The Bill That Arrives Before the Datacentre Does

Wisconsin ratepayers owe more than $1 billion on power plants that are already shut. We Energies estimated remaining value of over $700 million across three recently retired units as of December 2024, and WEC Energy Group will likely have over $1 billion in recently retired assets by the end of 2026. The meter runs whether the turbines turn or not. The debt is real; the electrons are not.

That is the inheritance. Now add the forecast: utilities across the state are scheduled to request permission in early 2026 to build new power plants or expand existing ones to serve datacentres. Some states, such as Minnesota, have adopted laws prohibiting the costs of stranded assets from data centers being passed onto ratepayers; Wisconsin has no such laws. A finance owner opening an electricity bill in suburban Milwaukee this autumn faces two liabilities stacked in the same line item. She is still paying off the last generation cycle that ended early, and she is about to underwrite the next one on a demand forecast nobody will defend in writing.

The question is not whether datacentres need power. The question is what happens when the baseload they promised in the interconnection application does not show up on the schedule the utility used to size the turbine order, and who receives the invoice when the asset goes cold.

The arithmetic already in the rate base

US utilities plan $1.4 trillion in spending through 2030, a PowerLines analysis released on 14 April 2026 reveals, jumping 27% from last year's $1.1 trillion projection and effectively doubling the $700 billion invested over the previous decade. Duke Energy plans to spend an industry record of $103 billion over five years, and CEO Harry Sideris told Fortune he expects that number to grow because the AI surge is only beginning. PowerLines estimates that residential customers could bear approximately $700 billion of the $1.4 trillion total through rate hikes.

That $700 billion is not a conditional forecast. It is the fraction of utility capex that will land in residential tariffs because the regulatory formula says so. The structure works like this: a utility builds generation or transmission capacity, capitalises the cost, earns a regulated return on the asset for its expected life, and recovers both through the rate base. If the asset stays in service for thirty years, the residential share of the recovery gets spread across three decades of billing cycles. If the asset is retired early or if the load it was built to serve does not materialise, the unrecovered balance does not disappear. It moves to the next rate case as a stranded cost, and unless a state law or a commission order says otherwise, it gets allocated to whoever is still connected to the system.

Since 2020, residential electricity prices in the US have risen by more than 36%, from 12.76 cents per kilowatt-hour to 17.44 cents per kilowatt-hour in February 2026. Average residential rates rose 7.3% between April 2025 and April 2026, and utilities requested a combined $18.6 billion in rate increases in the first half of 2026 alone. Areas with high concentrations of datacentres saw electricity prices jump 267% over the past five years, and nearly three-quarters of Virginia voters blame the facilities for rising costs, according to a January 2026 survey conducted by Global Strategy Group and the Chesapeake Climate Action Network Action Fund.

The Wisconsin finance owner is not paying for capacity that will serve her house. She is paying down debt on generation that was forecast for industrial load that either moved or changed its mind, plus a forward commitment to generation the utility has already ordered on a datacentre demand curve it admits it cannot model with confidence.

The demand nobody modelled

The IEA reports that global electricity demand from datacentres grew by 17% in 2025, and electricity consumption from AI-focused datacentres surged 50%. The Big Five hyperscalers will collectively spend approximately $725 billion on AI infrastructure in 2026, a 77% increase over the $410 billion deployed in 2025. US datacentres consumed more than 4% of total electricity in 2023, and that figure is projected to reach 9% by 2030.

Those percentages describe an aggregate. They do not describe a contract. A hyperscaler signs a power purchase agreement or a tariff with a fixed minimum take, the utility models load growth over twenty or thirty years, the commission approves the rate structure, and the utility orders the equipment. The binding number in that sequence is the one the utility gives to the turbine manufacturer, because that purchase cannot be unwound if the customer renegotiates two years later.

Datacentres currently account for less than 1% of Duke Energy's peak demand in the Carolinas, but the company expects them to represent about 10% of its total electricity sales by 2030. Duke believes datacentre load could account for 25% or more of total system demand in its Carolinas footprint by 2030. The span between 10% and 25% is not a rounding error; it is several gigawatts of generating capacity that either get built or do not, and either run at the load factor the model assumed or do not.

The risk is asymmetric. If Duke under-builds and the hyperscalers show up on schedule, the queues lengthen and the projects stall, but Duke does not carry stranded assets; it just loses the revenue it forecast. If Duke over-builds and the hyperscalers delay, down-size, or cancel, the generation still gets paid for, and the Carolinas rate base absorbs it. Utilities face stranded-asset risks with regards to generation and transmission buildout; if infrastructure is built to serve projected datacentre demand and said demand does not materialise, these assets could be underutilised.

The state that wrote the out-clause

Minnesota's new law prohibits the costs of such stranded assets from being passed to other ratepayers. In 2025, the Legislature passed one of the strongest datacentre laws in the country, and Governor Tim Walz signed it; the law puts the largest electricity users in a class of their own and makes them pay the full cost of the power and infrastructure built to serve them.

The mechanics are straightforward. The Minnesota Public Utilities Commission must review tariffs or Electric Service Agreements to ensure all costs attributable to the very large customer are assigned to that class, and that other customers are not at risk of paying for stranded assets caused by the utility serving the very large customer. If a datacentre signs a fifteen-year agreement, leaves after twelve, or reduces its draw by half because the next-generation chips run cooler than the spec the utility planned for, the unrecovered infrastructure cost stays with the datacentre customer class. It does not migrate to residential.

