PJM Data Center Electricity Costs Hit $23.1 Billion, and Ratepayers May Not Get It Back

PJM data center electricity costs have reached $23.1 billion across three capacity auctions, according to the regional market’s independent monitor. Those costs represent payments to power suppliers, driven by existing and forecast data center demand. Customers across PJM’s territory will keep absorbing the auction results through May 2028.

The finding turns a familiar technology story into a utility-bill fight. Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI have promised that their infrastructure will not raise household electricity bills. Yet parts of the cost are already embedded in regional power markets before many planned facilities have opened.

The central conflict is no longer whether AI infrastructure needs more electricity. It is whether existing market rules can identify who caused each cost and charge that party accordingly. State regulators can redesign future rates, but they cannot easily unwind completed auctions or recover payments already committed to power suppliers.

PJM Data Center Electricity Costs Are Already Locked In

The $23.1 billion figure describes regional capacity-market costs, not a single nationwide surcharge placed directly on household bills.

PJM Interconnection coordinates the wholesale electricity market across all or parts of 13 states and the District of Columbia. Its territory includes Northern Virginia, the country’s largest concentration of data centers, along with major markets in Pennsylvania, Ohio, and Maryland.

PJM runs capacity auctions to ensure enough generating resources will be available several years later. Capacity payments compensate generators for being ready to supply electricity during periods of high demand. Customers ultimately fund those payments through utilities and competitive electricity suppliers.

PJM’s independent market monitor, Monitoring Analytics, examined how forecast data center demand affected three capacity auctions. It estimated that including existing and projected data center load produced a combined $23.1 billion increase in system costs.

The finding does not mean every dollar has appeared on a residential bill. Capacity costs move through wholesale contracts, state-approved rates, and utility billing schedules at different speeds. However, the auction obligations are real and cannot simply be canceled after regulators adopt better policies.

The market monitor said large data center additions had created a significant impact that customers would pay through May 2028. It also warned of further effects through transmission expenses, energy prices, and future capacity auctions.

The scale became clearer during 2025. Total wholesale power costs in PJM climbed to $67 billion, up from $43.5 billion one year earlier. Capacity costs rose 262 percent and increased their share of total wholesale costs from 6.5 percent to about 16 percent.

Energy costs also rose, and transmission expenses continued growing. Data centers were not the only factor affecting every category. Fuel prices, power-plant retirements, transmission constraints, and market rules also influenced the final numbers.

Still, Monitoring Analytics identified forecast data center load as the principal reason for the capacity-market increase. Its capacity market analysis also found a widening gap between PJM’s available capacity and its reserve target.

That gap was about 210 megawatts for the 2026 to 2027 delivery year. It expanded to approximately 6,520 megawatts for 2027 to 2028. A tighter reserve margin makes additional demand more expensive because the market must secure increasingly scarce capacity.

These results matter beyond PJM. The region offers an early view of what happens when large computing projects collide with slow power-plant development and constrained transmission infrastructure.

The costs also reveal an important timing problem. A data center can be announced, delayed, resized, or canceled after its expected demand has already influenced grid planning. Power-market commitments, however, often remain after the original forecast changes.

That mismatch explains why recovering the money will be difficult. Regulators can allocate future investments differently, but auction charges already flowing through the system have become part of the cost of serving everyone.

Why AI Power Demand Is Rewriting Grid Forecasts

Data centers are forcing utilities to plan for sustained electricity growth after roughly two decades of comparatively flat national demand.

AI systems require large clusters of specialized processors, cooling equipment, networking hardware, and backup systems. These facilities can consume power continuously, unlike homes and offices whose demand rises and falls more predictably.

The Department of Energy reported that American data centers used about 176 terawatt-hours of electricity in 2023. That represented approximately 4.4 percent of total national consumption, up from 58 terawatt-hours in 2014.

An updated Lawrence Berkeley National Laboratory assessment projects an even larger expansion. Its reference case estimates data center consumption at 649 terawatt-hours in 2030, equal to about 11.8 percent of national electricity use.

The laboratory’s scenarios range from 578 to 782 terawatt-hours. The variation depends on chip shipments, server utilization, hardware lifetimes, and the power consumed while AI equipment remains idle.

These are forecasts, not guaranteed outcomes. Yet utilities must make investment decisions before anyone knows which scenario will materialize. New power plants and transmission projects can take years to permit, finance, and build.

