Do Data Centers Raise Electric Bills? What the Data Shows

The strongest public objection to data centers—that they are raising household electric bills—is well supported in the PJM region and only weakly supported as a national average story. Water stress is a real siting problem in the arid West, but onsite cooling volumes are often smaller than the politics imply, and the bigger water footprint is usually the power plants behind the meter. State tax exemptions are large, frequently poorly targeted, and buy relatively few permanent jobs; local property-tax gains can still be real where abatements are limited. In the 2026 midterms these issues were highly salient as a local face of the cost-of-living vote, not as the country’s top national issue.

Electricity bills: a clear PJM mechanism, a contested national one

PJM’s capacity market is the cleanest causal path from data centers to residential bills. The auction does not buy electricity. It pays generators years ahead to be available at peak, and the cost is socialized across load in 13 states and D.C. When the peak forecast jumps, everyone pays a higher insurance premium—even households nowhere near a campus.

Prices collapsed the old baseline. The 2024/25 delivery year cleared at $28.92 per MW-day. The 2025/26 auction cleared at $269.92, then 2026/27 at $329.17, 2027/28 at the cap of $333.44, and the July 2026 auction for 2028/29 at the cap of $325. Each of the recent auctions cost about $14.7–16.4 billion. PJM itself attributed nearly 5,100 MW of a 5,250 MW forecast increase for 2027/28 to data centers, and both of the latest auctions finished roughly 6.6–6.8 GW short of the 20% reserve-margin target. [1] [2] [3] [4]

The independent market monitor, Monitoring Analytics, has quantified the data-center share as a counterfactual: auction outcomes with versus without data-center load in the forecast. Joseph Bowring’s figures, repeated across Utility Dive and the monitor’s own presentations, put data centers at $6.3 billion, or 38%, of the $16.4 billion 2028/29 auction, and $29.4 billion, or about 46%, of $63.6 billion across the last four auctions. Earlier vintages were even more data-center-driven on the increase: inclusion of existing and forecast data-center load raised 2026/27 capacity revenues by about $7.3 billion (82%). In the first half of 2026, that capacity effect alone raised PJM’s wholesale cost by $11.11/MWh, or 9.7%, from $103.40 to $114.50—explicitly excluding energy and transmission. Most of the latest attributed dollars are for facilities that have not been built; the monitor has called the forecast uncertainty “unprecedented.” [5] [6] [7] [8] [9] [10]

Household translations of that wholesale shock are estimates, not meter reads, but they cluster. Researchers and IEEFA-linked estimates put the mid-2025 to mid-2026 capacity increase at roughly $16 a month on a typical Ohio bill and $18 in western Maryland, with about 63% of the 2025/26 price jump attributed to data centers. Those figures are the ones campaigns used. They are directionally consistent with the auction math; they are not a full decomposition of every utility bill. [11] [12] [13]

Outside PJM, the evidence is thinner and often points the other way. Lawrence Berkeley National Laboratory and the Brattle Group found that national average retail prices rose about 29% from 2019 to 2025 in nominal terms but only about 3% after inflation, and that state-level load growth has historically been associated with lower average prices because fixed grid costs are spread over more kilowatt-hours. They flag PJM’s capacity market as the clear exception, driven by load growth colliding with constrained supply. An MIT Center for Energy and Environmental Policy Research working paper estimated that data-center entry from 2010–2024 raised U.S. residential rates by about 2.1% on average (5.6% at investor-owned utilities; essentially zero at cooperatives). A Dallas Fed dispatch model found existing data centers had already raised wholesale prices by roughly 2–6% nationwide by 2024–25, with secondary summaries citing about 6.9% nationally and about 15% in Northern Virginia’s Dominion zone and the Northwest—wholesale, not the full retail bill, and sensitive to utilization assumptions. [14] [12] [15] [16] [12] [17]

