Data-center water use is always a crisis—or always negligible.
Depends on the project
Virginia · 2024 review
What the evidence says
JLARC found wide variation in water use. Virginia is water-rich overall, but some local supplies are constrained.
What future projects should show
Publish annual consumption, peak summer demand, water source and drought operating limits for each campus. Compare with the supplying watershed and utility, not just national totals.
Limits & sources
A finding about Virginia cannot establish safety in an arid watershed or for a denser future campus.
Zero-water cooling means a data center has no water footprint.
Needs qualification
Microsoft new designs · December 2024 announcement
What the evidence says
Microsoft says its new closed-loop designs avoid evaporation for cooling and could avoid over 125 million liters annually per data center. Water is still used for administrative needs. This is a design claim, not proof that the existing fleet or its electricity supply uses no water.
What future projects should show
Track commissioning, measured make-up water and hot-weather electricity demand. Distinguish an internal recirculating liquid loop from the method used to reject heat outdoors.
Limits & sources
The disclosure does not provide a complete measured lifecycle water account for every future site.
One water-use number applies to every AI prompt.
Not established
Cooling and electricity supply vary by location
What the evidence says
Microsoft’s change in cooling design illustrates why a site-independent water-per-query claim needs assumptions. A single prompt is not a uniform unit of computation.
What future projects should show
Require model, task length, hardware utilization, location, season and an allocation method. Report direct cooling water separately from electricity-related and manufacturing water.
Limits & sources
We have not verified a universal per-prompt estimate. Withdrawals, water returned and water consumed must not be treated as interchangeable quantities.
Data centers are consuming most U.S. electricity.
Contradicted at national scale
United States · DOE/LBNL 2024 report vintage
What the evidence says
DOE reports an estimated 4.4% share in 2023 and a projected 6.7–12% in 2028. These are all data centers, not AI alone. A national minority share can still be a large increment on a local grid.
What future projects should show
Compare each project’s MW and annual MWh with regional available supply and upgrade schedules. Carry demand scenarios beyond 2030 without relabeling older forecasts as current observations.
Limits & sources
The 2028 range is a forecast, not capacity already delivered or a guaranteed demand outcome.
Data centers typically lower everyone’s electricity bills.
Not established
Continental U.S. wholesale markets · March 2026 working paper
What the evidence says
Dallas Fed researchers estimate existing data centers raised wholesale prices by 3–5% on average in their dispatch model. This is a modeled wholesale effect, not an observed percentage increase in household bills. Effects depend on location, utilization and supply.
What future projects should show
Track retail tariffs, fuel and capacity charges, delivered generation and bills at fixed household usage. Identify who pays incremental costs before claiming savings.
Limits & sources
E3’s Amazon-commissioned facility studies identify potential allocated-cost surpluses; the Dallas Fed studies wholesale market effects. Those different boundaries can coexist. Neither establishes typical national retail bill savings.
Every builder must bring and pay for all its own infrastructure.
Jurisdiction-specific
AEP Ohio · PUCO order July 2025
What the evidence says
Ohio ordered a dedicated data-center tariff intended to protect other customers from costs of underused infrastructure. Such an order establishes a specific utility arrangement; it is not evidence of a universal national self-supply rule.
What future projects should show
Link the applicable tariff and signed service agreement. Identify responsibility for substations, transmission, generation, minimum bills, security and early exit costs.
Limits & sources
A connection payment does not by itself settle every shared-system cost or prove additional generation is operating.
An announced campus guarantees utility investment will be recovered.
Not established
Large-load cost allocation · Ohio example
What the evidence says
PUCO’s customer-protection rationale explicitly addresses underused infrastructure. The need for protection persists when requested load does not materialize.
What future projects should show
Record binding contracted demand, commissioning phases, credit support and termination terms. Track decommissioning responsibility and land restoration arrangements separately.
Limits & sources
A public investment announcement is not a disclosed enforceable payment guarantee.
The buildout can finance more clean power.
