A data center is easy to ignore when it is somewhere else.
From a distance, it is part of an abstract “cloud”: a place where photographs are backed up, businesses run software, hospitals store records, banks process transactions, streaming services deliver video, scientists run simulations, and artificial-intelligence systems are trained and operated. From nearby, it is an industrial building with substations, cooling equipment, backup generators, security fencing, construction traffic, water pipes, and transmission lines.
That difference in perspective explains much of the controversy surrounding new data-center construction.
The companies building these facilities describe essential digital infrastructure, enormous private investment, construction employment, and new tax revenue. Opponents describe industrialization, electrical demand, water consumption, noise, public subsidies, and damage to landscapes that may never receive an equivalent local benefit. Both descriptions can be accurate.
The debate becomes distorted when either side treats the issue as morally simple. A data center is not harmless merely because its final product is digital. It is also not uniquely destructive merely because it uses a great deal of electricity. The relevant questions are more concrete:
How large is the project?
Where will its electricity come from?
Who pays for the grid expansion?
What cooling system will it use?
How much water will it consume during drought or extreme heat?
How close will it be to homes, schools, natural areas, farms, and historic sites?
What will residents hear and see?
What tax concessions is the company receiving?
What economic benefits remain after construction ends?
What happens if the projected computing demand never materializes?
Those are questions about infrastructure, land use, engineering, and public policy—not about whether one approves of every social use of artificial intelligence.
1. Why the controversy has intensified
Data centers are not new. Banks, universities, governments, and large corporations have operated them for decades. What has changed is their scale, concentration, and rate of construction.
A traditional enterprise data center might occupy part of an office or a modest industrial building. A cloud campus can contain several warehouse-sized buildings. An AI-oriented campus may request hundreds of megawatts of electrical capacity, with individual proposals reaching toward a gigawatt.
The U.S. Department of Energy reported that utilities were receiving requests for hyperscale facilities in the 300-to-1,000-megawatt range , sometimes with requested delivery schedules of only one to three years.1
Nationally, U.S. data centers consumed an estimated 176 terawatt-hours of electricity in 2023 , or about 4.4 percent of total U.S. electricity consumption . Lawrence Berkeley National Laboratory projected a wide range of 325 to 580 terawatt-hours by 2028 , equivalent to approximately 6.7 to 12 percent of U.S. electricity use, depending on hardware shipments, utilization, and cooling practices.2
The uncertainty is itself part of the problem. Utilities normally plan power infrastructure over decades. AI demand is growing on a technology-industry timetable, with rapidly changing hardware, business models, and forecasts. A transmission line, substation, gas plant, or nuclear plant may operate for 40 to 80 years. An AI accelerator may be commercially outdated within several years.
Communities are therefore being asked to approve durable physical infrastructure around an unusually uncertain demand forecast.
2. A data center is an industrial use, even when its product is invisible
A distribution warehouse stores physical goods. A factory transforms raw materials. A power plant produces electricity. A data center processes information.
Because the output leaves through fiber-optic cable, the site may generate little truck traffic after construction and employ relatively few people compared with its size. That can make it appear cleaner and quieter than many other industrial uses.
But the building still has an industrial character:
Large windowless structures, often several stories tall.
Extensive grading and impervious surfaces.
Electrical substations and high-voltage transmission connections.
Cooling towers, dry coolers, chillers, pumps, and fans.
Dozens or hundreds of backup generators.
Fuel storage, battery rooms, and fire-protection systems.
Continuous operation, 24 hours per day.
Security fencing, controlled access, and limited public interaction.
Virginia's Joint Legislative Audit and Review Commission, after an unusually detailed study of the world's largest data-center market, concluded that the facilities have industrial characteristics and can affect nearby residential areas through noise, visual impacts, construction, and associated infrastructure.3
That does not mean a data center is equivalent to a refinery, steel mill, or chemical plant. Its routine emissions and truck traffic are generally lower. It means it should be treated as serious industrial development rather than as an oversized office building.
3. The central conflict: national benefits, local burdens
The services produced by a data center may be used across the country or around the world. The physical burdens are concentrated around one utility territory, water system, watershed, road network, and community.
This geographic mismatch creates a recurring political problem.
A resident living near a proposed campus may receive no special benefit from the project beyond ordinary access to online services already available. The company, its customers, and distant users receive the primary economic value. The host community receives some combination of tax revenue, construction work, permanent jobs, infrastructure investment, and indirect economic activity—but also accepts the building, power lines, noise risk, water demand, and opportunity cost of the land.
