🌿 Community Awareness Initiative  |  Your voice matters. Report. Engage. Protect.
Research & Fact-Check

AI Data Centers:
What's True, What's Not

A lot of claims are circulating about what AI data centers actually do to communities. Here's an honest, sourced look at the facts — including the things the industry doesn't advertise.

💧 Water Usage

Water is used primarily to cool servers. The scale varies enormously depending on the cooling technology used — but for large AI facilities, the numbers are significant.

6.1B
gallons used by Google's data centers in 2023[1]
~500K
gallons per day for a typical hyperscale facility[2]
~1 pint
of water evaporated per 10–50 ChatGPT conversations[3]
Myth
"Data centers don't use much water — they're just buildings full of computers."
False. Large AI data centers are among the most water-intensive industrial facilities being built today. Cooling towers evaporate enormous volumes of water to remove heat from servers.[1] Microsoft and Amazon have reported comparable year-over-year growth in water use tied to AI data center expansion.[4]
True
AI queries use significantly more water than standard web searches.
Researchers at UC Riverside estimated that a short ChatGPT conversation of 10–50 questions consumes roughly 500 ml — about a 16 oz bottle — of water once cooling and the regional power-plant water footprint are included.[3] The authors note this is a conservative estimate built from public data, and real-world use could be higher depending on location and season.
Nuanced
"Tech companies are going 'water neutral' — so it's not a problem."
Some companies, like Microsoft, have pledged water neutrality by 2030 — meaning they'll replenish more water than they consume. However, neutrality programs often involve purchasing water credits in different watersheds than where the data center actually operates. A community in drought-stressed central Virginia doesn't benefit from a water restoration project in the Pacific Northwest.
True
The water draw is concentrated in specific areas, straining local supplies.
Because data centers cluster in corridors (Ashburn, VA; Phoenix, AZ; Columbus, OH), the combined water demand in those regions is far above national averages. Phoenix data centers are expanding in one of the most water-stressed metro areas in the US, raising serious concerns from local water authorities.

⚡ Electricity Consumption

Power demand from data centers is rising faster than at any point in the internet era — driven by the computational demands of AI training and inference.

~4.4%
of total US electricity consumed by data centers in 2023[5]
7–12%
projected share of US electricity by 2028[5]
$427B
spent by the four largest hyperscalers on AI infrastructure in 2025 alone, nearly doubling in 2026[7]
Myth
"AI is going to collapse the power grid."
This is an overstatement of a real risk. The grid won't "collapse," but rapid data center growth is straining regional grids, delaying clean energy timelines, and leading utilities to reopen or extend the life of fossil fuel plants. In some markets (Virginia, Texas), utilities are scrambling to approve new gas capacity to meet data center demand — reversing clean energy progress.
True
Residential electricity bills are rising partly because of data center load growth.
In Virginia, Dominion Energy has requested significant rate increases tied in part to the infrastructure costs of serving its massive data center customer base. Ratepayers shoulder the cost of grid upgrades that primarily benefit commercial customers. This dynamic is playing out across multiple utility territories.
Nuanced
"Big Tech data centers run on 100% renewable energy."
This is technically true under certain accounting methods, but misleading in practice. Most companies purchase renewable energy certificates (RECs) that "match" their usage on paper but don't guarantee the electrons running their servers are actually clean. True 24/7 carbon-free energy — where clean power is available every hour of every day — is a much harder standard that only a handful of facilities currently meet.
True
Training a single large AI model can emit as much carbon as five cars over their lifetimes.
A 2019 study from UMass Amherst found that training one large Transformer model with neural architecture search produced roughly 626,000 lbs of CO₂ equivalent — about 5× the lifetime emissions of the average American car.[4] Newer, larger models are widely presumed to be more energy-intensive to train, though companies rarely disclose per-model figures.
Outdated
"A single ChatGPT query uses 10× more energy than a Google search."
This widely repeated figure traces to a 2023 back-of-envelope estimate (~3 watt-hours per query) that has not held up. More recent analysis of ChatGPT's current models puts typical query energy use around 0.3 watt-hours — in the same range as older per-search estimates from Google.[6] The comparison isn't necessarily wrong for every model or workload, but the flat "10×" figure gets cited far more confidently than the underlying data supports.

