Who Pays for AI?

Statistics docket

Not every AI number holds up.

28 figures on the electricity, water, chips, transformers and money behind the AI buildout. Each one is sourced, dated and graded for how much of the effect is actually AI. 2 are cleanly AI-driven. 1 is an outright scapegoating.

How the verdicts are graded
AI-drivenAI is the main documented cause in the market and period studied.
AI-contributingAI has a measurable effect alongside other documented causes.
ContestedThe evidence does not isolate AI or produce a consistent causal result.
ScapegoatedThe evidence contradicts the claim or does not show how AI caused it.
BackgroundMarket or policy context with no AI attribution assigned.

Showing 28 of 28 statistics.

Start here

The five numbers behind the claim report cards.

AI-drivenElectricitychecked 2026-07-17#
11.8%of U.S. electricity projected for data centers in 2030, LBNL reference case

Data centers really are driving national demand growth. Whether that reached your bill is a separate question.

Lawrence Berkeley National Laboratory’s reference case projects data centers will use 649 TWh in 2030 — 11.8% of U.S. electricity — with a scenario range of 9.5% to 15.3%.

What this proves
National data-center demand growth, with AI servers as a major documented driver.
What this does not prove
That AI raised any specific household’s electric bill. That effect depends on the utility, its rate agreements and who funds dedicated upgrades.
AI-drivenChips & memorychecked 2026-07-17#
3:1HBM-to-DDR5 production capacity trade ratio reported by Micron

AI buyers compete for the same memory fabs as your laptop, but no consumer surcharge has been isolated.

Micron reports that producing high-bandwidth memory for AI consumes roughly three times the manufacturing capacity needed for the same amount of standard DDR5.

What this proves
A documented production trade-off between AI memory and conventional DRAM at a leading producer.
What this does not prove
A specific dollar or percentage effect on retail RAM prices. That pass-through has not been measured.
AI-contributingGrid equipmentchecked 2026-07-17#
+41%growth in U.S. distribution-transformer demand since 2019, per DOE

Data centers added to transformer demand, but they did not start the shortage.

The Department of Energy reports distribution-transformer demand up 41% since 2019, driven by post-pandemic demand, aging infrastructure and new loads including data centers.

What this proves
Data centers adding pressure to an already strained transformer market.
What this does not prove
That AI started the transformer shortage. DOE traces it to causes that predate the AI buildout.
ContestedWaterchecked 2026-07-17#
>10,000×variation in measured workload-level data-center water use, per LBNL

There is no defensible universal water figure for an AI prompt.

LBNL’s review of data-center water use found more than 10,000-fold variation across workloads, depending on cooling design, climate, utilization and grid mix.

What this proves
Why no single per-prompt water figure is reliable.
What this does not prove
The claim that every AI prompt uses a bottle of water. No universal per-prompt estimate fits the measured range.
BackgroundGrid equipmentchecked 2026-07-17#
≈2/3of newly manufactured distribution transformers used for planned replacements, per DOE

Most new transformers replace aging equipment, so the shortage has major non-AI causes.

DOE says about two-thirds of newly manufactured distribution transformers go to planned replacements of aging equipment rather than new demand.

What this proves
The transformer shortage having large non-AI components.
What this does not prove
Data centers compete for the manufacturing capacity left after planned replacements.

The full docket

The remaining evidence, grouped by where it lands.

Power

10 statistics
AI-contributingElectricitychecked 2026-07-17#
80–90% / 36%modeled generation / transmission capacity growth needed in Virginia by 2040

Virginia’s data-center forecast requires a grid buildout of exceptional scale, but the forecast is not a bill.

Full statement and scope

Virginia’s legislative auditor estimated that serving even half of unconstrained demand would require generation capacity to grow 80% to 90% and transmission capacity to grow 36% by 2040.

What this proves
The extraordinary infrastructure scale implied by Virginia’s data-center-driven demand forecast.
What this does not prove
That every requested load will materialize, that this exact resource mix will be built or that the forecast is a bill impact.
AI-contributingGrid equipmentchecked 2026-07-17#
1–2+ yrtypical distribution-transformer lead time in 2024, per DOE

Transformer waits are long across the grid, but this number cannot blame any one AI project.