This is not unique to data centers. Communities hosting power plants, airports, rail yards, landfills, mines, factories, pipelines, and distribution centers face similar questions. The distinguishing feature of data centers is the extreme ratio between their physical resource requirements and their relatively small permanent workforce.
The fair question is not whether local residents personally use cloud computing or AI. Nearly everyone does, directly or indirectly. The fair question is whether the host community is being asked to absorb disproportionate costs for broadly distributed benefits.
4. Electricity: the largest and most difficult issue
Electricity is the strongest legitimate concern surrounding modern data-center construction.
The national number is large, but concentration matters more
Data centers' 176 terawatt-hours in 2023 represented a significant but not dominant share of U.S. electricity use. For context:
Residential air conditioning used about 254 terawatt-hours in 2020.4
Cooling in commercial buildings used about 170 terawatt-hours in 2018.4
Total U.S. utility-scale generation was about 4,178 terawatt-hours in 2023.5
These figures come from different survey years and should not be treated as an exact side-by-side ledger. They do, however, place the industry in scale. Data centers already consume electricity on the order of a major nationwide building end use. They are not consuming most U.S. electricity, but they are no longer a marginal load.
The more important problem is that they cluster.
A national share of 4.4 percent can translate into a very large fraction of demand within one utility territory. Northern Virginia demonstrates the point. Virginia's data centers were estimated to use about 5,050 megawatts in 2024, with most of that demand concentrated in three counties.3
New demand can require more than a wire to the property
A data-center proposal may require:
A new utility substation.
New high-voltage lines or expanded rights-of-way.
Reconductoring or rebuilding existing lines.
Additional regional transmission.
New generating capacity.
Grid-scale batteries or other firming resources.
Natural-gas pipeline expansion.
New transformers, breakers, and switchgear.
Reinforcement of local distribution networks.
The building may occupy one parcel, but its electrical shadow can extend across counties and states.
This is why “the company will pay for its own substation” is not a complete answer. The direct connection can be privately funded while broader generation and transmission costs are recovered through utility rates. Regulators must decide which investments are dedicated to one customer, which benefit the wider grid, and what happens if the customer uses less power than forecast.
The stranded-cost problem
Utilities recover large infrastructure costs over many years. If a developer requests power, causes new facilities to be built, and then delays, reduces, or abandons its project, some costs may remain.
Virginia's JLARC found that utilities can reduce this risk through minimum-payment requirements, construction contributions, collateral, phased capacity commitments, and special tariffs. It also concluded that the risk cannot be eliminated entirely, particularly for long-lived generation and transmission projects.3
This issue deserves more attention than the simple claim that data centers will either “raise everyone's rates” or “pay their own way.” Either outcome depends on rate design.
A defensible arrangement should require large-load customers to bear the costs and financial risks that are uniquely caused by their projects. Ordinary customers should not be forced to finance speculative infrastructure for companies capable of negotiating sophisticated power contracts.
At the same time, not every grid upgrade benefits only the data center. A new transmission project can improve reliability, relieve existing congestion, or support other development. Cost allocation should reflect actual beneficiaries rather than functioning as either a subsidy or a punitive surcharge.
Reliability is a real constraint
Data centers require highly reliable electricity, but they can also improve grid resilience under the right agreements.
Potentially useful capabilities include:
Reducing nonessential computing during grid emergencies.
Moving flexible jobs to another region.
Temporarily adjusting temperature or server power limits.
Using on-site batteries to reduce peaks.
Coordinating generator tests to avoid stressed periods.
Building dispatchable generation or storage that also supports the grid.
Not all computing is equally flexible. A long AI training job may be delayed or moved. A hospital system, financial transaction, emergency service, or live inference platform may require continuous operation.
Utilities should therefore assess actual load flexibility rather than accepting vague promises that “AI can run whenever renewable energy is available.”
Electricity source determines much of the environmental impact
The same data center can have very different carbon and water footprints depending on where it is connected.
A facility supplied mainly by coal and natural gas indirectly causes more greenhouse-gas emissions than one supplied by nuclear, hydroelectric, wind, solar, or geothermal power. The difference is not merely contractual. It depends on what generators actually respond to the load over time.
Corporate renewable-energy purchases can help finance cleaner generation, but annual accounting can obscure hourly reality. A company may buy enough renewable energy credits to match annual consumption while drawing fossil-heavy grid power during calm nights or winter peaks.