That uncertainty creates two opposite risks. Underbuilding can leave the grid short of capacity and expose customers to reliability problems. Overbuilding can create stranded assets, which are facilities whose costs remain after the expected customer disappears.

The updated energy use forecast shows why neither risk is theoretical. Even its lower scenario implies a major change in how much electricity the country must generate and deliver.

A Dallas Federal Reserve working paper reached a similar conclusion through a different method. Its researchers modeled hourly power dispatch across wholesale markets in the continental United States.

They estimated that existing data centers had already raised average wholesale electricity prices by 3 to 5 percent nationwide. Effects were substantially larger in regions containing major data center corridors.

Under a moderate construction scenario, the model projected wholesale prices could rise 20 percent by 2028. A high-utilization scenario produced an increase approaching 50 percent.

Those estimates should not be read as a guaranteed increase in household bills. Retail prices also include distribution, transmission, taxes, existing contracts, and decisions by state regulators. Each utility passes wholesale changes through differently.

However, the wholesale price model strengthens the case that data center concentration matters. A large facility creates more pressure when it enters a constrained market than when it locates near surplus generation.

Data center operators can also change their electricity use more easily than households. They can shift computing jobs between facilities, reduce workloads during critical hours, or rely on batteries and backup generation.

That flexibility can benefit the grid when contracts reward genuine demand reductions. It can also complicate cost allocation when rates focus heavily on a customer’s demand during a few system peaks.

A facility that avoids those hours may reduce its assigned share of certain costs. Yet its anticipated year-round demand might still have influenced decisions to build generation, substations, and transmission lines.

Residential customers have little comparable flexibility. Families cannot relocate their electricity demand to another state or negotiate a custom tariff. Their heating, cooling, cooking, and medical needs also continue during grid stress.

The regulatory challenge is therefore larger than measuring total energy consumption. Officials must determine which investments each customer caused, who benefits from shared upgrades, and who carries the risk if projections fail.

The Real Fight Is Cost Allocation, Not Electricity Use

The primary dispute pits technology companies’ promise to pay their way against utility rules that routinely spread infrastructure costs across many customers.

Regulated utilities recover approved expenses through rates. State commissions examine power plants, transmission lines, substations, fuel, maintenance, salaries, and purchased electricity. They then divide those costs among residential, commercial, and industrial customer classes.

Some investments clearly serve one customer. A short line connecting a new data center to a nearby substation is an obvious example. Regulators can require the developer to fund that dedicated connection.

Shared infrastructure is harder. A new transmission line might be triggered by a data center but later improve reliability for surrounding communities. A power plant built for one project can also supply the wider system.

Utilities and developers can argue that those broader benefits justify sharing costs. Consumer advocates can respond that the project would not exist without the new industrial load.

The dispute becomes more difficult when a project receives economic-development incentives. State leaders may value construction spending, tax revenue, and technology investment. Those goals do not automatically align with the commission’s obligation to maintain affordable electricity.

In March 2026, seven major technology companies signed a federal Ratepayer Protection Pledge. They committed to build, bring, or buy new power and cover delivery upgrades required by their data centers.

The pledge also called for separate rate structures and payments for infrastructure, whether or not the companies eventually use all the electricity. That final commitment addresses the risk of a project being abandoned after a utility starts construction.

The principles are significant, but a pledge does not change a utility bill by itself. State commissions must translate those promises into tariffs, contracts, minimum payments, collateral requirements, and exit charges.

A tariff is a regulator-approved set of prices and service conditions for a customer class. Its details determine whether a data center covers only its direct connection or also pays for wider generation and transmission expansion.

Brookings researchers argued that separate tariffs offer a workable framework for protecting other customers. They also stressed that enforcement depends on state legislatures, governors, and utility commissions.

Their ratepayer protection review highlights the core reversal. Industry and government have reached broad agreement on the desired outcome, but existing institutions still control implementation.

Microsoft’s 2026 proposal in Nevada shows what a more detailed system can look like. The framework would divide infrastructure between a customer-funded share and a broader system-benefit share.

Project-specific assets would appear on a public schedule. Large customers could pay upfront or through continuing facility charges. An exit payment would cover unfinished obligations if a customer departed early.

The proposal would also allow a developer to bring generation from another provider. Accredited capacity could then be included in the utility’s planning, reducing the chance of duplicate construction.

Yet even this approach leaves room for disagreement. Regulators must decide when an asset provides a system benefit and how much of its cost belongs in the general rate base.