Full statement and scope

DOE-reported lead times for distribution transformers reached one to two years or longer in 2024, with large power transformers reaching three to four years.

What this proves
Long waits for grid equipment affecting utilities, housing and industrial projects alike.
What this does not prove
Attribution of any single delayed project to AI without project-specific evidence.
AI-contributingGrid equipmentchecked 2026-07-17#
+5.9%rise in the BLS power and distribution transformer PPI, Oct 2024 to Jun 2026

Transformer prices rose, but this index cannot separate AI orders from other grid demand.

Full statement and scope

The BLS producer price index for power and distribution transformers rose 5.9% between October 2024, the first observation of the current series, and June 2026.

What this proves
Continued price pressure in grid equipment markets.
What this does not prove
Separating AI-driven orders from utility replacement and electrification demand.
AI-contributingPower generationchecked 2026-07-17#
100 GWGE Vernova gas-turbine backlog plus slot reservations, Q1 2026

Gas turbine orders surged during the data-center buildout, but AI’s share is not measured.

Full statement and scope

GE Vernova reported about 100 GW of gas-turbine backlog and slot reservations in the first quarter of 2026.

What this proves
A historic order book for gas generation during the data-center buildout.
What this does not prove
How much of the backlog is AI. Electrification, coal-to-gas replacement and general load growth are also drivers.
AI-contributingPower generationchecked 2026-07-17#
$2.4BGE Vernova electrification orders to support data centers in Q1 2026

Data-center equipment orders are accelerating, but their effect on other buyers’ prices is unmeasured.

Full statement and scope

GE Vernova booked $2.4 billion of data-center-support orders in its electrification segment in Q1 2026 — more than in all of 2025.

What this proves
Rapid growth in equipment demand directly tied to data centers.
What this does not prove
A measured price effect on non-AI buyers of the same equipment.
ScapegoatedGrid accesschecked 2026-07-17#
5 yearsmedian request-to-operation time for generation projects completed in 2023

The five-year grid queue predates the AI buildout, so blaming AI for that delay gets the history wrong.

Full statement and scope

Generation projects completed in 2023 spent a median of five years in interconnection queues, according to LBNL — a backlog that predates the current AI buildout.

What this proves
Slow grid processes as a long-standing structural problem.
What this does not prove
The claim that AI caused the five-year queue. The delay predates the AI buildout.
BackgroundElectricitychecked 2026-07-17#
1.6 → 1.4estimated improvement in average U.S. data-center PUE, 2014 to 2023

Data centers became more efficient per unit of facility power, even as their total electricity use kept growing.

Full statement and scope

LBNL estimates average U.S. data-center power usage effectiveness improved from about 1.6 in 2014 to 1.4 in 2023, reducing infrastructure overhead from roughly 40% to 30% of facility electricity.

What this proves
Meaningful efficiency gains in cooling and power delivery during a period of growing computation.
What this does not prove
A decline in total data-center electricity use, useful computation per kilowatt-hour or the footprint of any individual facility.
BackgroundGrid equipmentchecked 2026-07-17#
~40,000distribution-transformer configurations DOE identifies as a manufacturing obstacle

Manufacturing complexity slows transformer supply regardless of who creates the demand.

Full statement and scope

DOE identifies roughly 40,000 distinct distribution-transformer configurations as a key obstacle to scaling U.S. manufacturing.

What this proves
Why transformer supply responds slowly to demand, whoever drives that demand.
What this does not prove
Any claim about AI. This is a supply-side constraint that predates the buildout.
BackgroundGrid accesschecked 2026-07-17#
≈2,600 GWof proposed generation capacity waiting in U.S. queues, over 95% zero-carbon

A vast supply pipeline is waiting for grid access, but data centers are not shown to be displacing it.

Full statement and scope

Nearly 2,600 gigawatts of proposed generation was waiting in U.S. interconnection queues as of LBNL’s 2024 study, more than 95% of it zero-carbon resources.

What this proves
The scale of supply waiting to connect while new demand grows.
What this does not prove
That data centers displace those projects. Generation queues and large-load processes are different procedures.
BackgroundGrid accesschecked 2026-07-17#
14%of queued generation capacity in the studied cohorts ultimately built

Most proposed generation never reaches operation, so queue size is not the same as available supply.