A stronger standard asks:
Is new clean generation being added?
Is it located on the relevant grid?
Does it produce power during the data center's operating hours?
Is firm capacity available when wind and solar output are low?
Does the project delay retirement of fossil generation?
Are storage and transmission included?
Who pays for reliability?
A data center connected to a cleaner grid is not impact-free, but the grid mix matters more than slogans about the cloud being either inherently clean or inherently dirty.
5. Water: a problem that is often measured badly
Water disputes are especially vulnerable to misleading numbers.
A report may refer to water that is:
Withdrawn from a river, lake, or aquifer.Consumed and not immediately returned, usually through evaporation.Used on-site by the data center.Used indirectly by power plants generating its electricity.Potable , reclaimed, saline, or otherwise nonpotable.Measured during an average year, a peak summer day, or a drought emergency.
Those are not interchangeable.
Direct water use
Lawrence Berkeley National Laboratory estimated that U.S. data centers directly consumed about 66 billion liters , or roughly 17.4 billion gallons , of water in 2023.2
In isolation, that sounds enormous. In comparison with national water withdrawals, it is small. The U.S. Geological Survey estimated total U.S. withdrawals at 322 billion gallons per day in 2015. Thermoelectric power accounted for 133 billion gallons per day, irrigation for 118 billion, and public supply for 39 billion.6
That comparison requires caution. Much thermoelectric water is withdrawn and returned, while evaporative data-center cooling consumes water. Withdrawal figures also include saline water and sources that may have no relationship to the municipal system serving a proposed data center.
The purpose of the comparison is not to dismiss data-center water use. It is to show that data centers are not the largest national water-using industry. Agriculture and power generation dominate national totals.
Local water use can still be decisive
National percentages are nearly useless when a facility is proposed in a water-stressed basin or a small utility district.
Virginia's study found that data centers used less than 0.5 percent of statewide water withdrawals in 2023. Yet they represented between 2 and 21 percent of total use at six water utilities examined by the state. Most individual buildings used no more than an average large office building, but a small number used far more; one building consumed approximately 243 million gallons in a year .3
Both statements are true:
The industry was a small share of statewide water use.
It could be a major customer for a particular local utility.
This is why “data centers use less water than agriculture” is not a sufficient defense. A proposed facility is not competing with all U.S. agriculture. It is competing with residents, businesses, ecosystems, and future development connected to the same water sources.
Indirect water use may exceed on-site use
Many power plants consume water. Berkeley Lab estimated that the electricity serving U.S. data centers in 2023 carried an indirect water footprint of nearly 800 billion liters , much larger than direct site consumption.2
The indirect figure varies sharply by grid. Some thermal plants use significant cooling water; wind and solar photovoltaics use little water in operation. Hydroelectric water accounting is methodologically complicated because reservoir evaporation can be allocated in different ways.
A project advertised as using “almost no water” on-site may therefore still have a water footprint through its electricity. Conversely, a facility using evaporative cooling may reduce its electrical demand by avoiding compressor-heavy cooling. There is often a water-energy tradeoff rather than one universally superior design.
Cooling choices matter
Common approaches include:
Evaporative cooling , which can be energy-efficient but consumes water.Chilled-water systems , which may use cooling towers and compressors.Dry cooling , which uses little site water but may require more fan energy and perform less efficiently during hot weather.Direct-to-chip liquid cooling , which circulates coolant in a closed loop but still needs a method to reject heat outdoors.Hybrid systems , which use water only under hot or high-load conditions.Reclaimed-water systems , which reduce demand for potable water but require pipelines, treatment capacity, and careful water chemistry.
“Liquid cooled” does not necessarily mean “water consuming.” A closed loop can circulate the same fluid repeatedly. The important question is how the heat is ultimately rejected and how much makeup water is required.
A responsible application should disclose annual consumption, peak-day demand, drought-stage demand, source water, wastewater discharge, cooling technology, and expected changes when the facility is fully built out.
Water should be judged against local scarcity
One million gallons in a water-rich region with reclaimed-water infrastructure is not equivalent to one million gallons from a stressed aquifer.
Good policy therefore evaluates water by:
Source.
Consumptive fraction.
Seasonal timing.
Drought conditions.
Ecological flow requirements.
Utility treatment and delivery capacity.
Competing future needs.
Availability of reclaimed or nonpotable supplies.
A universal gallon limit is less useful than a transparent basin-level water budget.
