The AI Surcharge on Your Power Bill: How Data Centers Reshape Electricity Costs

01 The Line Item That Didn't Exist Five Years Ago

Open a utility bill in a fast-growing data-center market and you may find a charge that did not exist five years ago: a tariff rider or surcharge tied to large computing loads. A rider, in utility practice, is a narrow, provision-specific charge approved by a regulator and itemized on the statement, sitting on top of the base rate. The names vary by state, but the mechanism is the same. A power company asks its public utility commission for permission to bill some or all customers for the cost of serving massive new computing facilities, and the resulting line item lands on millions of bills.

The proposals run in both directions. Some utilities frame surcharges as protection, a way to guarantee that data centers pay their own way rather than leaning on everyone else. Others propose discount structures that attract the loads in the first place, betting that enormous steady consumption will spread fixed costs so widely that rates fall for all. Both directions appear in regulatory dockets filed since 2024.

What no one in these proceedings disputes is the trigger. American electricity demand was nearly flat for two decades; it is now rising, and utility filings increasingly attribute the growth to computing. The surcharge debate is simply the billing expression of that shift.

02 Why Data Centers Move Utility Math

A large AI training campus can draw several hundred megawatts continuously, which is the demand profile of a small city concentrated at one interconnection point. Engineers describe this with load factor, the ratio of average demand to peak demand. A data center runs close to unity around the clock; a residential neighborhood peaks on hot afternoons and idles overnight. The grid investment built to serve each is sized very differently per kilowatt-hour delivered.

That profile collides with how utilities recover their costs. Most of a power company's expenses are fixed: wires, substations, transformers, and generation capacity that must exist whether electrons flow or not. Those costs have historically been spread across kilowatt-hours sold. A customer that buys enormous volumes nonstop is, under that arithmetic, cheap to serve per unit and a generous contributor to the fixed-cost pool.

The trouble is timing and growth. Serving new load requires capital spent before the revenue arrives: new transmission lines, new substations, sometimes new generation. If the projected load shows up on schedule, the math can work for everyone on the system. If it does not, the utility is left holding stranded investment, and the fight over who absorbs that risk is precisely what the surcharge proposals are about.

03 How a Rate Case Works

Electricity pricing in most of the United States is set through a rate case, a formal proceeding in which a utility files proposed rates with its public utility commission, evidence is exchanged under procedural rules, and commissioners issue an order that governs prices for years. Interveners can include consumer advocates, industrial customers, environmental groups, and increasingly the data-center companies themselves, each filing testimony on how costs should be divided.

Residential customers sit in this process mainly through state consumer advocacy offices and organized intervenors, institutions that are typically outnumbered and outspent by utility and corporate counsel. The quality of that representation shapes outcomes far more than most bill payers ever see. How fixed costs are allocated between customer classes is decided in negotiation and cross-examination, not on the bill itself.

Since 2024, large-load tariff dockets have opened across the country, in states including Virginia, Georgia, and Oregon, as commissions try to write rules for loads this size before the interconnection queue fills in. The proceedings share a vocabulary of minimum demand charges, long-term contract terms, and exit fees, and their outcomes will define who pays for the buildout for a decade.

04 The Numbers Behind the Bills

The best available baseline comes from the Lawrence Berkeley National Laboratory, whose December 2024 report to Congress estimated that US data centers consumed about 176 terawatt-hours in 2023, roughly 4.4 percent of national electricity use, up from about 76 terawatt-hours in 2018. Those figures are estimates assembled from operator reporting and modeling rather than meter readings, and they are the numbers most regulatory filings cite.

The same report projected 2028 consumption between 325 and 580 terawatt-hours, or between 6.7 and 12 percent of US electricity. A range that wide is not timid forecasting; it is an honest statement that the total depends on how fast AI capacity is actually built, how efficient the hardware becomes, and how much training demand materializes at all. The band is the finding.

The direction is what matters for rate design. Even the low projection implies adding demand comparable to a mid-sized state's entire consumption within a few years, concentrated in a handful of utility service territories rather than spread evenly. That concentration is why bills are changing in specific places first, not everywhere at once.

05 Who Pays: The Design Question

The design question for regulators is how to price a load this size. The menu includes special large-load tariffs with their own rate classes, minimum-take contracts in which a data center commits to paying for a floor of capacity whether it uses it or not, collateral requirements against abandoned projects, and staged ramp schedules that match billing to construction. Each tool aims at the same risk: utilities building ahead of demand that may not arrive on schedule.

The demand shock is already visible in wholesale markets. PJM, the grid operator serving much of the Mid-Atlantic and Midwest, runs a base residual capacity auction that pays generators to be available years ahead, and its clearing price jumped from $28.92 per megawatt-day for the 2024/25 delivery year to $269.92 for 2025/26, then to $329.17 for 2026/27 under a settlement cap agreed with regulators and market participants.

Attribution is contested. Data-center load growth is one driver cited in the auction results alongside generator retirements and tightening supply, and analysts disagree on the weighting. What is measured is the clearing price itself, which flows into the bills of customers across the PJM footprint; what is interpretation is how much of the increase belongs to computing rather than to everything else changing at once.

06 The Utility-Scale Bargain

None of this makes data centers a simple loss for a region. Construction brings years of skilled work, permanent on-site employment is small but real, and the property tax base is often transformative for the rural counties that host campuses. Utilities also argue, correctly as far as the arithmetic goes, that large steady loads spread fixed costs over more kilowatt-hours and can pull average rates down if capacity keeps pace with demand.

The backlash is equally real. Ratepayer advocates point out that the cost of transmission upgrades and new generation arrives on the system before the data center's revenue does, and that residential customers have little leverage if a tenant leaves or scales back its plans. Water use for cooling and land-use conflicts add local friction that a rate case does not resolve, and the politics of both are visible in town hearings across hosting regions.

The honest summary is that both stories are true at once, and which one dominates is decided by contract design. A campus bound by a stiff minimum-take agreement, posting collateral, and funding its own grid interconnection is a different proposition from one billed under a legacy tariff written for factories, and the difference lands on neighbors' bills.

07 What to Watch

Watch three fronts. First, co-location: proposals to connect data centers directly at power plants rather than through the grid, which has put large questions in front of the Federal Energy Regulatory Commission about who pays for the network when a giant customer opts out of it. Behind-the-meter generation raises the same allocation question from the opposite direction.

Second, contract terms. The minimum-take structures now being negotiated in individual utility dockets will harden into precedent, and their details, including exit fees, ramp schedules, and credit requirements, will matter more to future bills than any single headline price. Once a standard template spreads, it will be very hard to rewrite.

Third, the numbers themselves. Berkeley's 2028 projection band, 325 to 580 terawatt-hours, will start resolving within a year or two as actual consumption data arrives, and whether surcharges survive regulatory review, get rewritten into permanent tariffs, or quietly disappear as load materializes will tell you which theory of the buildout was right.