Center for Practical AI
AI and the Environment · Guide 1 of 10Start here

Ten arguments, one word.

Before you can decide whether AI’s environmental cost is acceptable, you have to know which cost you mean. They are not the same size, not in the same place, and not equally well understood. Two of them pull against each other.

12 min read · Includes an interactive: Which Question Are You Asking?

The split

One phrase, ten questions.

Electricity, carbon, water, air, cost, participation, hardware, enforcement, sound, disclosure. Different evidence bases, different geographies, and different rooms where the answers get decided. This guide is a thinking tool: the numbers here demonstrate the split, and the other nine guides carry the depth.

1How much electricity does this take?
National totalUtility commissionWell established
Read the guide
2What does that electricity actually emit?
GlobalUtility commissionUnresolved
Read the guide
3How much water, and whose?
WatershedState agencyWell established
Read the guide
4What does it do to the air people breathe?
AirshedState agencyWell established
Read the guide
5Who pays for the buildout?
Utility territoryUtility commissionContested
Read the guide
6Who gets a say before this gets built?
County / parcelLocal governmentWell established
Read the guide
7What did it take to make the hardware?
Supply chainNo working venueContested
Read the guide
8What happens when the rules are broken?
AirshedCourtsIn litigation
Read the guide
9What does a data center sound like next door?
NeighborhoodLocal governmentWell established
Read the guide
10What are we not told?
County / parcelNo working venueWell established
Read the guide
The pedagogical heart

Each one is true at a scale.

Most arguments in this space are scale errors delivered with conviction. Here is each scale in the series, widest to narrowest, with the questions that actually operate there.

Global

Carbon

A ton of CO₂ emitted in Virginia and a ton emitted in Ireland do the same thing. This is the one place in this entire topic where “we're all in it together” is literally, physically true. It is also why carbon's framing gets borrowed for questions it doesn't fit.

Grid region

Electricity

A facility is served by a specific balancing authority: a specific set of power plants that actually ramp up when it draws. “3% of global electricity” therefore answers almost nothing about what any particular facility does at 7pm in July.

Watershed

Water

Water is not fungible across basins. Moving a facility from southern Arizona to Ohio meaningfully changes the water story and barely changes the carbon story. That sentence gets misheard as dismissal more than any other in this section. Naming water as local is what hands you a named water system, a drought plan, and a permit hearing.

Airshed

Air and health

Pollution travels downwind, unevenly, and not forever. Per-household burden in the worst-affected counties runs about seven times the national average. A national figure cannot show you that, because averaging is the operation that hides it.

Utility territory

Money

The people who pay are a defined, nameable list: everyone on one utility's bill. Not a nation, not a generation. A service territory, with a commission that sets its rates in public.

County / parcel

Participation, enforcement, and what's disclosed

Zoning, permits, moratoriums, and whether anyone was asked, or could be. Three of the ten questions live here: who gets a say, what happens when rules are broken, and what nobody is required to report. Whether a hearing legally exists for a given project is decided at this scale, and increasingly that is the whole fight.

Neighborhood

Sound

The narrowest scale in the series, added because the complaints demanded it. A data center's hum is felt by the specific houses within earshot of the fence line, measured at a property line in decibels, and governed by a local code that either has a working limit or doesn't.

Supply chain

Materials

Global in flow, concentrated in harm, and usually somewhere else. This was the series' declared gap until August 2026. The research is done now, and the honest summary is: mostly copper for grid infrastructure, an e-waste projection that was recalibrated roughly tenfold, and several key numbers nobody publishes at all.

Scale is where an impact lands. Since August 2026 the series tracks a second axis: the venue where each question is decided, whether a planning board, a state agency, a utility commission, a courtroom, or, for two of the ten, no working venue at all. The hub’s question matrix carries both axes, and the participation guide is about what happens to the venues themselves.

A claim that’s true at one scale, asserted at another, is how a real problem becomes an unwinnable argument.

The proof they're separate

Two of them pull against each other.

If these were really one issue with several faces, they would move together. Two of them demonstrably don't.