FactCheck.org’s October 2026 review is the right frame: data centers have likely raised bills in PJM; nationally, fuel prices (especially gas) and replacement of aging wires remain the larger documented drivers, and “data centers raised your bill” is not uniformly true. Regulators have started treating cost-shift as a design problem rather than a settled fact. Virginia’s State Corporation Commission created a GS-5 class for loads of 25 MW or more, with long contracts and minimum charges, effective 2027, in the same order that raised a typical Dominion residential bill by roughly $11 in 2026 for reasons that included but were not limited to data centers. Ohio’s AEP tariff requires new large data centers to pay for at least 85% of contracted capacity even if they use less. Those tariffs are evidence that the risk is real and that the historical socialization of costs is no longer politically acceptable. [12] [18] [19]

What this means for anyone entering or defending the industry: in PJM, “we pay for the power we use” does not answer the objection. Capacity is priced on the forecast peak, and that cost is spread. Bring-your-own-generation, bilateral contracts, and large-load tariffs are the mechanisms that actually change who pays. Elsewhere, a blanket claim that data centers are the main reason national residential rates rose will not survive contact with LBNL or utility-level rate cases.

Water: location is the well-supported claim; “they are draining the West” is not

The siting fact is solid. A 2026 Environmental Science & Technology mapping of 9,347 global data centers, led by researchers at Stanford and the Carnegie Institution, found that 25% already sit in places with at least one month a year when demand exceeds supply, and that share could rise to about 33% under warmer hydrology. About 6% face stress most of the year. Hotspots include the western United States, Spain, Italy, India, and the Gulf. The authors are explicit that exposure does not yet equal a measured harm to other users: “the extent to which water-stress exposure translates into operational impacts… remains an open question.” A separate U.S. inventory finds more than one in three American data centers in high or extremely high water-stress areas, and all 57 of Arizona’s in extremely stressed basins. [20] [21] [22] [23]

Onsite volumes can be large in absolute terms and small relative to a metro’s total use. Google, the only hyperscaler that publishes site-level water, reported U.S. data centers consuming about 6.1 billion gallons in 2024 and about 10.5 billion gallons consumed (13.6 billion withdrawn) in 2025. Council Bluffs, Iowa—not a desert site—is routinely the thirstiest campus, on the order of 1.0–1.7 billion gallons a year. Google has said only a small share of its U.S. withdrawal is from high-scarcity sources, and its Phoenix-area sites use air cooling or closed loops precisely because the basin is stressed. Microsoft has piloted near-zero-water designs in the Phoenix area. That shift is the non-obvious mechanism: rejecting evaporative cooling in a drought region cuts onsite water and raises electricity use, which can increase both bills and the indirect water used to generate power. [24] [21] [25] [26]

Indirect water is where the numbers diverge and should be treated as contested. A 2024 Lawrence Berkeley estimate of U.S. data-center electricity-related water was about 211 billion gallons; environmental analyses such as Ceres have produced figures an order of magnitude higher, depending on how thermoelectric cooling and grid mix are counted. An arXiv mapping of 472 U.S. hyperscale sites put total operational water near 300 billion liters a year under baseline assumptions, with electricity-related water about three-quarters of the total, concentrated in fossil-heavy eastern grids—not the same map as onsite cooling stress in the West. [27] [28]

Phoenix is the political test case and also the best check on exaggeration. Ceres estimated active Phoenix-area data centers used on the order of 385 million gallons a year and projected that could rise toward 3.7–3.8 billion if the pipeline is built—an almost tenfold jump, and enough for the group to warn of higher regional water stress. Independent readings of that same projection put even the high case at roughly 1% of regional residential use and under 0.5% of total regional water use. Power-plant water tied to serving that load was projected far higher, up to about 14.5 billion gallons a year. Agriculture, golf, and municipal demand remain larger onsite users. A federal cut to Arizona’s Colorado River allocation and ongoing groundwater depletion are real; data centers are an incremental, highly visible claimant, not the basin’s main drain. [28] [29] [30] [31]

For a project in a drought basin, the defensible position is not “we use less than farms.” It is metered onsite use, a stated cooling technology, a reclaimed-water plan, and an honest accounting of the electricity-water tradeoff. Communities are rejecting projects on water even when the gallons are modest, because the Colorado River and local aquifers have no slack and nondisclosure agreements read as concealment.