🔊 Noise & Environmental Disruption

Noise is often the first complaint from residents near new data centers — and it's frequently dismissed by developers. Here's what the evidence actually shows.

Myth
"Data centers are quiet — you won't even notice one in your neighborhood."
Consistently false. Industrial HVAC units, chillers, backup generators, and cooling towers run 24 hours a day, 7 days a week. Resident- and press-reported noise measurements near active data centers in Loudoun County, Virginia have registered in the 50–65 dB range at property lines — comparable to a busy office or a running dishwasher, but constant and unrelenting. WHO night-noise guidance recommends outdoor residential levels stay below 40 dB to protect sleep, so sustained levels in this range are well above the health-based target.[8]
True
Low-frequency "hum" from cooling systems can penetrate walls and is harder to block than standard noise.
Infrasound and low-frequency noise (below 200 Hz) from industrial cooling equipment travels farther and passes through standard building insulation more easily than higher frequencies. Residents report feeling a constant vibration or pressure sensation inside their homes even when standard dB measurements appear within zoning limits.
Nuanced
"Developers are required to do noise studies before building."
In many jurisdictions, environmental impact assessments including noise studies are required — but enforcement is inconsistent. Studies are often conducted before full operational load is reached, and modeled noise levels frequently underestimate real-world impacts. Some permits are granted without any public hearing, and residents may have limited recourse once construction begins.
True
Light pollution from data centers disrupts sleep and local ecosystems.
Large data centers are secured facilities with extensive exterior lighting that operates all night. In rural or suburban areas where facilities are increasingly being built (due to cheaper land and power), 24-hour floodlighting represents a significant change to the light environment — affecting migratory birds, nocturnal wildlife, and nearby residents' sleep quality.

🔄 Closed-Loop Cooling Systems

Closed-loop cooling is often presented as the clean solution to data center water consumption. The reality is more complicated.

What is a closed-loop cooling system?

Traditional data center cooling uses open-circuit cooling towers: water is pumped over hot surfaces, some evaporates (removing heat), and the remainder is recirculated. Because water evaporates, it must constantly be replenished from local supplies.

A closed-loop system keeps the coolant (usually water or a refrigerant) in a sealed circuit — it never evaporates or contacts the outside air directly. Heat is transferred to a secondary system, such as a dry cooler (large fans blowing outside air over tubes) or a heat exchanger, and rejected without water consumption.

Some data centers use a hybrid approach: closed-loop primary cooling supplemented by evaporative cooling only during peak summer heat.

✅ Why Closed-Loop Can Be Good

  • Dramatically reduces or eliminates water consumption — up to 95% less water than conventional cooling towers
  • No cooling tower blowdown (wastewater discharge)
  • No chemical water treatment required, reducing chemical runoff risk
  • Less noise from water tower fans in some configurations
  • Better for drought-prone regions where water scarcity is a real concern

⚠️ Why Closed-Loop Can Be Problematic

  • Dry coolers require significantly more electricity to achieve the same cooling effect, especially in hot climates
  • In extreme heat events, closed-loop systems may be unable to cool adequately — requiring backup evaporative cooling anyway
  • Higher upfront capital cost, which some operators use to justify choosing conventional cooling instead
  • Some companies claim "closed-loop" status but still use water in the secondary heat rejection stage
  • The higher electricity draw of closed-loop cooling can increase carbon emissions depending on the grid mix
Nuanced
"We're building a closed-loop facility — so there are no water impacts."
This framing is common in developer communications and is often misleading. A true zero-water closed-loop system (using dry coolers only) does dramatically reduce water impacts — but it draws more electricity, which has its own environmental cost. And in hot climates, many operators quietly add supplemental evaporative cooling during summer months, reintroducing water consumption that isn't always disclosed. Ask for full technical specs before accepting a "closed-loop" claim at face value.
Context
What questions to ask about any data center's cooling claims
1. Is the system truly closed on both the primary and secondary loops, or does the secondary loop use evaporative cooling?
2. What is the facility's water usage effectiveness (WUE) rating — ideally close to 0 for closed-loop systems?
3. Is supplemental evaporative cooling used during peak summer temperatures? If so, at what frequency and volume?
4. How does the increased electricity demand of dry cooling compare to the carbon footprint of conventional water-cooled systems on the local grid?