Supported opportunity
Global and U.S. supply · IEA 2025 Base Case
What the evidence says
IEA projects renewables meeting nearly half of global data-center demand growth through 2030. Some new capacity is financed through technology-company power contracts. In the U.S. it projects 110 TWh of additional annual renewable supply to data centers between 2024 and 2030.
What future projects should show
Track named wind, solar, storage, geothermal and nuclear projects from contract through operation, with annual delivered MWh. Keep factory investment separate from power generation.
Limits & sources
A contract can support financing; it does not establish commissioning or prove all contracted generation is additional.
A renewable purchase means the campus runs on clean power every hour.
Needs qualification
Global physical supply · IEA 2025
What the evidence says
IEA separates physical electricity supply from contractual sourcing. Its Base Case still has gas and coal meeting over 40% of additional global data-center demand through 2030.
What future projects should show
Report hourly, local carbon-free matching alongside annual purchases, storage duration and firm supply. Count operating clean generation and avoided emissions using a stated counterfactual.
Limits & sources
Annual contractual matching cannot establish the physical mix at every hour or zero lifecycle emissions.
Better chips and cooling will make aggregate electricity demand fall.
Not guaranteed
Global · IEA 2035 scenarios
What the evidence says
IEA’s High Efficiency Case still projects about 1,100 TWh of electricity generation supplying data centers in 2035, below its Base Case but above 2024. This supply measure includes losses and is not the same as electricity consumed at the facility.
What future projects should show
Measure energy per completed task and aggregate delivered work together. Preserve the source’s boundary when comparing TWh, and test demand under multiple adoption paths.
Limits & sources
Efficiency reduces input per task; total demand also depends on how much work is performed.
The buildout is creating demand for skilled trades.
Supported, with scope
U.S. Indeed postings · July 2026 study
What the evidence says
Indeed reports data-center postings more than doubled over two years. Installation and maintenance account for about a quarter of openings; advertised installation wages carry a 42% premium, around $10 an hour.
What future projects should show
Follow electricians, HVAC, pipefitters, lineworkers and commissioning technicians through hires, paid hours, apprenticeships and retention. Report the share hired locally.
Limits & sources
Postings are not filled jobs. A cross-sectional advertised pay premium is not a 42% raise for the same worker or an electrician-only estimate.
Thousands of construction workers mean thousands of permanent campus jobs.
Different measures
Virginia industry interviews · JLARC 2024
What the evidence says
Interviewees described roughly 1,500 workers at peak construction, versus about 50 full-time workers, half contractors, for a typical 250,000-square-foot facility.
What future projects should show
Publish peak construction headcount, worker-years and permanent operating jobs as separate measures. Include construction duration and contractor scope; never add overlapping shifts or phases.
Limits & sources
These illustrative facility figures are not a staffing forecast for every AI campus.
Data centers can create substantial wider economic benefits.
Supported estimate
Virginia · JLARC 2024 economic modeling
What the evidence says
JLARC estimates annual statewide contributions of 74,000 jobs, $5.5 billion in labor income and $9.1 billion in GDP, with benefits concentrated in construction.
What future projects should show
Seek supplier contracts, actual local payments, wages and operating payroll. Trace power equipment, construction materials and services into the region’s businesses.
Limits & sources
Modeled statewide contribution is not observed onsite staffing, net national employment, or a multiplier transferable to another county. Labor income and GDP overlap.
Data-center taxes can materially fund local public services.
Supported local example
Loudoun County · FY2026 and FY2027
What the evidence says
Loudoun reports $1.2 billion in FY2026 real and personal property tax revenue from data centers, or 39% of its budget. Its FY2027 adopted budget forecasts $1.3 billion and 40%. These are county-reported figures, not an independent audit by Stack Ledger.
What future projects should show
Compare budgeted receipts with year-end collections, separated by tax type. Show schools, public safety and infrastructure spending without attributing every budget improvement to one industry.
Limits & sources
Local tax rules, exemptions, equipment depreciation and public costs determine how much value each community retains.