Tax exemptions and jobs: big forgone revenue, few permanent jobs, real local tax base in some counties

The fiscal objection is the best-documented of the three, because state auditors—not advocates—have run the numbers.

Virginia’s sales-and-use exemption for data-center equipment is the largest single economic-development incentive in the state. JLARC found it accounted for $2.7 billion, or 53%, of incentive spending from FY2015 through FY2024, and about $1.0–1.02 billion in FY2024 alone. The state’s annual financial report put FY2025 forgone revenue above $1.6 billion. JLARC’s return calculation is about 48 cents of state revenue per dollar forgone—better than the average Virginia incentive, still a net loss to the state treasury. A typical 250,000-square-foot building employs about 50 people, roughly half of them contractors, against about 1,500 workers at peak construction. Industry-sponsored totals of roughly 74,000 “supported” jobs and $9.1 billion in GDP are mostly construction, equipment, and multiplier effects, not permanent on-site employment. Loudoun County’s local tax share has historically been large; that is a local win financed in part by a state exemption. [32] [33] [34] [35] [36]

Georgia’s audit is the cleanest “but for” test. The University of Georgia’s Carl Vinson Institute, for the Department of Audits and Accounts, estimated that only about 30% of data-center activity depended on the high-tech equipment exemption; roughly 70% would have been built anyway. FY2025 state revenue forgone was about $474 million. A summary initially credited the program with 28,350 construction jobs and 5,471 operations jobs; the corrected, exemption-attributable figures were 8,505 and 1,641. Net state fiscal impact is negative. Later official projections put the high-tech exemption near $1.4 billion in FY2027 ($762 million state, $645 million local in one accounting), with broader computer-equipment breaks pushing combined losses higher still. Good Jobs First’s reading—that the break is a poor bargain once additionality is applied—tracks the auditor’s own but-for finding, even if advocates and the university differ on how to present gross economic output. [37] [32] [38] [39] [37]

The pattern repeats. NCSL reports that at least 10 states forgo more than $100 million a year, with Texas and Virginia each in the billion-dollar range in some years, and that Illinois subsidized centers often created exactly the statutory minimum of 20 jobs. An Illinois Economic Policy Institute projection of $57 billion in investment through 2035 produced about 121,000 jobs, almost all temporary or induced, and only about 2,800 permanent direct positions. A Gargano and Giacoletti study of incentives in 38 states found they do cause more large cloud and AI facilities to be built—fiscal competition works—but found no significant local tech-employment effect in most states. Their back-of-the-envelope for Ohio, the one state with a detectable tech-job response, was more than $140,000 and possibly $500,000 in incentive per induced job. Their model implies a coordinated repeal would raise combined state welfare. [40] [41] [42] [43]

Independent job studies cut against both “no jobs” and “transformational employer.” A study of 1,584 U.S. data centers opened from 2000 to 2023 (Buechler, Rolheiser, and colleagues, circulated via MIT’s Center for Real Estate and The Conversation) found counties had about 1,000–1,300 more non-construction jobs than comparable counties three to five years after opening, or about 2.7 non-construction jobs per megawatt once later capacity growth is netted out. Gains showed up in transportation, technical services, and real estate, not in local retail and restaurants. A Brookings analysis of roughly 1,500 built facilities versus 52 announced-but-canceled projects found the first large data center raised data-processing employment 56% and telecommunications 43% over a decade—implying only about 100–200 jobs in a typical treated county—with no wage effect. Both designs are stronger than industry multipliers. Neither asks whether the jobs are worth the power, water, and tax expenditure. Industry figures such as PwC’s 5.5 million “supported” jobs for the Data Center Coalition count direct, indirect, and induced employment on a national input-output basis and are not comparable to auditor counts of permanent on-site staff. [44] [45] [46] [47]