“Water use by data centres can be negatively coupled with CO2-equivalent emissions, with methods of reducing water consumption increasing carbon emissions in some cases.”

Chien, Gupta, Ren, Sriraman & Tomlinson (2026), Nature Reviews Clean Technology

WatershedGlobal

Closed-loop “zero water” cooling reduces water use and raises electricity demand, and therefore carbon. Anyone demanding both zero water and zero carbon is asking for something the engineering does not currently offer.

The method

Two arguments, taken apart.

Every guide in this series names the overstated version and the dismissive version before saying what the evidence supports. You should not be able to tell from the page which side we're on, only which claims survive.

AI is using up our water.

The version that goes too far

Treats water as a global stock being drawn down, and leans on a per-query figure the authors themselves retired. There is no shared global pool to use up. There is your basin, on a hot day, with a specific amount of spare capacity.

The version that waves it away

Points at small annual totals at particular facilities — Meta's Forest City site used about 4.2 million gallons in all of 2024 — and concludes there's nothing here. Annual totals are the wrong unit. In Ren's words: “Only comparing the annual totals can obscure the real water challenge.”

What the evidence supports

US data centers could require 697–1,451 million gallons a day of new peak water capacity through 2030, against New York City's entire daily supply of about 1,000 MGD, at a build cost of roughly $10B to $58B. Whether any of that lands on you depends entirely on your basin.

Watershed

Sources for this split: smallBottle · renSpectrum · ncWaterUse — full citations below.

How much electricity is this, really?

The version that goes too far

Treats every kilowatt-hour as maximally dirty and every projection as a forecast. LBNL's own range for 2028 is 6.7% to 12% of US electricity, nearly a factor of two, because it depends on assumptions about hardware efficiency and growth in the rest of the economy.

The version that waves it away

Uses a global denominator to answer a local question. About 3% of world electricity by 2030 is real and proportionate, and it tells someone in a county facing a multi-gigawatt interconnection queue nothing at all about their grid or their bill.

What the evidence supports

Both the demand surge and the renewable supply response are documented in the same reports. Renewables are growing about 22% a year for data centres and meeting nearly half of demand growth. That is a materially different story from “AI is burning the planet.”

National total

Sources for this split: lbnl2024 · ieaEnergyAI — full citations below.

The paragraph this section turns on

Precision is not indifference.

Saying that water is a local problem is not a way of saying it doesn’t matter. This gets misheard often enough that it is worth stating as plainly as possible: naming the scale correctly is the move that makes the problem actionable at all.

A global problem gives you guilt. A watershed problem gives you a phone number.

Specifically: a named municipal water system, a drought contingency plan, a permit hearing with a date on it, and a disclosure requirement that either exists or doesn’t. Every one of those has a person attached whose job is to answer questions about it. None of that is available to you at planetary scale, which is why the planetary framing, however sincerely meant, reliably produces people who care a great deal and do nothing.

The same applies to your own use. Your prompt count is not the lever. The per-response footprint is small, the reason many people believe otherwise is a statistic its own authors retired, and the decisions that determine the actual impact — where a facility is sited, what it is cooled with, what it must disclose, who pays for the interconnection — are made in public proceedings. Whether you can reach those proceedings depends on the approval pathway, and that question now has its own guide.

A pattern worth noticing

The retelling has a direction.

When we traced the most-repeated figures in this literature back to their sources, something consistent turned up.

A cooling-tower engineering constant

Originally: 80% evaporation, specific to towers with good water quality

In retelling: became an industry-wide average, overstating withdrawals by roughly 60%

An IEA figure

Originally: 100 MW ≈ 2 million litres a day

In retelling: became Shaolei Ren's, a researcher who publishes a different and more careful number

A 2023 GPT-3 estimate

Originally: 500 ml per 10–50 responses, modeled at specific facilities

In retelling: became a fact about today's models

A federal range

Originally: 325–580 TWh by 2028

In retelling: became “about 600 TWh,” the top of the high scenario rounded up

Every figure that drifted, drifted upward and toward AI causation. Not one drifted the other way.