❓ Common Questions

Quick answers to questions we hear most often from community members.

Can I find out how much water a specific data center uses?

Sometimes. Large publicly traded companies (Google, Microsoft, Amazon, Meta) publish annual sustainability reports that include aggregate water and energy figures — but typically not broken out by individual facility. For local data, you can file a public records request with your state environmental agency or water utility, request permit applications from your county zoning board, or contact your state's Department of Environmental Quality. Water discharge permits (NPDES) are public record and list facility-specific usage.

Are data centers required to notify the community before being built?

Requirements vary significantly by state and county. In some jurisdictions, a data center qualifies as a "by-right" use in an industrial zone and requires no public hearing at all. In others, conditional use permits trigger public comment periods. Many communities have discovered data centers only after construction began. If you're concerned about a facility in your area, check your county's zoning board meeting minutes and permit applications — these are usually public record.

Don't data centers bring jobs and tax revenue? Isn't that good for communities?

Data centers do generate tax revenue and construction jobs — but the long-term employment picture is more complicated. Once built, a large data center may employ only 20–50 full-time workers. Many localities also grant significant property tax abatements to attract facilities, reducing the net tax benefit for years. The question communities should ask is: are the environmental and infrastructure costs fairly offset by the economic benefits, or are they being asked to subsidize profits that largely flow elsewhere?

What's the difference between a "data center" and an "AI data center"?

Traditional data centers store files and run standard software — they're relatively efficient. AI data centers are purpose-built to run machine learning workloads on specialized chips (like Nvidia GPUs or custom AI accelerators), which generate significantly more heat per square foot and consume far more electricity and water than conventional computing. A hyperscale AI training cluster can draw 100–500 megawatts — enough to power a city of 75,000–375,000 homes.

Is there any regulation coming that could address these concerns?

Federal action has been slow, but some states are moving ahead. California, Texas, Georgia, and Virginia have all seen proposed legislation requiring data centers to disclose energy and water use, meet efficiency standards, or undergo environmental review. The EU has already enacted mandatory reporting requirements under its Energy Efficiency Directive. Community pressure — including documented reports like the ones on this site — is a key driver of regulatory action.

📚 Sources

Every numbered citation above links back to one of these. Primary sources and major outlets only — no aggregator blogs.

  1. Google, 2024 Environmental Report (covering 2023 data) — 6.1 billion gallons of water used by Google data centers in 2023. Google Environmental Report
  2. Estimates for typical hyperscale facility daily water draw, as reported in EESI, "Data Centers and Water Consumption" and industry water-use surveys.
  3. Li, Yang, Islam & Ren, "Making AI Less 'Thirsty'" (UC Riverside), first posted 2023 and later published in Communications of the ACM; see also UC Riverside News coverage.
  4. Strubell, Ganesh & McCallum, "Energy and Policy Considerations for Deep Learning in NLP," ACL 2019. ACL Anthology
  5. International Energy Agency, Energy and AI (2025) — US data centers at ~4.4% of electricity demand in 2023, projected 6.7–12% by 2028. IEA: Energy demand from AI; see also Pew Research Center summary.
  6. Epoch AI, "How much energy does ChatGPT use?" — reassessing the widely cited "10× a Google search" estimate against current model efficiency. Epoch AI
  7. Hyperscaler 2025–2026 AI capital expenditure figures as reported by CNBC and Fortune.
  8. World Health Organization, Night Noise Guidelines for Europe — 40 dB Lnight target for residential areas. WHO Compendium, Ch. 11: Environmental Noise

Sources were checked at the time of writing (July 2026) — data center statistics change quickly, so figures are revisited periodically.

Seen Something in Your Community?

Self-reporting from residents is how we build a real picture of data center impacts. Your firsthand account matters.

Submit a Report →