Full statement and scope

Across the queues for which LBNL had completion data, 19% of projects and 14% of capacity requesting connection from 2000 to 2018 had been built by the end of 2023.

What this proves
Treating the 2,600 GW generation-queue headline as developer interest rather than supply certain to reach operation.
What this does not prove
That the remaining queue has no value or that generation already under construction cannot help serve new demand.

Money

8 statistics
AI-contributingRetail billschecked 2026-07-17#
$14–$37/momodeled Dominion residential generation and transmission increase by 2040

Virginia ratepayers face a real modeled cost risk, not a current or inevitable bill increase.

Full statement and scope

Virginia’s legislative auditor estimated that a typical Dominion residential customer could pay $14 to $37 more per month in constant dollars by 2040 for generation and transmission under the modeled demand buildout.

What this proves
A quantified forward ratepayer risk in a state where data centers are the main forecast demand driver.
What this does not prove
A current bill increase or an inevitable outcome; the estimate predates later large-load rate protections and depends on demand, construction and cost allocation.
ContestedRetail billschecked 2026-07-17#
+2.1%estimated residential-price effect of data-center entry, 2010 to 2024

One study links data-center entry to higher residential prices, but it does not isolate AI.

Full statement and scope

An MIT CEEPR working paper estimates that data-center entry increased residential electricity prices by 2.1% between 2010 and 2024, with larger effects at investor-owned utilities.

What this proves
A measurable historical relationship between data-center entry, infrastructure investment and residential price increases in one causal study.
What this does not prove
That every data center raises rates, that the estimate applies to future supply constraints or that the effect came from AI rather than data centers generally.
ContestedRetail billschecked 2026-07-17#
≈−4%estimated residential-price effect of doubling data-center capacity, 2015 to 2024

Large data centers can lower average rates when spare capacity lets utilities spread fixed costs.

Full statement and scope

A competing 2026 study estimates that doubling data-center capacity reduced residential electricity prices by roughly 4% from 2015 to 2024 by spreading fixed system costs across more sales.

What this proves
Evidence that durable large loads can reduce average rates when spare capacity and economies of scale dominate.
What this does not prove
That new data centers will lower future bills where power is scarce or infrastructure is overbuilt; the authors explicitly warn that supply constraints could reverse the result.
BackgroundRatepayer protectionschecked 2026-07-17#
85% / 60%minimum contracted T&D / generation demand charges for qualifying Virginia large loads

Virginia is making large loads bear more of the risk when their contracted demand goes unused.

Full statement and scope

Beginning in 2027, qualifying Dominion large-load customers must pay for at least 85% of contracted transmission and distribution demand and 60% of contracted generation demand.

What this proves
A regulator using minimum charges and a separate customer class to put underuse and stranded-asset risk on large loads.
What this does not prove
That every residual system cost has been isolated from other customers or that comparable protections exist outside Dominion’s Virginia territory.
BackgroundLocal financechecked 2026-07-17#
38%of Loudoun County’s FY2026 General Fund revenue generated by data centers

Data centers fund a large share of Loudoun’s budget, but that does not settle their net local benefit.

Full statement and scope

Loudoun County reports that data centers generate 38% of its General Fund revenue, and it maintains a stabilization reserve for volatility in those receipts.

What this proves
Large local fiscal benefits—and meaningful revenue dependence—in the country’s most concentrated data-center market.
What this does not prove
A net-benefit calculation after incentives and public costs or a result that less mature markets can expect to reproduce.
BackgroundLocal financechecked 2026-07-17#
$0.4M–$10.8Mfive-year local tax range for the same $150M equipment example

Local tax policy can change the public return on identical equipment by more than twenty-five-fold.

Full statement and scope

JLARC found that Virginia localities could collect between $0.4 million and $10.8 million over five years from the same $150 million of data-center equipment because tax rates and depreciation schedules differ.

What this proves
Local policy changing the public return from otherwise identical equipment by more than twenty-five-fold.
What this does not prove
The complete fiscal impact of a real project; the comparison excludes real-property taxes, incentives, infrastructure and service costs.
BackgroundLocal financechecked 2026-07-17#
$928M / 90%Virginia FY2023 sales-tax savings / industry capacity using the exemption

Virginia’s tax exemption is large and widely used, but its net public cost and AI share are unknown.

