from outright bribery to rate-case manoeuvring. The defining recent case is FirstEnergy and Ohio’s House Bill 6, in which from 2017 to 2020 the company paid roughly $60 million through a dark-money nonprofit controlled by Larry Householder, then Ohio’s House Speaker, now a federal inmate. The payments bought $1 billion in subsidies for FirstEnergy’s struggling nuclear and coal plants. The company also paid $4.3 million to the man who would go on to chair the state utility commission.
The newest mechanism is secret data-centre contracting. You might think utilities at least file these things publicly. You can guess that they don’t.
A March 2025 Harvard Electricity Law Initiative paper, Extracting Profits from the Public by Eliza Martin and Ari Peskoe, reviewed nearly fifty regulatory proceedings and found a pattern. Utilities negotiate special contracts with hyperscale customers and file them with state PUCs under confidential treatment.
The infrastructure costs are then spread to all ratepayers through ordinary rate cases. Federal antitrust litigation against Duke Energy established that the company sought to shift the costs of a $325 million data-centre discount onto its other ratepayers. Similar patterns appear at AEP and Dominion, and at Entergy in the South. The rate structures that once spread the cost of reliable power across everyone are now forcing the public to pay for infrastructure built to supply a handful of wealthy corporations.
Utilities looking at a hyperscale interconnection is not looking at a problem. It is looking at the largest rate-base-growth opportunity in half a century. So, they will offer favourable terms and spread the infrastructure costs across its broader ratepayer base, filing the whole thing under confidential treatment. The hyperscaler, faced with grid lead times that cannot wait and just don’t care. Of course they take the deal.
Why doesn’t either side fix itself?
The deregulatory account is that American electricity got expensive because regulators got in the way. That Texas is the proof case (the most market-based major grid in the country) with the lowest industrial rates in the lower 48. Get government and the legacy utilities out of the way, the argument goes, and supply responds. The second Trump administration’s Executive Order 14154 and the “Speed to Power” agenda are its policy expression.
The progressive account runs the other way. American electricity got expensive because corporate concentration captured the public interest. IOUs extract excessive returns through regulatory capture. In this case, it is the hyperscalers externalise their power costs onto residential ratepayers. The alignment between Big Tech and the Trump administration is the visible quid pro quo. The ideal remedy is some sort of enforcement. It’s also a popular idea as Abigail Spanberger won Virginia by 14 points in November 2025 campaigning on this account. No to mention, 78% of Americans told Consumer Reports they are worried data centres will raise their bills.
Each account have a degree of truth. Permitting and interconnection lead times are just absurd. IOUs do extract roughly $50 billion a year through a comprehensively gamed system. But both share a deeper failure as they assume the architecture of American electricity governance is adequate. The operational (not political) problem is which actors operate within it.
ERCOT runs the most market-based major grid in the country. In twelve months it requested a federal emergency order during Winter Storm Fern. The legislature afterwards passed Senate Bill 6 to mandate large-load interconnection rules. The state also leaned on the Texas Energy Fund (a $9 billion voter-approved programme) to fund dispatchable generation. That is an energy and industrial policy in everything but name. Texas (reluctantly and in pieces) is reverting to state coordination because the market alone cannot do the planning.
Even the Texas Public Policy Foundation (ever hostile to coordinated state action) documents that the state’s winter capacity gap is now larger than during Winter Storm Uri in 2021. ERCOT added 31 GW of solar and 15 GW of batteries but only 3.1 GW of new natural gas between 2021 and 2026. While peak winter demand grew roughly 20%.
PJM (the largest grid operator in the country and serving 67 million people) hit its FERC-approved capacity price cap for the third auction in a row in December 2025. Still 6,625 MW short of its reserve-margin target. Absent the temporary cap negotiated with Pennsylvania Governor Josh Shapiro, PJM estimates the price would have cleared roughly 60% higher. It’s own market monitor attributed about $6.5 billion of the $16.4 billion auction cost to data-centre load.
The progressive account isn’t any better. If it was simply that Big Tech and IOUs extract from ratepayers. The cure would be to discipline them and nothing else. But hyperscalers choose gas turbines because no institution exists to deliver the alternative. State regulators are patching, late and unevenly. The Virginia State Corporation Commission approved a new GS-5 rate class in November 2025 shifting more buildout cost onto large data-centre customers. A the same time approving a $16/month residential rate increase.
Who got what they paid for?
So why do utilities and hyperscalers, whose deeper interests actually diverge, end up coordinating against the public anyway? American political economy is fragmented across layers of government and empowers local self-rule. Which privileges actors with long time horizons and the ability to operate across many venues at once. As JS Tan has argued, hyperscalers have both. It turns out, so do investor-owned utility holding companies. Almost nobody else does.
In practice this means venue arbitrage. A hyperscaler shopping a $5 billion project can play states and counties against one another, and play their regulators. A utility shopping a rate case can find a friendly PUC or legislature and play it against federal pressure. Neither needs the other’s enthusiastic cooperation. The short term incentives just happen to align.
Hyperscalers’ deeper interest is in cheap power, which is the same for you and. Given a working developmental institution, that interest would dominate. In the system we have now, it loses to something more proximate. Utility wants rate-base growth while the hyperscaler wants compute on capex timelines grid interconnection cannot meet.
You can see where this is going as the deeper interest is slow, but there is a shortcut. Most of the industry are going to take the shortcut.
They are coordinate with the utilities, sometimes through secret contracts. Other times through siting decisions that route load toward whatever utility offers the least scrutiny, not because their interests align but because nothing better is on offer.
The political surface is loud enough to mistake for the cause. Five major hyperscalers each gave $1 million to the Trump inauguration fund. Marc Andreessen and Ben Horowitz pledged $50 million to an AI super PAC.
Trump’s Executive Order 14148 fast-tracks any data centre over 100 MW, covering gas turbines and coal equipment while pointedly excluding wind and solar. I am not the “Big Tech bought the election” type. I am also not going to pretend the contributions and the EO are unrelated.
The temptation is to read this as straightforward corruption (I am not ruling it out). Hyperscalers bought the deregulation they wanted. I think that reading is too simple. Hyperscalers really did want was to scale. What they bought (and what the administration delivered) was permission to bypass the regulatory institutions that were not delivering it. The utilities were happy to receive their share.
The permission is not durable, and the political surface is already shifting. The backlash is not confined to the affluent-liberal-NIMBY caricature big tech finds convenient. As Taylor Lorenz has put it, “Data centers are one of the only physical manifestations of ‘big tech’ encroachment on ppl’s lives and broad swaths of the public are very angry and resentful of these companies and the power they have,” adding that “AI cos have done a terrible job explaining why they need these data centers.” Some industry observers see the reckoning coming: Thomas Hochman predicts that “The next Dem admin will look to ban data center scale deployment of diesel gen sets.” mattparlmer puts the strategic failure more sharply: “If you want to maximize the rate of progress over the long run you need to take probable political opposition into account and take steps to mitigate it by making sure it won’t derail you.” The bypass strategy works only until it doesn’t.
There are good proposals out there to build as much energy (clean btw) to bring down prices to penny a kilowatt/hour. It’s just as JS Tan said, the country has not lost the capacity to build. It has lost the capacity to build for the public.