6. Land, forests, farms, wetlands, and historic places
Data centers consume less land per unit of economic value than many low-density developments, but modern campuses are still large.
Virginia's approximately 150 data-center sites occupied about 7,200 acres and contained more than 63 million square feet of building space at the time of its 2024 study.3
A site may require:
Cleared and graded building pads.
Roads and loading areas.
Stormwater basins.
Fuel and generator yards.
Cooling equipment.
Electrical substations.
Setbacks and security zones.
Transmission-line corridors.
Water and sewer extensions.
The direct land impacts are not unique
Clearing a forest for a data center causes many of the same direct impacts as clearing it for a warehouse, shopping center, factory, or subdivision:
Habitat loss and fragmentation.
Soil compaction.
Increased runoff.
Stream sedimentation during construction.
Loss of agricultural land.
Wetland and stream disturbance.
Altered viewsheds.
Destruction of archaeological resources.
Reduced carbon storage in vegetation and soil.
Virginia's preservation review found that data centers' risks to historic resources were broadly similar to other large developments, though extensive grading and tall industrial buildings can make avoidance and viewshed protection difficult.3
The proper conclusion is not that data centers are harmless because warehouses also damage land. It is that land-use rules should apply consistently to large industrial developments.
The indirect footprint may be larger than the campus
A data center's transmission corridor can fragment forests, cross streams, affect private land, and alter historic or scenic views far beyond the property line. New generation has its own footprint.
For example:
Solar power requires substantial land unless built on roofs, parking areas, disturbed land, or in combination with agriculture.
Wind farms require broad spacing, though much of the land between turbines can remain in other uses.
Natural-gas generation requires pipelines and produces air emissions.
Hydroelectric projects alter rivers and reservoirs.
Nuclear plants use compact sites but require long development timelines, cooling, fuel-cycle infrastructure, and waste management.
Transmission is needed for nearly every remote energy source.
The physical cost of computing cannot be evaluated solely by looking at the walls of the data center.
Does pavement cause drought?
Impervious surfaces are a legitimate local concern, but they are sometimes assigned effects far beyond what they plausibly cause.
Roofs and pavement reduce infiltration, increase the speed and volume of stormwater runoff, contribute to stream erosion, carry pollutants, and can reduce local groundwater recharge. The U.S. Geological Survey describes these effects as standard consequences of watershed urbanization.7
A large campus can therefore worsen local drainage and water-quality problems if stormwater is poorly managed.
What it does not ordinarily do is create a regional meteorological drought. Drought is primarily driven by precipitation deficits, heat, evaporation, snowpack, soil moisture, and large-scale climate patterns. Paving one industrial site may affect a local watershed; it does not meaningfully redirect regional weather systems or explain widespread wildfire conditions.
The reasonable response is rigorous stormwater design:
Preserve natural drainage where possible.
Limit unnecessary parking and pavement.
Use infiltration basins, bioswales, wetlands, and permeable surfaces where soils permit.
Protect stream buffers.
Control construction sediment.
Monitor discharge temperature and quality.
Account for larger future storms.
Rejecting an exaggerated claim does not erase the real local hydrologic impact.
7. Noise: not always loud, but sometimes relentless
Data centers are often described as quiet because they have few vehicles and no assembly line. That description can be true at a well-sited and well-designed facility. It can also be badly misleading near certain cooling plants.
Potential noise sources include:
Cooling-tower and dry-cooler fans.
Chillers, pumps, compressors, and transformers.
Air-handling equipment.
Backup-generator testing.
Electrical hum.
Construction equipment and blasting.
Trucks delivering fuel or equipment.
The most difficult complaints often involve continuous low-frequency sound. It may not be loud enough to damage hearing or violate an ordinance written for parties, vehicles, or barking dogs. Yet a steady hum can be noticeable indoors and outdoors, particularly at night when background sound is low.
Virginia investigators reviewed measurements at complaint locations in the range of approximately 40 to 59 A-weighted decibels . They found that most data centers did not generate complaints, but some nearby residents reported sleep disruption, headaches, loss of outdoor enjoyment, and reduced quality of life.3
This is an area where both extremes fail.
It is inaccurate to imply that every data center produces intolerable noise. Location, terrain, equipment, enclosures, fan speed, barriers, and building orientation make a large difference.
It is equally inaccurate to tell residents that a noise cannot be disruptive merely because it falls below a conventional A-weighted limit. A-weighting deemphasizes low frequencies. C-weighted measurements, octave-band analysis, tonal penalties, nighttime limits, and limits on cumulative sound from multiple facilities may better capture the problem.