Meanwhile the findings that complicate the story were not exaggerated in the opposite direction: renewables meeting nearly half of demand growth, the water and carbon tradeoff, the researchers’ own moderation about their numbers. They were simply ignored.

A one-directional error pattern is a signal about the conversation, not about the technology.

This guide names its own gaps for the same reason. The materials question sat here as a declared blank until August 2026, when the research landed as the hardware guide. The marginal-emissions question (does new load delay fossil retirements, and what clears at the margin?) remains an open thread we have not closed. If you work on it, we’d like to hear from you.

What you can do

Action for every level of influence.

1

For yourself

  • Next time you hear the claim, ask which of the ten. Not as a gotcha, but as the thing that makes the conversation possible at all.
  • Look up which balancing authority and which river basin serve your county. Both are public, and both take about five minutes.
  • Notice whether a statistic arrives with a scale attached. A number without a place is not yet a fact about anything.
2

For a conversation or a classroom

  • Split the topic before you debate it. Agree on which of the ten questions is on the table, out loud, first.
  • Assign different questions to different people. The disagreement usually dissolves into two people being right about different things, which is a much better place to end up than a winner.
  • Use the interactive on this page as the opening exercise. The count on the screen does the work. A sentence containing five questions is visibly a sentence that cannot be answered as one.
3

For your community

  • Find out whether your municipal water system knows what its large industrial customers use. In many places it folds into city totals and no facility-level figure exists at all.
  • Ask whether your utility has an approved large-load tariff, and who pays for interconnection upgrades. If the answer is "that's still being decided," you have found the room that matters.
4

For policy

  • The single highest-leverage ask in this literature is mandatory peak water reporting, not annual totals. Annual volume is the unit that hides the constraint.
  • Second: require disclosure as a condition of tax incentives. A decade of North Carolina determinations required no company to report a single number, and the state now reads its own buildout off a commercial database.

Where this leads

Reading is one thing. Practicing it is another.

The Applied AI Certification builds practical AI fluency across all six domains — the working competence that advances toward proficiency, with structured practice, feedback, and a cohort on the same problems.

Sources

Research & further reading.

This guide's job is the split, not the depth. Every figure here is carried in full by one of the other nine guides.