A lower property-tax rate means every homeowner’s bill fell.
Needs qualification
Loudoun County · tax years 2016–2026
What the evidence says
Loudoun reports its real-property rate fell from $1.145 to $0.805 per $100 of assessed value. It credits data-center revenue with supporting services and lower rates, while noting that individual bills vary with property values.
What future projects should show
Track rates and assessed values together. Compare household bills on a consistent property basis and account for exemptions.
Limits & sources
The rate comparison alone cannot establish how much an individual homeowner saved because of data centers.
Headline tax revenue proves every incentive package pays for itself.
Not established
Local fiscal assessment
What the evidence says
Loudoun cautions against dependence on a single fast-growing revenue source. Gross receipts alone do not establish net fiscal return.
What future projects should show
Publish receipts less abatements, grants and attributable public service and infrastructure costs, over a stated horizon. Include a lower-growth scenario and the timing of equipment depreciation.
Limits & sources
A reviewed project-level fiscal counterfactual is still needed before claiming incentives caused a net gain.
Noise and land-use concerns are simply misinformation.
Contradicted as a blanket claim
Virginia · JLARC 2024
What the evidence says
JLARC documents some residential impacts and challenges regulating continuous low-frequency noise.
What future projects should show
Show setbacks, nighttime measurements, tonal noise and enforcement remedies. Choose compatible industrial sites and publish the applicable zoning decision.
Limits & sources
A countywide economic benefit does not measure exposure at the nearest home; impacts vary by site design.
Backup generators are either harmless or the dominant regional polluter.
Needs qualification
Northern Virginia · JLARC 2024
What the evidence says
JLARC reports backup generators contribute under 4% of regional nitrogen oxides and at most 0.1% of carbon monoxide and particulate emissions; local impacts still require scrutiny.
What future projects should show
Separate emergency backup from routine primary generation. Track permits, operating hours, fuel, local exposure and cleaner backup alternatives.
Limits & sources
These regional shares cannot establish the effects of a future campus or engines operated continuously.
Mass unemployment caused by AI has already been demonstrated.
Not established
United States · August 2026
What the evidence says
BLS reports positive August 2026 payroll growth and 4.1% unemployment; these aggregates cannot establish causation. Yale finds no clear aggregate AI labor-market footprint in its August update. Regional New York Fed evidence shows both AI-related hiring and reduced hiring, with layoffs uncommon. Stanford finds a narrower entry-level employment warning. Together, these do not establish mass unemployment caused by AI or prove that AI has no employment costs.
What future projects should show
Track employment, hiring, separations, hours and wages by occupation and age. Compare exposed and less-exposed groups with transparent methods.
Limits & sources
National employment growth does not prove every worker benefits, rule out displacement, or guarantee future outcomes.
Healthy aggregate employment means young workers face no disruption.
Needs qualification
ADP sample · workers aged 22–25 · through June 2026
What the evidence says
Stanford reports employment in highly AI-exposed occupations about 19% below a counterfactual tracking less-exposed peers. The pattern is mainly weaker hiring, not elevated separations. The analysis is descriptive; controls and sample composition affect estimates.
What future projects should show
Follow entry-level hiring and apprenticeship pathways alongside new technical jobs. Evaluate whether training connects displaced workers to opportunities.
Limits & sources
This is not a 19% national unemployment rate, a count of workers fired by AI, or proof of causation.
A single headline establishes how many net jobs AI created.
Estimate not reproduced
BizNews / The Economist · September 2026
What the evidence says
The linked article attributes around one million new U.S. jobs to AI, compared with roughly 200,000 AI-attributed layoffs. Its synthesis is not an official causal net-employment series.
What future projects should show
Require reproducible component series, a common period, an explicit counterfactual and checks for overlapping roles. Distinguish company explanations for layoffs from independently measured causes.
Limits & sources
We have not reproduced the estimate. Gross additions and announced cuts from different sources cannot simply establish net employment.
AI makes every worker more productive immediately.