Noise is much easier to prevent than correct. Once a campus is operating, retrofitting barriers, silencers, fan controls, or enclosures can be difficult and expensive.
Local approval should therefore require:
Baseline sound measurements.
Worst-case modeling with all cooling equipment operating.
Separate daytime and nighttime standards.
Low-frequency and tonal analysis.
Modeling of future campus phases.
Independent review.
Post-construction verification.
Enforceable corrective-action requirements.
Setbacks are useful, but distance alone is not a substitute for acoustic design.
8. Air pollution and backup generators
Most data centers rely on diesel generators for emergency power. A large campus may have scores or hundreds of them because backup capacity is distributed among electrical blocks.
Diesel engines emit nitrogen oxides, carbon monoxide, particulate matter, and greenhouse gases. Their presence is a legitimate air-quality concern, particularly in regions already struggling with ozone or particulate pollution.
Routine operation is usually limited. Generators are tested periodically and run during power failures. Virginia found that actual data-center generator emissions in 2023 were only about 7 percent of permitted levels , with most emissions caused by maintenance testing. In Northern Virginia, data centers accounted for less than 4 percent of regional nitrogen-oxide emissions and 0.1 percent or less of carbon-monoxide and particulate emissions.3
That evidence tempers claims that backup generators normally operate like a continuous fossil-fuel power plant. They generally do not.
Several concerns remain:
Local concentration: Regional averages may not detect effects near a campus containing a dense cluster of engines.Testing: Monthly and annual tests produce recurring noise and emissions even without a grid failure.Long outages: A regional emergency could cause many generators to run simultaneously.Technology: Older Tier 2 engines emit more pollution than Tier 4 engines or units equipped with selective catalytic reduction.Demand response: Allowing emergency engines to run for economic grid programs can increase operating hours unless permits and equipment are upgraded.
New projects in populated or polluted regions should use the lowest-emission practical backup technology, coordinate testing, publish schedules, monitor cumulative impacts, and evaluate alternatives such as batteries, fuel cells, cleaner engines, or shared resilient generation.
A battery is not automatically a complete substitute. Multi-day backup would require enormous storage capacity, and batteries themselves involve mining, manufacturing, fire protection, and eventual replacement. The practical near-term solution may be a layered system rather than the elimination of all engines.
9. The visible plume is usually not smoke
Cooling towers can release a visible white plume, particularly in cool or humid weather. The plume is generally condensed water droplets—essentially fog—not smoke from combustion.8
That distinction matters. A visible plume is not evidence that the data center is continuously burning fuel.
It can still have effects:
Fogging or icing under certain weather conditions.
Mineral drift if water treatment and separators are inadequate.
Visual impacts.
Water consumption through evaporation.
Potential microbial risks if cooling-water systems are poorly maintained.
The correct response is neither panic nor dismissal. The plume should be identified accurately, and the cooling system should be regulated for water quality, drift, maintenance, and siting.
10. Construction: intensive, disruptive, and temporary
Data-center construction is economically significant precisely because the buildings are expensive and mechanically complex.
Construction can involve:
Mass grading and rock blasting.
Heavy truck traffic.
Concrete pours and structural steel.
Installation of substations and underground utilities.
Continuous or extended work schedules.
Large temporary workforces.
Dust, mud, lighting, vibration, and road wear.
Multiple phases extending across years.
Virginia's review estimated that a typical 250,000-square-foot data center might have approximately 50 full-time operational workers , while as many as 1,500 construction workers could be present at the peak of construction. Statewide, most employment and economic effects attributed to the industry came from capital investment and construction rather than ongoing operation.3
That does not make the construction employment unreal. Electricians, pipefitters, equipment operators, engineers, technicians, concrete workers, and other trades receive real work and wages.
It does mean officials should distinguish among:
Temporary construction jobs.
Permanent employees of the data-center operator.
Permanent on-site contractors.
Indirect jobs at utilities and suppliers.
Induced jobs supported by worker spending.
Jobs that would have existed elsewhere without the subsidy.
Announcements often combine these categories into one impressive number.
A community evaluating a project should ask for annual employment by phase, expected wages, residency assumptions, apprenticeship commitments, and a clear definition of “job.”
11. Taxes, incentives, and the low-employment paradox
Data centers can generate substantial tax revenue even with relatively few employees because the buildings and equipment are extraordinarily valuable.
