The data-center ledger: who gets the benefit, who pays?

Growth can spread costs. It can also outrun supply.

A 2026 working paper estimates that data-center growth lowered average U.S. residential electricity rates in 2015–2024 relative to the alternative without that growth. Its proposed mechanism is spreading fixed costs over more electricity sales. This is a statistical estimate, not a finding that every household's bill fell. Its causal interpretation depends on assumptions about where facilities locate. The authors warn that constrained supply or demand that never arrives could reverse the benefit. It adds a countercase, not a guarantee for today's buildout. Local charges and payment commitments still need checking. Watten, Bistline and Blanford: working paper, sections III–IV ↗

Oregon changed the rate split, not every household's bill.

Oregon's July decision shows why cost allocation matters. Portland General Electric's approved package lowered average residential rates while raising data-center rates. These are July 8 changes against prior rates, not surveyed bills paid or a comparison with Virginia. The package also changed power costs and other charges, so it cannot isolate data centers' contribution to the decrease.

The September 15 compliance filing says it changes no existing rates. Its contract terms include minimum payments and exit-cost protection, with possible refunds when capacity is reused or exit charges exceed stranded costs. The docket records a September 10 judicial-review petition, not a court decision. Adopted terms do not prove future costs cannot shift. Ask locally which costs a large customer must cover if it leaves.

The crew and the payroll are not the same number.

Virginia's Joint Legislative Audit and Review Commission (JLARC) reports that representatives described a typical 250,000-square-foot data center as having roughly 50 full-time operating workers, about half of them contractors. A single building can have around 1,500 workers on site at the peak of construction, which usually lasts about 12 to 18 months. Those are reported examples, not a headcount for every proposed project. A campus built in phases may sustain trade work longer; it should not relabel peak crews as permanent staff.

JLARC's statewide economic model estimated 74,000 jobs supported by the industry annually, including indirect and induced work. Most estimated benefits arose during construction. The audit also reported average annual pay of about $100,000 for data-center employees and contractors in Virginia, varying by role and region. That is a statewide figure, not a promised wage for every applicant. Read the state audit ↗

The local receipt: Loudoun County, Virginia.

Loudoun County reports $1.2 billion in real-estate and equipment-property taxes from data centers for fiscal 2026, or 39% of its overall budget. The county says the revenue has helped it increase school funding, support public safety, libraries, parks and mental-health services, and pay for capital work such as schools, roads and bridges. Over time, the county says its residential real-property tax rate generally fell from $1.145 per $100 of assessed value in tax year 2016 to $0.805 in 2026, with data-center revenue helping make that possible. That is the rate, not everyone's bill: an individual homeowner can still pay more when assessed value rises. County revenue figures ↗ County budget and tax-rate explanation ↗

Loudoun's school division says its approved fiscal 2027 budget funds additional teachers and instructional support as well as salary and benefit investments. Together, these records support the conclusion that data-center tax revenue contributes to a school system with more resources. They do not trace a particular teacher's raise to a particular server building. The county collects into a broader budget, and elected officials decide what to fund. School division's approved-budget account ↗

A tax receipt does not cancel a water bill.

JLARC documented substantial local revenue and real work. The same audit warned that rapid load growth could require difficult grid expansion and raise costs for other electricity customers. Water use varies by cooling system and site; some facilities use far more than a large office building. Noise and proximity to homes matter, too. Virginia's sales-tax exemption for qualifying data-center equipment is also a public cost to count beside revenue. None of this proves a facility near you is a bad deal. It means a local deal needs its own complete accounting. JLARC's benefits and costs ↗

A national estimate does not settle a neighborhood debate: Berkeley Lab models all U.S. data centers at 192 terawatt-hours in 2024, 4.7% of national electricity use. That is not an AI-only count or a meter reading for a proposed site. EIA's higher-demand scenario explores possible regional generation and wholesale-price effects; it is not an observed household bill. The decisive local documents are the utility tariff and rate case, interconnection and load records, water agreement, tax terms and actual operating reports. Berkeley Lab estimate ↗ EIA scenario ↗

An October 9 Senate staff report challenges how grid-expansion costs are divided; Google's original response describes infrastructure payments and binding commitments. Berkeley Lab's tariff review identifies protective mechanisms, not measured household savings. The local test is the approved cost allocation and who pays if demand falls short. Read the grid-cost record and its countercase →

October 13 follow-up: FERC accepted PJM's large-load definition and registry, effective October 12, but rejected the linked bring-new-capacity and mandatory-curtailment package. Tracking large customers is not the same as adding supply or guaranteeing lower bills. This October 9 tariff decision is separate from DOE's temporary emergency authority. Read the decision and its limits →

If a gallon counts here, count the other gallons too.

Virginia's auditors estimated that data centers used 2.1 billion gallons of water statewide in 2023, just over a third of it reclaimed rather than newly withdrawn. That was less than 0.5% of statewide withdrawals. The state's single largest industrial water user used about 36.5 billion gallons that year. That comparison punctures the idea that data centers are Virginia's dominant water user. It does not settle a local water dispute: at the six utilities the auditors reviewed, data centers accounted for 0.2% to 21% of customer water use after reclaimed water was excluded. Eleven individual buildings each used more than 50 million gallons. Place, source, season and cooling design change the answer. JLARC water findings, pp. 62–63 ↗

And yes, the rest of the water ledger deserves ink. EPA estimates U.S. residential outdoor use at nearly 8 billion gallons a day, mainly for landscaping, with as much as half lost to wind, evaporation or runoff from inefficient irrigation. That is a national estimate, not a measurement of Loudoun's lawns or leaky mains; we cannot subtract it from a Virginia data-center total. A street puddle is a storm-drainage problem, not automatically lost drinking water. To compare a proposed facility with local irrigation or pipe leakage, get the same utility's annual figures and put them in the same units. EPA irrigation facts ↗

The diesel comparison deserves the same discipline. If exhaust matters at a generator yard, it matters at a truck stop. JLARC found data-center backup generators made up less than 4% of Northern Virginia nitrogen-oxide emissions and 0.1% or less of its carbon-monoxide and particulate emissions. The auditors also warned that regional averages may miss concentrated exposure next to a facility, especially during a prolonged outage. We have no verified count of idling trucks at a comparable local stop, so we won't invent one. Compare the same pollutant, area and time period before declaring either source harmless or uniquely intolerable. JLARC air findings, pp. 58–60 ↗

Where the water goes matters.

Withdrawal is not the same as consumption: some water returns; some is lost, including through evaporation. ICPRB's March 2026 summary estimates Washington Metropolitan Area data centers at 1% of withdrawals but 9% of annual consumptive use for 2025, reaching up to 12% in summer. That is a different region and denominator from Virginia's audit, not a revised statewide figure. Its estimates combine utility records and modeling, with a 75% consumptive-use assumption. Reclaimed cooling supply can reduce drinking-water demand without eliminating evaporative losses. Small annual shares do not replace a drought-season water plan. Water summary, March 2026 ↗ Study methods, December 2025, pp. 6-12–6-20 ↗

Regional direct on-site estimates, not measurements of today's local conditions. Both links describe the same study.

Missing water returns are not zero water use.

Virginia's October 1, 2026 SCC return lists contracted demand in megawatts and generator permitting information. These figures are not annual electricity use or measured emissions. Municipal water responses were not received in time; the investigation remained ongoing. A missing return is not zero water use. The older audit and regional estimates above keep their own dates and scope. SCC return, pp. 1–2 →