Federal government reportModeled estimate · not metered measurementNational total
Shehabi et al., Lawrence Berkeley National Laboratory / DOE (Dec 2024)2024 United States Data Center Energy Usage ReportUS data centers used 4.4% of national electricity in 2023 — 58 TWh (2014) → 176 TWh (2023) → a projected 325–580 TWh by 2028, or 6.7–12% of US electricity. Cite the range, never the top of it. Published December 2024, so it predates the 2025–26 capital expenditure surge. The report's own limitations section notes it says nothing about where the load lands, which is why its figures are badged national rather than grid.
International agency reportModel projection · scenario-dependentGlobal
International Energy Agency (2025)Energy and AIGlobal data centre electricity demand roughly doubling by 2030 — about 485 TWh to about 950 TWh, near 3% of world electricity. Also the source of the finding that renewables are growing about 22% a year and meeting nearly half of that demand growth, which belongs on the page for the same reason the alarming figures do. The hardware guide additionally cites this report's critical-minerals estimate: roughly 512 kt of copper demand from data centres by 2030, about 2% of global copper demand, the first institutional attribution of mineral demand to data centres with a defensible denominator.Citation still being verified against our research files.
Peer-reviewed studyModeled estimate · not metered measurementWatershed
Li, Yang, Islam & Ren (2023/2025)Making AI Less "Thirsty"Source of the retired 500 ml figure and of its replacement: one GPT-3 output of 150–300 words consumed 16.9 mL total in an average US data center — 2.2 mL onsite cooling plus 14.7 mL at the power plant. Most of the water is not in the building. Ren notes later models are likely more efficient.
Preprint · not yet peer-reviewedModeled estimate · not metered measurementWatershed
Han, Li, Wierman & Ren (2026)Small Bottle, Big PipeUS data centers could require 697–1,451 million gallons per day of new peak water capacity through 2030 — New York City's entire daily supply is about 1,000 MGD — at a build cost of roughly $10B to $58B; or 227–604 MGD if water intensity falls 10% a year. Ren: "Only comparing the annual totals can obscure the real water challenge." The constraint is peak capacity, not annual volume.
Journalism · secondary reportingWatershed
Shaolei Ren, IEEE Spectrum (September 2025)Where the fleet-wide water figures actually come fromCompanies report between 45% and 60% of withdrawn water is consumed, and Ren explicitly assumes 50% fleet-wide when converting LBNL's consumption figures to withdrawals. Also the source of the seasonality framing: in Arizona, a data center's monthly water use during the summer can run nearly twice its average level. Ren's moderating line lives here too: "on the national level, data centers' water use is relatively modest." Trade press, not a peer-reviewed paper.
Peer-reviewed studyGlobal
Chien, Gupta, Ren, Sriraman & Tomlinson (2026)Strategies and design for increasing AI sustainabilityNature Reviews Clean Technology. Quote in full, never truncated: "Water use by data centres can be negatively coupled with CO2-equivalent emissions, with methods of reducing water consumption increasing carbon emissions in some cases." The dropped second clause is both the mechanism and the paper's own hedge. Closed-loop "zero water" designs raise electricity demand.
Peer-reviewed studyModeled estimate · not metered measurementAirshed
Han, Wu, Li, Wierman & Ren (2026)Health-Informed ComputingCommunications of the ACM. $20.9B in public health costs is the high-growth 2028 scenario; low-growth is $11.7B and the 2023 baseline is $6.67B — always state which. About 1,300 premature deaths and 600,000 asthma symptom cases in the 2028 high scenario. Scope 2 (generation) dominates Scope 1 (onsite backup diesel) by roughly 12 to 1. The worst-affected counties carry about 7× the national average per household.
Journalism · secondary reportingWatershed
Guajardo, WUNC (April 10, 2026)What North Carolina facilities actually useMeta's 30 MW Forest City facility used about 4.2 million gallons in all of 2024 — roughly 11,500 gallons a day. Forest City town manager Janet Mason: "We were shocked at how little water they used, because I think we were hoping to sell a little bit more water to them." The figure exists because a reporter asked, not because anyone had to publish it.
Journalism · secondary reportingWatershed
WRAL (April 17, 2026)What nobody has to report36 states proposed data center transparency and oversight legislation in 2026. "In North Carolina, there are no statewide proposals specifically requiring data centers to disclose water use or setting limits during drought conditions." Large facilities are served by municipal systems, so their use folds into city totals and no facility-level figure exists.
Statute / session lawUtility territory
N.C. Gen. Stat. §143-215.22HRegistration of water withdrawals and transfers requiredRegistration is required of "any person who withdraws 100,000 gallons per day or more of water from the surface or groundwaters of the State." This is the threshold Meta's Forest City facility falls below at roughly 11,500 gallons a day, which is why no state filing describes it.
Government agency memo or determinationCounty / parcel
North Carolina Department of Commerce (April 6, 2026)Memo to the Governor's Energy Policy Task ForceAbout 800 MW operational in North Carolina as of December 2025, with roughly 6,300 MW in the pipeline. Commerce issued 37 written data center tax-exemption determinations between 2015 and 2025; companies are not required to report actual investment or exemption value, and the list of 37 is not published. Commerce states openly that its MW figures come from Baxtel, a commercial tracker — a disclosure worth noticing, because it means the state is reading its own buildout off a private database.
Last reviewed: August 2026We review this page quarterly. Statistics in this category change rapidly.One research gap is named on this page rather than papered over: the marginal-emissions question, which is open in the literature and has no confident answer here. The materials gap this note used to name was closed in August 2026 by the hardware guide.

Want CPAI to teach this in your community or classroom?

The ten-questions split works as a 45-minute session and travels well beyond AI. It is a general method for arguments where a number is true at one scale and asserted at another.