Local fiscal math can still favor a host city if the facility pays full property tax. A Maryland Tech Council study of a hypothetical 660,000-square-foot center in Washington County estimated 86 direct operating jobs and about 256 direct-plus-secondary jobs, against more than $12 million a year in city and county property tax—several times a warehouse, and equivalent to thousands of new homes the county would otherwise have to service. That is the actual bargain communities are offered: a small permanent payroll, a large taxable building, and a state that may be giving away the equipment tax. [48]

How salient this was for 2026 voters

Opposition to a data center near me became a majority position in 2026 and stayed there through Election Day. It was not the top national issue. It was one of the most concrete ways voters talked about affordability, and it showed up in ballots.

Pre-election surveys:

• Heatmap Pro / Embold Research, October 3–15, 2026 (2,537 registered voters): 71% would oppose a data center planned for their area, down slightly from 75% in August—the first time the series did not worsen. The number-one reason to oppose was that it could raise power or water bills. The number-one reason to support was tax revenue. A plurality wanted building to continue only with stronger protections for electricity customers and neighbors. [49] [50] [49]
• POLITICO Poll, late October 2026: 50% opposed a data center within three miles of home, up from 28% in January. Voters had flipped from a slight preference for pro-data-center candidates to a preference for opponents. [51] [52]
• America Is All In / Global Strategy Group, September 2026, likely voters: data centers were 28% favorable and 57% unfavorable. Seventy-two percent had heard a lot or some about them; 84% said at least some of what they heard was negative, driven by water, electricity, bills, and noise. Support for a local project was net −33. Thirty-eight percent named rising energy demand, including data centers, as a top-three reason their bill was up—behind gas prices (59%) and politicians failing to contain costs (41%). [53]
• Consumer Reports, released October 21, 2026 (2,003 adults; fieldwork in spring): 81% of people in the same home for at least three years said their electric bill had risen; 38% cited data centers and AI as a cause, second to utility profits at 45%. Eighty-four percent supported a separate rate class for large users. Half said data centers should pay all infrastructure built mainly for them. [54] [55]
• USC Annenberg, early November 2026: nearly one-third of Americans ranked AI and data centers among the three most important issues facing the country—behind inflation and the cost of living (86%) and political polarization (35%), and ahead of immigration, climate, crime, and unemployment. Seventy-seven percent said government is doing too little to regulate data centers. [56]

Election Day sharpened rather than invented the pattern. The Verge, citing an NBC News Election Day poll, reported 76% opposed data centers in their communities. A Washington Examiner read of NBC exit polls found 56% said AI and data centers played a crucial or major role in their vote, with about two-thirds of that group voting Democratic. The AP Voter Poll, by contrast, put the economy and the state of democracy on top, and described AI data centers as less often “crucial” than corruption or the cost of living—consistent with data centers mattering as a subset of affordability, not as a standalone national priority. Local results were lopsided: trackers counted on the order of 33 to 37 restrictive ballot measures passing, including multiple Ohio bans on centers above 25 MW, against one or two approvals. Campaigns in PJM states explicitly tied opponents to data-center deals; reporting on November 4 described the issue as the way the cost-of-living vote showed up in the Midwest and Mid-Atlantic. [11] [11] [57] [58] [59] [60] [61]