Local revenue may come from:
Real-estate taxes.
Business personal-property taxes on computers and machinery.
Utility taxes.
Construction-related sales and use taxes.
Permit and connection fees.
Payments negotiated through development agreements.
This can be attractive to local governments. Data centers generally add fewer students, commuters, and daily service demands than a large residential development or labor-intensive office district. A valuable industrial property can therefore expand the tax base without proportionately increasing many local operating costs.
Virginia localities reported using data-center revenue to lower real-estate tax rates, build schools, create reserves, and support affordable housing.3
The controversy arises because states and localities also compete through tax exemptions and reduced rates.
Virginia's data-center sales-and-use tax exemption provided an estimated $928 million in savings in fiscal year 2023 .3
A subsidy can be defensible when it changes a company's location decision and produces net benefits. It is wasteful when every competing state offers similar concessions, the company would have built there anyway, or public costs exceed the additional economic activity.
The low-employment paradox sharpens this question. Governments often subsidize industry to create jobs, but the principal local advantage of a data center may be tax base rather than employment.
That bargain should be stated honestly. A project should not be marketed as both an enormous automated facility and a major source of permanent mass employment.
Tax revenue is not automatically community consent
A project can be fiscally valuable and still be poorly located.
Tax revenue does not cancel:
Noise at nearby homes.
Loss of a historic landscape.
Water stress.
Ratepayer exposure.
Destruction of high-quality habitat.
Industrialization of a rural view.
Construction disruption.
Opportunity cost of scarce power capacity.
Neither should those impacts automatically cancel genuine fiscal benefits. Local government exists partly to weigh competing goods.
The process becomes illegitimate when costs are minimized, benefits are inflated, or key details are withheld until after zoning approval.
12. Transparency is unusually important
Data-center companies often protect operational details for security and commercial reasons. Some confidentiality is reasonable. Public disclosure of network architecture, customer identities, or detailed building vulnerabilities would be inappropriate.
However, “proprietary” should not become a blanket answer to infrastructure questions.
Communities need enough information to evaluate:
Full campus buildout.
Maximum and expected electrical demand.
Ramp-up schedule.
Cooling technology.
Annual and peak water use.
Source and quality of water.
Generator count, type, fuel, and testing schedule.
Expected emissions.
Sound modeling.
Building height and visual simulations.
Tax incentives.
Permanent and construction employment.
Required utility upgrades.
Decommissioning and restoration plans.
A common source of distrust is phased approval . A developer presents one building, but the power, land purchase, or master plan anticipates a much larger campus. Each phase may appear manageable while the cumulative project creates the true impact.
Environmental and infrastructure review should therefore evaluate reasonably foreseeable full buildout, not only the first shell.
13. AI makes forecasting harder, not every concern larger
Artificial intelligence is driving much of the current construction boom, but AI should not be used as a vague multiplier for every fear.
AI facilities generally require denser computing, more electricity, high-speed networks, and increasingly liquid cooling. That can intensify power and cooling requirements.
Other impacts do not rise automatically in the same proportion:
A denser building may use more power without occupying proportionately more land.
Closed-loop liquid cooling may reduce or avoid evaporative water use.
A high-density campus may have fewer buildings than an equivalent amount of older computing.
Permanent employment may not rise much with electrical load.
Routine truck traffic remains limited after construction.
Noise depends more on the heat-rejection design than on whether the chips run AI.
The term “AI data center” is therefore not a complete environmental description.
The public needs engineering details, not labels.
14. Common claims that deserve correction
A balanced debate requires rejecting weak arguments regardless of which side makes them.
“Data centers are just warehouses”
No. They may resemble warehouses externally, but their electrical, cooling, generator, and network infrastructure is much more intensive. Their grid and utility impacts can be far larger.
“Data centers use most of the country's electricity”
No. They used an estimated 4.4 percent of U.S. electricity in 2023. That is substantial and rapidly growing, but it is not a majority.2
“Their electricity use is insignificant compared with everything else”
Also no. Data centers already consume electricity on the scale of major national building uses, and projected growth is large enough to shape utility planning. Local concentration can matter more than the national percentage.
“Every data center drains an aquifer”
No. Water use varies by cooling design, climate, source, and operating strategy. Some facilities consume large quantities; others use little site water. Many are connected to surface-water utilities rather than pumping groundwater directly.
“Water use is trivial because agriculture uses more”
No. Agriculture dominates national freshwater use, but that does not resolve whether one local water system can support a large new industrial customer during drought.
