There’s an uncomfortable political-economy implication here: AI infrastructure expands fastest where permitting and incentives are friendliest,...

Source: LinkedIn

Data centers and water: what we can (and can’t) responsibly say

by ChatGPT-5.2

Karen Hao’s LinkedIn post is essentially a pointer to her public correction note about her book Empire of AI and its discussion of data-center water impacts.

The correction itself is instructive—not just about one Chilean case, but about how easy it is to get water claims wrong when the underlying reporting mixes units, definitions, and accounting boundaries.

1) The “unit error” is a cautionary tale about how water narratives get distorted

Hao explains that a government document for Cerrillos (Chile) labeled figures as liters but they were actually cubic meters—an error that understated residential water use by a factor of 1,000 and, downstream, inflated the comparison between a proposed Google data center and the local community. She corrected the line to: the data center could use slightly more water than the entire population of Cerrillos over a year—based on Google’s permitted maximum of 169 liters/second, which is ~5.33 billion liters/year, versus ~5.10 billion liters/year for the municipality (2019).

Two takeaways matter beyond the Chile example:

• Water claims are brittle: one labeling mistake turns into a viral “data centers steal X thousand people’s water” soundbite.

• Permitted maximum vs expected use: EIA/permit documents often state maxima. Using them can be defensible (as Hao argues), but it should be expressed as “could use up to” and not as actual consumption.

2) The big conceptual trap: “withdrawal” vs “consumption” (and why both matter)

Hao’s second change is even more important: she says she incorrectly used “consume” where the correct term was “use” or “withdrawal,” and she clarifies that the water footprint is driven both by on-site cooling and by electricity generation (which may require water depending on the grid mix).

This is the core reason reasonable people talk past each other:

• Withdrawal = water taken from a source (some may be returned).

• Consumption = water not returned to the same watershed (often evaporated or incorporated elsewhere).

A hyperscale operator can market “water-free cooling” while the system still has large indirect water impacts via the power it consumes (especially on grids with thermoelectric generation). That exact dynamic is now showing up in reporting and critique: “water-free” can be true for direct use, while indirect water remains large.

3) Water use varies wildly by cooling design, climate, and grid—so averages mislead

A rigorous 2021 review in npj Clean Water finds data-center water efficiency spans an enormous range, depending on system characteristics and what’s being counted (withdrawal vs consumption): consumption from ~0 to 4.4 L/kWh, and withdrawal from ~0.31 to 533.7 L/kWh.

This is why “one number” is nearly always propaganda—whether it’s minimising or catastrophising.

A more grounded way to speak:

• Site specifics dominate: evaporative cooling in a hot/dry place can drive high local consumption; closed-loop/chiller approaches shift burdens toward electricity (and potentially upstream water).

• Water stress matters more than volume: withdrawing “moderate” amounts in a basin under drought conditions can be more socially destabilising than higher volumes in water-abundant regions.

• Temporal spikes matter: heat waves drive peak cooling demand exactly when communities face restrictions.

4) AI changes the water conversation because it changes load shape and location choices

Hao cites an estimate (University of California, Riverside researchers) that surging AI demand could drive ~1.1–1.7 trillion gallons/year of predominantly fresh water use/withdrawal globally by 2027, and she compares that scale to about half the UK’s annual water withdrawals.

That estimate originates from the “Making AI Less ‘Thirsty’” line of work (Li, Yang, Islam, Ren), which explicitly argues the water footprint is both direct (cooling) and indirect (power generation), and that it has been under-measured and under-reported.

There’s an uncomfortable political-economy implication here: AI infrastructure expands fastest where permitting and incentives are friendliest, not necessarily where water and grid constraints are lowest. That pushes real, localised conflict—even if global averages look “manageable.”

What else counts as pollution and environmental burden from data centers?

Below is a practical inventory of burdens that recur across credible reporting, permitting debates, and environmental assessments—especially once you look beyond operator sustainability PR.

Air pollution (local public-health burden)

• Diesel backup generators: NOx, PM2.5, CO, VOCs—especially during testing, outages, and grid-stress events. Regulators explicitly permit and track diesel emissions from data centers (e.g., Washington State).

• Investigations and public-health analyses increasingly connect AI data centers to asthma, cardiovascular harms, and other externalities, via both generators and power plants serving the new load.

• A live example: a proposed UK hyperscale site associated with nearly 600 diesel generators has raised concerns about air pollution near homes and a school.

Greenhouse gases (direct + indirect)

• Scope 2 emissions from electricity use remain the dominant driver in many regions; “100% renewable” claims often rely on contracting/offset structures that may not match local grid reality hour-by-hour.

• Scope 1 refrigerant leakage can be a high-GWP climate risk where HFCs are used in cooling systems.

PFAS and other persistent chemical burdens

• Fire suppression foams (AFFF) historically contain PFAS; releases can contaminate soil/groundwater and are extremely difficult to remediate.

• PFAS/F-gas related concerns in data-center operations are now being raised explicitly by environmental advocates and media, including claims of on-site leakage pathways and end-of-life disposal problems for fluorinated materials.
  Note: “PFAS” is sometimes used loosely in public debate; refrigerant HFCs are fluorinated gases but not always PFAS in the strict chemical sense. The governance point stands either way: fluorinated chemistries can create long-tail liabilities through leaks and disposal.

Water quality impacts (not just “how much”)

• Thermal pollution: water returned warmer can disrupt ecosystems (especially in smaller receiving waters).

• Chemical treatment and biocides in cooling systems can create wastewater management burdens.