What is well supported, what is contested, and what follows

Claim	Status	Best evidence
Data centers are the primary driver of the PJM capacity-price spike and a large share of recent capacity costs	Well supported	PJM auction reports; Monitoring Analytics counterfactuals ($6.3B / 38% latest; $29.4B / ~46% over four auctions)
That spike has added on the order of $15–20/month to some PJM residential bills	Supported as an estimate, not a universal meter result	IEEFA and academic translations of capacity charges; utility-specific rate cases differ
Data centers are the main reason U.S. residential rates rose	Contested / overstated	LBNL/Brattle: real national prices ~+3% since 2019; load growth often lowers average prices outside constrained markets; MIT ~+2.1% residential from entry
Data centers sit disproportionately in water-stressed places	Well supported	Stanford/Carnegie global map (25%); U.S. stress inventories; Arizona concentration
They are a major driver of Western water scarcity relative to agriculture and cities	Contested	Phoenix high-case onsite use still ~1% of residential demand in one reading of Ceres; indirect power-plant water is the larger, less certain term
State sales-tax exemptions cost hundreds of millions to billions and often fail a but-for test	Well supported	Virginia JLARC (~48¢ back per $1; $1B+ a year); Georgia UGA (70% would have located anyway; net fiscal loss)
Data centers create large numbers of permanent local jobs	Poorly supported	Auditors: ~50 staff per large building; Brookings ~100–200 county jobs; minimum-job statutes often met exactly
Data centers create some local employment beyond the building, mostly in trades, logistics, and technical services	Supported, modest	1,584-facility study: ~2.7 non-construction jobs per MW; not retail
Voters in 2026 treated this as a top-tier local affordability issue	Well supported	61–76% local opposition across polls; bill and water as top reasons; restrictive ballot measures passed widely
Data centers were the number-one national voting issue	Not supported	AP and other exit polls: economy and democracy first; USC: AI/data centers a top-three issue for about one-third, far behind cost of living

The political implication is narrower than the slogan. Voters are not mainly objecting to AI as such. They are objecting to a cost-allocation regime in which forecast peaks, shared wires, and automatic equipment exemptions socialize risk while the permanent payroll stays small. The evidence says that objection is correct in PJM and in state capitols that copied Virginia’s exemption, overstated as a national retail-price story, and only partly correct on water—right about where the buildings are going, often wrong about how many gallons a modern desert campus actually evaporates. Projects that bring new generation, take a dedicated large-load tariff, publish water and cooling choices, and stop asking for exemptions they do not need are responding to the part of the backlash the record actually sustains.

---

Recent Findings Supplement (November 2026)

PJM capacity auctions have driven substantial, data-center-attributed cost increases passed to ratepayers across 13 states, with the independent market monitor providing consistent quantification.[1]

The July 2026 auction (for 2028/29 delivery) cleared at the $325/MW-day FERC-approved cap for a total of ~$16.4 billion but fell 6,831 MW short of reliability targets (reserve margin 14.7% vs. 20% goal). Monitoring Analytics attributes $6.3 billion (38%) of these costs directly to data center load forecasts; across the last four auctions, the figure reaches $29.4 billion (46% of $63.6 billion total). Data center load also accounted for 9% of PJM wholesale power costs ($10.48/MWh) through July 2026.[2]

These costs are socialized across all customers via capacity charges. Analyses estimate additions of roughly $16/month to average residential bills in Ohio and $18/month in western Maryland, with data centers responsible for ~63% of recent capacity-driven price spikes in some estimates. Broader PJM residential rates could rise 30-60% by 2030 per ICF modeling.[3]

Well-supported in PJM: Market monitor reports and auction results are corroborated across multiple outlets and directly tie forecast data center demand to price spikes and shortfalls.[4]

More contested nationally: A MIT working paper (using 2010–2024 utility data, published Oct 2026) estimates data centers raised average U.S. residential rates by 2.1% (2.7% overall retail; higher at investor-owned utilities). A CEPR analysis finds an additional data center raises residential prices by ~0.182 ¢/kWh. Some studies (e.g., E3 whitepaper) note mixed evidence outside PJM or attribute part of increases to other factors like fuel costs and retirements.[5]

Implications for competitors/entrants: Expect continued regulatory pressure on cost allocation (e.g., FERC scrutiny of backstop procurements and “bring-your-own-capacity” rules). Large-load tariffs, minimum bill commitments, and curtailment requirements are proliferating in PJM states.[6]

New global and U.S. studies quantify data center exposure to water stress, showing 25–37% already in high-risk areas with competition for municipal supplies.[7]