A substation can begin as project-specific infrastructure and become useful to other customers later. Conversely, a supposedly shared asset might mainly support a private computing campus for most of its operating life.

The fight therefore cannot be solved by saying data centers should pay their fair share. Regulators need rules that define the share before utilities commit capital and before political pressure reshapes the record.

Flexible Data Centers Can Help the Grid and Complicate the Bill

Demand flexibility can reduce peak stress, but it does not automatically repay infrastructure costs that customers have already incurred.

Electric systems are designed around peak demand, which is the highest level of consumption the network must serve during a defined period. Avoiding that peak can reduce the need for rarely used generation and network capacity.

Data centers possess several tools for doing so. Operators can delay nonurgent computing, shift workloads to another region, use batteries, or temporarily reduce processor activity.

Some facilities also maintain backup generators for reliability. When environmental permits and grid rules allow it, those resources can support the system during emergencies.

This flexibility has genuine economic value. A customer that reliably reduces demand during scarcity should not be charged as though it contributed fully to the peak.

The problem appears when a narrow measurement captures only one part of the customer’s impact. A data center might avoid selected peak hours while still driving the long-term forecast used for new infrastructure.

For example, a utility may plan generation around the facility’s expected contract demand. The regional market may also procure capacity based on the expectation that the load will exist several years later.

If the operator curtails at the right moment, it can lower some demand-based charges. That result does not erase the earlier planning decisions or the financing commitments behind them.

Regulators must distinguish between operational flexibility and infrastructure responsibility. The first concerns how a customer behaves on a particular day. The second concerns why the utility built long-lived assets.

A well-designed rate can reward both behaviors without confusing them. It can compensate verified demand reductions while retaining minimum payments for facilities built to serve the project.

Take-or-pay provisions are one option. They require a customer to pay for an agreed quantity of service even when actual consumption falls below that amount.

Long-term contracts, upfront contributions, and exit fees can serve related purposes. Each mechanism tries to prevent unused infrastructure from becoming a general customer obligation.

However, stronger commitments create another tradeoff. If requirements become too rigid, operators may build isolated power systems or choose jurisdictions offering looser rules.

That competition can weaken state protections. Governors want investment, utilities want new load, and technology companies want faster grid connections. Residential customers usually enter the process through small public-advocate offices with limited resources.

Utilities and hyperscalers can retain engineers, economists, lawyers, and rate-design specialists. Households participate indirectly and rarely have the information needed to challenge individual assumptions.

Regulatory proceedings also move slowly. A commission might approve a general rate case years after the utility began planning an asset. By then, canceling the investment can be more expensive than completing it.

This timing advantage helps explain why PJM data center electricity costs are difficult to recover. The market has already paid suppliers for capacity commitments based partly on projected demand.

A future data center tariff might prevent another transfer. It cannot retroactively identify every beneficiary of completed capacity auctions or demand repayment from facilities under older contracts.

The most credible policy response therefore combines two layers. Regional market reforms must address how large new loads enter capacity planning. State tariffs must govern how utilities recover local infrastructure costs.

Fixing only one layer leaves another path for cost shifting. A data center can pay for its dedicated substation while still influencing regional capacity prices. It can also secure capacity privately while leaving distribution upgrades in the general rate base.

The $23.1 Billion Claim Has Important Limits

The headline figure is consequential, but it should not be treated as a complete measure of every data center’s effect on every American bill.

First, the estimate covers the PJM market rather than the entire United States. PJM spans a large and economically important region, but its capacity rules and supply constraints are not universal.

Texas uses a different wholesale-market structure. Other regions rely more heavily on vertically integrated utilities, bilateral contracts, or different capacity arrangements.

Second, the estimate measures increased system costs in selected auctions. It is not a calculation of cash transferred directly from residential customers to technology companies.

Commercial and industrial customers also pay electricity charges. State policies, retail contracts, and utility structures determine how auction costs reach each class.

Third, the market monitor’s estimate relies on a counterfactual. Analysts must calculate what auction prices and quantities would have been without existing and forecast data center demand.

That approach is standard economic analysis, but assumptions matter. Changes in generation supply, plant retirements, reserve requirements, and market rules can alter the estimated difference.

Fourth, some infrastructure built for data centers can create public benefits. New transmission can reduce congestion, and additional generation can improve reliability when available to the wider market.

Large customers also spread existing fixed costs across more electricity sales. If a data center enters a system with ample capacity and pays appropriate rates, other customers can benefit.

