Seven 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 — and two of them pull against each other.
12 min read · Includes an interactive: Which Question Are You Asking?
One phrase, seven questions.
Electricity, carbon, water, air quality, cost, siting, materials. Different evidence bases, different geographies, and different people who can answer them. This guide is a thinking tool — the numbers here demonstrate the split, and Guides 2 through 5 carry the depth.
| # | The question you’re actually asking | True at the scale of | Where the evidence stands | |
|---|---|---|---|---|
| 1 | How much electricity does this take? | National total | Well established | Read the guide → |
| 2 | What does that electricity actually emit? | Global | Unresolved | Read the guide → |
| 3 | How much water, and whose? | Watershed | Well established | Read the guide → |
| 4 | What does it do to the air people breathe? | Airshed | Well established | Read the guide → |
| 5 | Who pays for the buildout? | Utility territory | Contested | Read the guide → |
| 6 | Who decided this gets built here? | County / parcel | Well established | Read the guide → |
| 7 | What did it take to make the hardware? | Supply chain | We haven't researched this | Tell us if you need this → |
Each one is true at a scale.
Most arguments in this space are scale errors delivered with conviction. Here is each of the seven, with the geography it actually operates in.
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 — and it is why carbon's framing gets borrowed for questions it doesn't fit.
Electricity
A facility is served by a specific balancing authority: a specific set of power plants that actually ramp up when it draws. Which is why “3% of global electricity” answers almost nothing about what any particular facility does at 7pm in July.
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 — so read the next one before deciding: naming water as local is what hands you a named water system, a drought plan, and a permit hearing.
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.
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.
Siting and consent
Zoning, water and sewer hookups, noise ordinances, and whether anyone was asked. The most local of the seven, and where most of the decisions that matter actually get made.
Materials
Global in flow, concentrated in harm, and usually somewhere else. We haven't done this research. What we'd need is a primary-source pass on extraction, manufacturing, and end-of-life — and until we've done it, we're not going to have an opinion.
Working on this? We’d like to hear from you →A claim that’s true at one scale, asserted at another, is how a real problem becomes an unwinnable argument.
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
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. Two things that trade off against each other were never one thing.
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 — New York City's entire daily supply is 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.
Sources for this split: smallBottle · renSpectrum · ncWater — 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,” and an honest account includes it.
Sources for this split: lbnl2024 · ieaEnergyAI — full citations below.
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 rooms that accept public comment and are currently attended almost entirely by people with a financial interest. That is not a reason to feel less. It is a reason to point it somewhere it moves something.
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 — renewables meeting nearly half of demand growth, the water and carbon tradeoff, the researchers’ own moderation about their numbers — were not exaggerated in the opposite direction. They were simply ignored.
A one-directional error pattern is a signal about the conversation, not about the technology.
Which is also why this guide names its own gaps. Question 7 — materials, mining, and e-waste — has no CPAI evidence base, and the marginal-emissions question (does new load delay fossil retirements? what clears at the margin?) is an open thread we have not closed. If you work on either, we’d like to hear from you.
Action for every level of influence.
For yourself
- Next time you hear the claim, ask which of the seven. Not as a gotcha — 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, both take about five minutes, and almost nobody has done either.
- Notice whether a statistic arrives with a scale attached. A number without a place is not yet a fact about anything.
For a conversation or a classroom
- Split the topic before you debate it. Agree on which of the seven 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.
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.
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.
Related
The Water Question
The most-quoted statistic in this debate has been retired by its own authors. What replaced it is more useful — and it points at your watershed, not the planet.
Electricity and Emissions
US data centers used 4.4% of national electricity in 2023, headed for 6.7–12% by 2028. What that does to emissions depends on something the number doesn’t contain.
Who Pays
The health and cost burden of the AI buildout lands on specific counties and specific ratepayers — and the algorithms optimizing for aggregate efficiency make that worse, not better.
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.
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 four guides.
Want CPAI to teach this in your community or classroom?
The seven-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.