In Oregon, an early datacentre hub due to its abundant and relatively cheap hydropower, rates paid by Portland General Electric customers have risen 50% over the past five years; Democratic Governor Tina Kotek signed a similar ratepayer protection law there last year. Virginia, the country's largest datacentre market, is among the states considering laws this year that would shift more costs onto tech companies after an independent study showed datacentres pushed up residential utility rates there.

Wisconsin has not moved. Ratepayers are on the hook for paying off the full debt of stranded assets unless a financial tool called securitization reduces the burden, which refinances the obligation at a lower rate but does not remove it. The finance owner in Milwaukee is not waiting for the datacentre to open. She is already paying the forward premium on the risk that it will not.

Three ways the forecast breaks

The stranded-asset scenario does not require a hyperscaler to go bankrupt. It requires any one of three ordinary shifts in the build cycle.

First, the efficiency gain. Power consumption per AI task is declining rapidly, with efficiency improving at a rate unprecedented in energy history. A campus spec'd for 500 megawatts in 2024 might deliver the same compute with 300 megawatts by 2028 if the next chip generation halves the energy per FLOP. The lease still runs; the minimum take clause still applies; but the actual draw drops, and the excess generation capacity the utility built to the original forecast sits idle. That idle capacity is a stranded asset the moment it stops earning the return the rate base assumed.

Second, the competitive site selection. From 2025 onward, the bottleneck has migrated from the server rack to the substation, with US interconnection queues delaying projects for years and utility providers warning of regional capacity shortages as early as 2026. A hyperscaler submits interconnection applications in three states simultaneously, signs conditional agreements in two, and exercises the option in the one where the grid study clears first. The two that do not clear still show up in the utility's capital plan until the contract gets formally terminated, and the termination usually happens after the long-lead equipment order has already been placed.

Third, the technology substitution. Constrained by slow grid connections, datacentre developers are advancing projects with onsite natural gas-based power generation in the US; IEA satellite-based tracking shows many of these projects remain in early stages, and one of the key challenges is that AI datacentres have rapid and large swings in demand. A developer signs a grid interconnection agreement, the utility starts the substation build, and eighteen months later the developer announces it will self-supply with on-site gas turbines plus battery storage instead. The utility is left with transmission infrastructure sized for a load that will never connect. That is a stranded asset by definition, and unless the state wrote a carve-out, it goes into the residential rate base.

None of these scenarios is speculative. All three have already occurred in European markets. Ireland has €5.8 billion in stranded datacentre investment: fully permitted projects on purchased land that cannot reach commercial operation because the grid cannot connect them.

The contract the hyperscaler will not sign

The structural solution is a power purchase agreement with a take-or-pay clause that runs for the full depreciation life of the generation asset, ideally with a termination fee that covers the unrecovered capex. Duke Energy introduced minimum take clauses requiring datacentres to pay for a certain amount of power regardless of consumption, and could also introduce contracts requiring datacentre operators to make up-front contributions to pay for new power infrastructure. Duke Energy is requiring upfront financial commitments from datacentre developers before approving grid connections, attempting to shift some infrastructure costs away from residential ratepayers and onto the companies creating the demand.

The hyperscalers are not signing those. They will commit to a seven-year minimum take, occasionally ten, never twenty. Their own capex cycles do not run past a decade, their lease terms are shorter still, and they optimise for exit optionality in every supply agreement they write. A utility building a combined-cycle gas plant with a thirty-year book life cannot recover the investment on a contract that expires in year seven unless the tariff loads the full recovery into those seven years, which produces a power price the hyperscaler will not pay because it can site the campus in the next state over.

FERC issued customised show cause orders via Section 206 of the Federal Power Act to each of the six regional grid operators on 18 June 2026, aiming to accelerate interconnection timelines. That solves the queue; it does not solve the mismatch between a thirty-year asset and a seven-year contract. Faster approvals mean utilities will over-commit faster, not more carefully.

The Wisconsin finance owner cannot short the utility or refuse the tariff increase. She can move, which reallocates the cost but does not reduce it, or she can wait for the state to pass a Minnesota-style cost assignment rule, which will not happen while the utilities are still lobbying the Public Service Commission on the same rate cases that are absorbing the stranded costs from the last cycle.

What the numbers say about who decides

About 1 in 6 US households were already behind on their utility bills entering 2026, and Americans could owe a combined $25 billion in unpaid electric and gas bills by the end of the year. John Steinbach was shocked to receive a $281 electricity bill in January 2026, a huge spike from the roughly $100 he'd paid the previous month. He lives in Northern Virginia, where Loudoun County supplied around 1 billion gallons of water to datacentres in 2023 and where the grid is rebuilding for a customer class that pays a different rate structure.

The finance owner's position is this: she did not approve the load forecast, she did not sign the turbine order, and she cannot audit the capital plan the utility submitted to the commission. She receives a bill that reflects all three, plus the unrecovered cost of the plants that were retired before she moved into the house. The only input she has is testimony at a public hearing the commission holds after the rate case has already been filed, and the only data she sees is the summary table in the local newspaper, which reports the proposed percentage increase but not the asset-level breakdown that would show her how much of the hike is attributable to datacentre infrastructure that has not yet been energised.

Utilities requested a record high $31 billion in rate increases in 2025, more than twice the near record from 2024. Combined with increases already implemented since 2021, consumer electricity bills have risen approximately 40%, with further increases expected as utilities file new rate cases to recover their capital investments through 2030. The ratchet only turns one way. The bill arrives whether the datacentre does or not.

Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.