A October 2026 study in Environmental Science & Technology (Rosa et al.) mapped 9,347 facilities worldwide: 25% experience at least one month of water scarcity annually (demand exceeds supply); 6% face it most of the year. Under 3°C warming scenarios, this rises to 33%, with most added exposure in the U.S. 41% of sites compete primarily with municipalities for water.[8]

U.S.-focused analysis finds 37% of data centers in high/extremely high water-stress areas and 26% in both stress and drought for at least half the year; all 57 Arizona sites are in extreme stress. A hyperscale inventory estimates ~300 GL/year operational water use (three-quarters electricity-related). Phoenix-area projections: ~3.8 billion gallons/year direct use.[9]

Well-supported: Consistent findings across peer-reviewed mapping studies on current exposure and climate projections.[10]

Contested/less quantified: Direct operational impacts on other users or measurable shortages attributable solely to data centers remain open questions; many facilities use reclaimed or non-potable water, though electricity-generation water use dominates totals in some models.[11]

Implications: Siting in drought-prone regions (e.g., Arizona, parts of the West) faces heightened local opposition and potential moratoriums. Strategies like air cooling, non-potable sources, or co-location with power plants using less water-intensive generation will be scrutinized.[12]

Recent cost-benefit analyses highlight high taxpayer costs from tax exemptions relative to limited permanent jobs, with net benefits turning marginal or negative under accelerated growth.[13]

Scioto Analysis (Ohio, Oct 2026) estimates 2013–2026 data centers delivered $62 billion net social benefit (96,000 jobs, $75 billion wages) but projects only $2.4 billion net (BCR 1.04) for 2027–2034 under current rules; a ban yields $35 billion NPV. Good Jobs First reports states lose >$3 billion annually to incentives (e.g., Virginia ~$1.6 billion FY2025; Georgia high-tech exemption $1.4 billion FY2027, up 61%; Minnesota >$1 billion over a decade; Pennsylvania projected $1.9 billion over seven years). Typical facilities employ ~50 full-time staff; subsidies can exceed $1 million per job.[14]

Some jurisdictions see revenue gains (e.g., Loudoun County, VA: data centers ~4% of parcels but 38% of general fund revenue).[14]

Well-supported: Low permanent employment and rising exemption costs are documented in multiple state analyses and reporting.[15]

Contested: Overall net benefits depend on policy (e.g., behind-the-meter generation or longer contracts improve ratios in simulations); some local tax-base boosts occur despite state-level losses.[13]

Implications: Expect pushback on incentives, including federal proposals like Sen. Wyden’s excise tax and restrictions on bonus depreciation/Opportunity Zones. States are already scaling back or adding conditions.[16]

2026 polling and midterm ballot results show electricity costs as the dominant driver of opposition, with broad bipartisan salience in affected regions.[3]

National polls (Oct 2026) found 70%+ opposition to a local data center (up sharply from earlier in the year), with rising electricity bills the top reason; Consumer Reports: 81% reported bill increases, 84% support separate large-user rates. In PJM states, 55–63% oppose in Ohio polls; Nevada: 57% oppose new centers, 83% concerned about water/growth. Electricity concerns outrank water or jobs in most surveys.[17]

Midterms featured 41 local measures across 10 states; voters approved restrictions/bans in at least 38 jurisdictions. Candidates emphasizing affordability and opposition gained traction. AI/data centers ranked as a factor for 56% of voters but trailed cost of living and other issues.[18]

Well-supported: Multiple independent polls (Heatmap, WaPo/ABC/Ipsos, POLITICO, state-specific) converge on high opposition tied to bills.[3]

Implications: Local referenda and ratepayer protections will constrain siting; messaging around “bring your own power” or cost-shifting resonates across parties.[19]

These developments reflect accelerating pushback since mid-2026, centered on verifiable PJM cost attribution and new water-mapping data, while job/tax tradeoffs remain policy-dependent.