“The white plume is toxic smoke”
Usually no. A cooling-tower plume is generally condensed water droplets. Backup-generator exhaust is a separate emissions source.
“Pavement from data centers causes regional drought and wildfire”
Not in the ordinary causal sense. Impervious surfaces worsen local runoff, infiltration, water quality, and heat-island effects. They do not explain broad meteorological drought by themselves.
“Data centers create no useful product”
False. They support communications, commerce, government, medical systems, scientific research, entertainment, storage, cloud software, and AI. Whether every workload is socially valuable is a separate question from whether the infrastructure as a whole has value.
“Because the services are useful, every proposed facility should be approved”
False. Necessary infrastructure can still be oversized, subsidized, badly located, poorly designed, or unfairly financed.
“Efficiency will solve the growth problem”
Not by itself. Servers, power supplies, and cooling systems have become much more efficient, but demand can grow faster than efficiency improves. Lower computing costs can also encourage more use.
“A renewable-energy contract makes the project impact-free”
No. New generation, storage, transmission, land, and backup capacity still have physical impacts. Annual renewable accounting may not match hourly consumption.
“Opposition is simply fear of technology”
Sometimes opposition includes exaggeration, but many conflicts are ordinary land-use disputes: proximity to homes, ratepayer risk, water allocation, industrial noise, public subsidies, and loss of valued landscapes.
15. What responsible approval should require
A serious data-center policy need not ban the industry or accept every proposal. It should make approval conditional on measurable performance.
1. Appropriate siting
Priority should generally go to:
Existing industrial districts.
Brownfields and previously disturbed land.
Former power-generation or heavy-industrial sites.
Locations near existing transmission capacity.
Sites with adequate setbacks from homes.
Areas with sufficient water or viable low-water cooling.
Locations where transmission and utility corridors minimize new fragmentation.
High-quality habitat, prime farmland, historic landscapes, and residential edges should face a higher burden of proof.
2. Full-buildout review
Applications should disclose the reasonably foreseeable campus, not only the first building. Power, water, traffic, sound, stormwater, tax, and visual studies should model cumulative development.
3. Large-load cost protection
Utilities and regulators should require:
Upfront contributions for dedicated facilities.
Minimum demand charges.
Financial security and credit requirements.
Phased capacity reservations.
Exit fees or stranded-cost protection.
Transparent allocation of regional upgrades.
Special tariffs reflecting load size and reliability requirements.
The objective is not to overcharge data centers. It is to prevent speculative private demand from becoming a public liability.
4. Verifiable energy and carbon plans
A credible plan should identify:
Expected annual and peak load.
Actual supply mix.
New generation associated with the project.
Hourly matching goals where practical.
Firm capacity and storage.
Demand-response commitments.
Backup strategy.
Reporting methods.
Promises should survive changes in corporate branding and renewable-accounting fashion.
5. Water budgets and drought plans
Projects should disclose normal, peak, and emergency water use. Approval should consider local basin conditions and competing needs. Reclaimed water, hybrid cooling, closed loops, and higher-temperature liquid systems should be evaluated rather than assumed.
Water commitments should include drought curtailment rules and public reporting.
6. Acoustic standards designed for continuous industrial sound
Modeling should include low-frequency and tonal components, nighttime conditions, all future equipment, and post-construction enforcement.
7. Cleaner backup power
New projects should use lower-emission engines and controls where feasible, maximize battery bridging, stagger testing, avoid unnecessary generator operation, and disclose test schedules. Dense generator campuses warrant local monitoring rather than reliance only on regional averages.
8. Stormwater and ecological protection
Projects should minimize impervious area, preserve buffers, control construction sediment, replace lost trees meaningfully, protect wetlands, and monitor downstream effects. Landscaping should function ecologically and acoustically rather than merely hide the building in renderings.
9. Honest economic accounting
Public presentations should separate:
Construction jobs from permanent jobs.
Direct jobs from indirect estimates.
Gross tax revenue from tax concessions.
Dedicated infrastructure costs from public costs.
New economic activity from activity displaced elsewhere.
10. Enforceable community benefits
Where a project imposes substantial local costs, agreements may fund road improvements, utility upgrades, conservation, workforce training, noise mitigation, emergency services, or other locally relevant needs.
These agreements should not become a method of purchasing permission for an unacceptable site. Mitigation is not the same as suitability.
11. Decommissioning and adaptive reuse
Computer equipment is replaced frequently, but the building and electrical infrastructure can last for decades. Approvals should address:
Equipment recycling.