Amazon has cited commissioned research arguing that data centers generally did not cause rate increases in areas where prices rose. The company also points to utility projections that additional demand can lower average residential rates in some service territories.

A 2026 academic paper reached a similarly skeptical conclusion about the national historical record. It estimated that data centers modestly reduced average retail rates between 2015 and 2024.

That finding does not directly invalidate PJM’s auction analysis. The studies examine different periods, geographic levels, market mechanisms, and outcome measures.

Historical data also reflects a smaller generation of facilities than the campuses now proposed for AI. A national average can hide sharp increases in constrained regional markets.

The Dallas Fed model found precisely that pattern. Average national wholesale effects were smaller than those in major data center corridors.

Both claims can therefore contain useful information. Data centers can lower average rates when they use spare infrastructure and reliably cover new costs. They can raise prices when rapid growth collides with limited supply.

The key variable is not the label “data center.” Location, timing, utilization, contractual commitments, and market design determine the impact.

Another uncertainty concerns whether announced projects will actually open. Developers often explore several sites before selecting one. Utilities can also receive overlapping requests from customers that will never build every proposed campus.

If planners treat all requests as firm demand, they can overstate future needs. If they discount legitimate projects too heavily, they can leave the system unprepared.

The resulting infrastructure has a much longer life than an AI hardware cycle. Power plants and transmission lines can operate for decades, while computing economics can change within a few years.

Efficiency improvements could reduce electricity needed for each AI task. Growing usage could offset those gains by making computation cheaper and more widely available.

These uncertainties argue for enforceable contracts rather than confident forecasts. The customer requesting new infrastructure should carry much of the risk that its forecast proves wrong.

They also argue against presenting $23.1 billion as a final national verdict. It is better understood as evidence that one major market has already made costly commitments before completing its rules for unprecedented load growth.

Three Signals Will Show Whether Ratepayers Are Protected

The next phase will be decided by enforceable market rules, not another round of voluntary promises.

The first signal is PJM’s response to the market monitor. Monitoring Analytics has proposed separating data center capacity needs from the ordinary capacity auction.

Under that approach, new data center demand would procure new generation through a dedicated auction and long-term contracts. The facilities would face curtailment when adequate supply was unavailable.

PJM has said the questions are moving through its stakeholder process. Proposals are expected to reach the Federal Energy Regulatory Commission, which oversees interstate wholesale power markets.

A rule that links new loads to new supply would strengthen the case that future costs can be contained. A broad backstop funded through existing customers would weaken it.

The second signal is whether states approve separate large-load tariffs with meaningful minimum payments. Microsoft’s Nevada tariff proposal offers one detailed test.

Regulators should examine project-specific asset tracking, contract demand, collateral, exit charges, and treatment of shared benefits. A tariff’s name matters far less than those provisions.

The strongest rules will require payment when usage falls below projections. They will also prevent customer-funded assets from quietly entering the general rate base without transparent review.

The third signal is the difference between forecast and operating demand. Regulators need public reporting on proposed loads, signed contracts, construction progress, actual consumption, and canceled projects.

A shrinking gap would show that utilities are planning around credible commitments. A widening gap would warn that customers are financing infrastructure for speculative requests.

The same reporting should identify how much capacity comes from new generation. Merely moving existing supply between customers does not solve the underlying shortage.

These signals will unfold across different institutions. FERC controls wholesale-market decisions, state commissions approve retail tariffs, and local authorities handle many permits and incentives.

That fragmentation is why recovery remains so difficult. No single regulator can trace every data center cost from a chip purchase to a household bill.

PJM data center electricity costs also demonstrate the price of acting after auctions clear. The region can improve future rules, but completed obligations will continue through May 2028.

For households and small businesses, the most useful action is to follow utility commission proceedings rather than corporate sustainability announcements. Rate cases and tariff dockets contain the terms that determine who ultimately pays.

For technology buyers, the issue belongs in AI’s total cost calculation. Cloud services can hide electricity infrastructure behind usage charges, but they cannot remove it from the economy.

Companies evaluating AI projects should ask providers where computing capacity is located, how its power is contracted, and whether local customers carry infrastructure risk. Those questions increasingly affect permitting, reliability, and long-term service availability.

The $23.1 billion finding is not proof that every data center harms consumers. It is proof that demand can enter power markets faster than protections enter utility rules.

The next test is concrete: Will regulators assign future costs before construction begins, or will households again discover them after the commitments become irreversible?