• Blowdown discharge from cooling towers can concentrate minerals and treatment chemicals, requiring controlled disposal.

Noise pollution (quality-of-life + health)

• Continuous sound sources: cooling fans, chillers, pumps; episodic peaks: generator tests.

• Noise has become a recurring vector of community backlash and permitting friction.

Land use, impermeable surfaces, and habitat fragmentation

• Hyperscale builds can cover hundreds of acres with concrete/steel/paved surfaces, displacing farmland/nature/housing and requiring new transmission corridors.

• Local effects include stormwater runoff changes, heat retention, and ecosystem disruption.

“Waste heat” and microclimate effects

• Data centers are effectively large heat engines; heat discharge can worsen local heat burdens.

• Heat-recovery exists but is geography-constrained and often oversold.

Materials, mining, and supply-chain extraction

• Large volumes of copper, steel, concrete, plus AI-adjacent demand for lithium and other critical minerals for backup power and grid infrastructure. Hao highlights these “extraordinary volumes” explicitly in the corrected paragraph.

• Upstream burdens: mining water use, tailings risk, and energy-intensive refining.

E-waste and toxic components

• Short refresh cycles (accelerated by AI hardware churn) increase volumes of end-of-life servers, batteries, and power electronics.

• Hazard pathways: heavy metals, brominated flame retardants, and improper recycling/export.

Grid stress and induced fossil fuel burn

• Even when a data center’s direct operations are “efficient,” rapid new load can:

  • trigger peaker plant dispatch,

  • delay coal/gas retirements,

  • justify new gas buildouts framed as “reliability.”

• This is where “digital growth” becomes an energy-system governance story, not an IT story.

Light pollution, traffic, and construction externalities

• Construction phase: heavy vehicle traffic, dust, noise, habitat disturbance.

• Operational phase: security lighting, 24/7 logistics, and community disruption.

Core sources on data centers, AI, and water use

Karen Hao – “Empire of AI – Water Footprint Changes” (correction note, Dec 17, 2025)
URL:
https://karendhao.com/20251217/empire-water-changes

Karen Hao – Empire of AI (book, water and environmental footprint discussion)
Author site:

https://karendhao.com

Li, Yang, Islam, Ren – “Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models” (arXiv preprint)
URL:
https://arxiv.org/abs/2304.03271

Li et al. – “Making AI Less ‘Thirsty’” (Communications of the ACM, DOI page)
URL:
https://dl.acm.org/doi/10.1145/3724499

Masanet et al. – “Data centre water consumption” (npj Clean Water, 2021)
URL:
https://www.nature.com/articles/s41545-021-00101-w

Environmental and Energy Study Institute (EESI) – “Data Centers and Water Consumption” (June 25, 2025)
URL:
https://www.eesi.org/articles/view/data-centers-and-water-consumption

Electricity generation, indirect water use, and grid impacts

Lawrence Berkeley National Laboratory – “The Environmental Footprint of Data Centers in the United States”
URL:
https://energyanalysis.lbl.gov/publications/environmental-footprint-data-centers

IEA – “Data centres and data transmission networks” (energy, emissions, and cooling impacts)
URL:
https://www.iea.org/reports/data-centres-and-data-transmission-networks

Air pollution, diesel generators, and public-health impacts

Washington State Department of Ecology – “Data Centers and Air Quality Permitting”
URL:
https://ecology.wa.gov/air-climate/air-quality/data-centers

IEEE Spectrum – “We Need to Talk About AI’s Impact on Public Health” (May 1, 2025)
URL:
https://spectrum.ieee.org/data-centers-pollution

The Guardian – “UK’s largest proposed datacentre ‘understating planned water use’” (Dec 19, 2025)
URL:
https://www.theguardian.com/environment/2025/dec/19/uk-largest-proposed-data-centre-planned-water-use-northumberland

Land use, water stress, and community impacts

Lincoln Institute of Land Policy – “Data Drain: The Land and Water Impacts of the AI Boom” (Oct 17, 2025)
URL:
https://www.lincolninst.edu/publications/land-lines-magazine/articles/land-water-impacts-data-centers

UNESCO – “Water stress and digital infrastructure” (background reports on water scarcity)
URL:
https://www.unesco.org/en/water-security

PFAS, firefighting foams, and persistent chemical risks

The Guardian – “Advocates raise alarm over PFAS pollution from datacenters amid AI boom” (Oct 4, 2025)
URL:
https://www.theguardian.com/environment/2025/oct/04/pfas-pollution-data-centers-ai

Clean Water Action – “PFAS-Containing Firefighting Foams”
URL:
https://cleanwater.org/pfas-containing-firefighting-foams

Interstate Technology & Regulatory Council (ITRC) – “PFAS Guidance: Firefighting Foams”
URL:
https://pfas-1.itrcweb.org/3-firefighting-foams/

Noise pollution and quality-of-life impacts

TechTarget – “Understanding the impact of data center noise pollution” (Dec 3, 2024)
URL:
https://www.techtarget.com/searchdatacenter/tip/Understanding-the-impact-of-data-center-noise-pollution

Climate, refrigerants, and fluorinated gases

European Environment Agency – “Fluorinated greenhouse gases (F-gases)”
URL:
https://www.eea.europa.eu/en/topics/in-depth/climate-change-mitigation-reducing-emissions/fluorinated-greenhouse-gases

U.S. EPA – “Greenhouse Gas Emissions from Data Centers”
URL:
https://www.epa.gov/energy/energy-and-you-affecting-climate-change