Task-dependent
Customer support field study and developer experiments
What the evidence says
A customer-support field study (November 2024 revision) reports 15% higher issues resolved per hour with AI assistance. METR’s early-2025 experiment found experienced open-source developers took 19% longer with AI. METR’s February 2026 update describes selection and measurement difficulties in estimating newer tools’ effects.
What future projects should show
Measure completed work, quality, review time and total cost for each deployment. Track productivity beyond coding in support, engineering, science and administration using task-appropriate evaluations.
Limits & sources
Different tasks, tools, workers and study designs cannot be pooled into a universal productivity multiplier or assumed job-loss rate.
How does data-center water compare with golf irrigation?
Useful context; different boundaries
United States · 2023 data centers / 2024 golf
What the evidence says
Berkeley Lab estimates 66 billion liters of direct data-center water consumption in 2023. GCSAA estimates 1.63 million acre-feet applied to golf facilities in 2024, about 2,011 billion liters after conversion. Data-center electricity-related water is excluded from the first figure; applied irrigation is not identical to consumption.
What future projects should show
Publish comparable consumption and withdrawal measures for the same region and period. Include reclaimed water, return flows and indirect electricity-related consumption.
Limits & sources
These totals include all data centers, not AI alone. They are context figures, not a matched impact ratio; neither settles the adequacy of a local water supply.
Can an almond provide an honest reference for AI water use?
Conditional calculation
California historical irrigation estimate / reported ChatGPT average
What the evidence says
The featured calculation uses only the almond report’s blue-water component and an explicitly assumed kernel mass. It excludes rainfall and the grey-water pollution-assimilation indicator. Altman’s reported query average lacks enough methodological detail to verify matching boundaries.
What future projects should show
Seek model-specific measured water consumption, task length, cooling location, indirect electricity scope and measurement dates. Recalculate when comparable data are disclosed.
Limits & sources
The ratio is arithmetic using a disclosed assumption and attributed inputs. It is not an independently verified environmental equivalence or a measurement of every GPT prompt. Food and computation provide different benefits.
Can a data center contribute more utility revenue than its allocated cost?
Supported modeling example
E3 study for Amazon · December 2025
What the evidence says
E3 evaluated Amazon facilities in four utility territories and found projected revenue covered estimated facility-level service costs. For an assumed 100 MW facility, it projected surplus utility revenue of $3.4 million in 2025 and $6.1 million in 2030. Such margins can reduce revenue required from other customers.
What future projects should show
Track actual utility receipts, cost allocation and how regulators apply any surplus. Update cost studies as regional demand and supply change.
Limits & sources
Amazon commissioned the study. Modeled allocated-cost coverage is not an observed household saving; macro-level generation and market-price effects remain a separate question.
Can co-location and flexible demand reduce grid costs?
Supported regulatory pathway
PJM · FERC December 18, 2025 action
What the evidence says
FERC directed PJM to develop co-location service rules. Rosner explains how limiting contracted grid withdrawals and accepting curtailment can avoid unnecessary upgrades while preserving cost responsibility. His concurrence describes intended benefits; it is not a measured savings series.
What future projects should show
For each project, verify the subsequently accepted tariff, effective date, contracted imports, curtailment obligations and generation delivery. Identify whether generation is new or redirected from existing grid service.
Limits & sources
The cited action does not establish that every proposed service is currently implemented or that every data center must self-supply. Review later orders before reporting current legal requirements.
Does the SCC overview demonstrate that Virginia households already save?
Background, not an outcome study
Virginia SCC institutional overview
What the evidence says
The SCC lists proceedings examining data-center cost responsibility and flexibility. That establishes regulatory attention; the overview is not a quantitative household-bill evaluation.
What future projects should show
Link specific utility orders, tariffs and cost-of-service studies. Follow approved rules through actual household bills at fixed electricity usage.
Limits & sources
A regulator’s remit or symposium topic does not independently demonstrate an economic outcome.
No matching claims. Try another topic or search term.