Hazardous materials.
Battery and fuel removal.
Site restoration.
Financial assurance.
Reuse of substations and utility connections.
Conversion to other industrial purposes if the computing market changes.
A community should not inherit a specialized empty shell and contaminated equipment yard after the tax incentives expire.
16. Context is not absolution
Comparisons with other industries are necessary because raw numbers are psychologically powerful.
Saying that a data center uses millions of gallons of water or hundreds of megawatts of electricity tells the reader that the facility is large. It does not establish whether the use is reasonable, unprecedented, or locally sustainable.
Context shows that:
Agriculture and thermoelectric generation dominate national water withdrawals.
Buildings, industry, and transportation consume much more total energy than data centers.
Warehouses, subdivisions, roads, factories, and shopping centers also convert land and increase runoff.
Many industrial facilities emit more routine air pollution.
Data centers can produce unusually high local tax value with comparatively little traffic and fewer permanent workers.
None of that absolves a particular project.
A modest national share can create a severe local problem. A cleaner use than a refinery can still be the wrong neighbor for a subdivision. A valuable tax base can still depend on an excessive subsidy. An industry serving real needs can still externalize costs.
The purpose of comparison is to improve judgment, not to make impacts disappear.
Conclusion: the cloud has an address
The core lesson of the data-center controversy is that digital services are physical.
Every stored photograph, cloud database, streamed video, scientific model, search result, and AI response depends on land, metals, concrete, electricity, cooling, networks, and human labor. The facilities concentrating those resources are becoming larger and more visible at the same time that their services become more deeply embedded in society.
That creates a genuine conflict. Communities cannot reasonably demand unlimited digital services while pretending the infrastructure should exist nowhere. Neither should technology companies assume that global demand entitles them to any parcel, any quantity of power, or any public subsidy.
The proper standard is not whether data centers are good or bad in the abstract. It is whether a proposed facility is built in a suitable place, supplied responsibly, financed fairly, designed to control its effects, and transparent about its full scale.
Over the long term, computing infrastructure may move toward more remote, unconventional, or even off-world locations. That possibility deserves separate treatment. For the foreseeable future, however, most computing will remain on Earth, connected to terrestrial grids and communities. The immediate task is therefore less dramatic but more important: deciding where these buildings belong, what they should be required to do, and which costs the public should refuse to inherit.
Sources and further reading
1 U.S. Department of Energy Secretary of Energy Advisory Board, Recommendations on Powering Artificial Intelligence and Data Center Infrastructure , July 2024.
2 Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report , December 2024. The report estimates 176 TWh of U.S. data-center electricity use, 66 billion liters of direct water consumption, and nearly 800 billion liters of indirect water consumption in 2023.
3 Virginia Joint Legislative Audit and Review Commission, Data Centers in Virginia , December 2024. This report is used repeatedly because it combines utility modeling, tax and employment data, local water records, emissions records, site visits, resident interviews, and land-use analysis in the world's largest concentrated data-center market. Its findings should not be assumed to describe every state.
4 U.S. Energy Information Administration, How much electricity is used for air conditioning in the United States? . EIA estimates 254 billion kWh for residential air conditioning in 2020 and 170 billion kWh for commercial-building cooling in 2018.
5 U.S. Energy Information Administration, What is U.S. electricity generation by energy source? . Utility-scale generation was approximately 4,178 billion kWh in 2023, with additional generation from small-scale solar.
6 U.S. Geological Survey, Estimated Use of Water in the United States in 2015 . USGS distinguishes withdrawals from consumptive use, an essential distinction when comparing industries.
7 U.S. Geological Survey, Impervious Surfaces and Flooding .
8 U.S. Environmental Protection Agency, Cooling Tower Plume Model . EPA describes visible cooling-tower plumes as fog formed when water vapor condenses into small droplets.
Additional references
U.S. Department of Energy, DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers .
U.S. Energy Information Administration, Annual Energy Outlook 2026 .
U.S. Energy Information Administration, Electricity use for commercial computing could surpass other building end uses .
Lawrence Berkeley National Laboratory, Water Efficiency in Data Centers .
Fairfax County, Virginia, Data Centers Staff Report .
ASHRAE, NEMA, and Pacific Northwest National Laboratory, AI Data Center Energy Performance Framework .
U.S. Department of Energy and National Renewable Energy Laboratory, Best Practices Guide for Energy-Efficient Data Center Design .