Full statement and scope

Virginia’s data-center sales-and-use-tax exemption provided $928 million in tax savings in FY2023 and covered about 90% of the industry measured by megawatts.

What this proves
The large fiscal scale and broad industry reach of one state’s data-center incentive.
What this does not prove
The exemption’s net cost after induced investment and local tax collections, or how much of the exempt capacity serves AI.
BackgroundLocal financechecked 2026-07-17#
84% / 68%share of announced Virginia development investment / share spent on IT and mechanical equipment

Data centers dominate Virginia’s announced investment, but much equipment spending leaves the state.

Full statement and scope

Data centers represented 84% of capital investment across Virginia economic-development projects announced in FY2022–24, while 68% of data-center investment was IT and mechanical equipment largely sourced outside the state.

What this proves
Both the dominance of data-center capital spending and the boundary between headline investment and locally retained activity.
What this does not prove
That 84% of the investment became Virginia income, or that out-of-state equipment spending creates no local construction or tax benefit.

Materials

2 statistics
AI-contributingChips & memorychecked 2026-07-17#
+30.4%rise in the BLS storage-device producer price index, Jan 2024 to Jun 2026

Storage prices rose sharply during the AI buildout, but the index cannot assign the increase to AI.

Full statement and scope

The BLS producer price index for computer storage devices rose 30.4% between January 2024 (49.891) and June 2026 (65.056).

What this proves
Sharply higher wholesale storage prices during the AI buildout.
What this does not prove
How much of the increase was caused by AI rather than the broader server replacement cycle. The index measures the storage market as a whole.
BackgroundChips & memorychecked 2026-07-17#
−6.3%change in the aggregate BLS semiconductor PPI, Jan 2024 to Jun 2026

The broad chip price index fell, so it cannot support a claim that AI made every chip more expensive.

Full statement and scope

The broad BLS producer price index for semiconductor manufacturing fell 6.3% from January 2024 to June 2026, even as memory and storage supply tightened.

What this proves
A caution: “AI made all chips more expensive” is not visible in the aggregate index.
What this does not prove
Price movements in specific segments such as HBM, which this aggregate series does not isolate.

Place & people

3 statistics
ContestedLaborchecked 2026-07-17#
757,220U.S. electrician jobs in May 2025, per BLS

Data-center construction draws from a large workforce, but a national AI wage effect is not measured.

Full statement and scope

The Bureau of Labor Statistics counted 757,220 electrician jobs in the United States in May 2025.

What this proves
The size of the national workforce that data-center construction draws from.
What this does not prove
A national AI effect on wages. BLS data does not separate data-center work from other construction.
BackgroundSiting & communitychecked 2026-07-17#
29% / 10%Virginia sites within 200 feet of residential zoning / sites with problematic-noise reports

Noise problems occur at a minority of Virginia sites, and proximity alone does not prove harm.

Full statement and scope

JLARC found 29% of operational Virginia data-center properties within 200 feet of residentially zoned land and identified problematic-noise reports at about 10% of operational sites.

What this proves
Problematic-noise reports at a minority of sites, concentrated by location and facility design.
What this does not prove
That every nearby site caused a nuisance, reduced property values or harmed health; the distance is measured property-line to property-line.
BackgroundAir qualitychecked 2026-07-17#
<4% / ≤0.1%regional NOx / carbon-monoxide and particulate emissions from data-center generators

Backup generators are not the air-quality story. The measured regional share is small.

Full statement and scope

JLARC estimated Northern Virginia data-center diesel generators contributed less than 4% of regional nitrogen-oxide emissions and 0.1% or less of carbon-monoxide and particulate emissions.

What this proves
Backup generators being a comparatively small share of measured regional air pollution under ordinary operating conditions.
What this does not prove
The absence of exposure near an individual facility or the impact of an unusual prolonged outage when many generators run together.

For journalists, researchers and analysts

Use the numbers. Cite Who Pays for AI.

This catalog is built to be quoted. When you publish a statistic from it, credit Who Pays for AI and link to that statistic’s permalink. Readers will land on the source trail, check date, attribution grade and evidence boundary behind the number.

The Copy citation button on every entry gives you a ready-made attribution with the direct link and original source included.

The original Who Pays for AI dataset compilation is licensed under CC BY 4.0. Source data remains subject to each cited publisher’s terms.