Is AI Bad for the Environment? Real Footprint, Wrong Unit: The Four Ledgers

Is AI bad for the environment is asked as a yes or no question, and it deserves a better answer than either camp usually gives. The footprint is real, growing, and unevenly shared. The famous numbers used to prove it are mostly wrong.

Here is the honest shape of it. AI’s cost does not live in your individual prompts, where current measurements are surprisingly small. It lives at system scale: total electricity demand, specific watersheds, hardware turnover, and the places data centers cluster.

This page audits the cost in four separate ledgers: electricity and carbon, water, materials and electronic waste, and place. Each ledger gets its best current figure, its source, and its date, because on this topic an undated number is decoration.

The savings side gets the same treatment later, at full strength, never subtracted from the costs. Measured savings and scenario potentials are different kinds of numbers, and mixing them is how both camps cheat.

The Short Answer: Yes in Four Ledgers, No in the Viral Ones

Yes: AI is driving the fastest demand growth the electricity system has seen in decades, landing on real grids, rivers, and communities. The direction of the total is up on official projections.

No: a text prompt does not drink a bottle of water, does not use ten times a search in current measurements, and is not where the problem concentrates. The viral unit is the wrong unit, and policy built on it would miss the levers.

Ledger	Best current figure	Source and date	Confidence
Electricity and carbon	Data centers about 415 TWh in 2024, heading to about 945 TWh by 2030 in the base case, AI the main growth driver	IEA Energy and AI, 2025	High for the past, scenario for the future
Water	GPT-3 training about 700,000 liters direct; global AI withdrawal projected at 4.2 to 6.6 billion cubic meters by 2027	Li et al., Making AI Less Thirsty	Modeled, location sensitive
Materials and e-waste	Generative AI could add 1.2 to 5.0 million tonnes of e-waste across 2020 to 2030 scenarios	Wang et al., 2024	Scenario range, wide
Place	Demand clusters in a few regions, straining local grids and basins out of proportion to national shares	IEA and national statistics	High qualitatively, local data varies

The Unit Problem: Per Prompt and Grid Total Answer Different Questions

Almost every argument on this topic is two people answering different questions at each other. One quotes a fraction of a gram per prompt and calls the concern hysteria. The other quotes a doubling grid and concludes every prompt is a small crime.

Both numbers can be accurate while both conclusions fail. Per prompt figures measure efficiency: how well the machine converts electricity into answers. Grid totals measure scale: how many machines, running how many hours, on which grids.

A graded evidence review of this exact question puts it in one line worth stealing: the per prompt figures and the grid totals answer different questions.

This page keeps them in separate columns throughout, and neither column is allowed to impersonate the other.

Ledger One: Electricity and Carbon

The Grid Totals, as Ranges

Start with the most solid ground available: the IEA Energy and AI report. Its 2025 report put data centers at about 415 terawatt hours in 2024, roughly 1.5 percent of world electricity, growing about 12 percent a year.

The 2026 update strands put 2025 at about 485 TWh, with AI focused data centers up about 50 percent in a year to roughly 155 TWh.

The base case for 2030 is about 945 TWh, just under 3 percent of world electricity, with AI focused facilities tripling to about 465 TWh.

Two honest qualifications belong on those numbers. First, data centers are not AI: the totals include cloud, storage, and streaming, with AI named as the main driver of the growth rather than the whole of the level.

Second, the future figures are scenarios. The same agency’s 2035 range runs from about 700 to 1,700 TWh depending on efficiency and adoption.

The American picture comes from the Lawrence Berkeley National Laboratory report for the Department of Energy: US data centers used about 176 TWh in 2023, 4.4 percent of national electricity, with a projected range of 325 to 580 TWh, or 6.7 to 12 percent, by 2028.

On emissions, the IEA projects data center electricity emissions rising from about 180 million tonnes toward 300 million tonnes by 2035 in its base case, still under 1.5 percent of energy sector emissions.

Context the agency itself supplies: data centers account for about a tenth of global electricity demand growth to 2030, less than industrial motors, air conditioning, or electric vehicles.

Training: The Locked In Bill

Training a frontier model is a single enormous computation run over weeks on thousands of accelerators. Its footprint is spent before the model answers its first public question, which is why training dominates headlines.

Published estimates for recent frontier training runs reach a small town’s annual electricity in scale, in Epoch AI’s strand for one 2025 model. Treat these as modeled illustrations, not meter readings.

Training also matters beyond volume: it is concentrated in time and place, drawn from one grid at full tilt for weeks, and it repeats with every model generation.

Inference: Where the Compute Actually Is

The quiet fact that reframes the debate: reporting by MIT Technology Review puts inference, the everyday answering of prompts, at 80 to 90 percent of AI computing power now. Training built the models. The world’s prompts are what runs them, billions of times a day.

That split means the footprint question is mostly about scale of use and efficiency per answer, not any single training run, and efficiency work on serving moves the biggest number.

The Per Prompt Numbers, With Assumptions Attached

For one ordinary text prompt, the credible 2025 estimates cluster tightly once their boundaries are stated. The table below is the whole controversy in five rows.

Estimate	Per text prompt	Who and when	What it assumes or counts
About 3 Wh	One ChatGPT style query	De Vries, 2023	Older A100 hardware, very long outputs, maximum server power
About 0.3 Wh	Typical GPT-4o query	Epoch AI, Feb 2025	About 500 output tokens on H100 hardware
About 0.34 Wh	Average ChatGPT query	Sam Altman, Jun 2025	Company statement, no methodology published
0.24 Wh	Median Gemini text prompt	Google, Aug 2025	Measured in production, full facility overhead included
0.10 Wh	Same Gemini prompt	Google, Aug 2025	Counting only the active AI chips

Read the first and last rows together and the archaeology is visible: the tenfold fall is mostly changed assumptions and newer hardware, not a measured tenfold efficiency miracle in one product. Microsoft’s own comparison table now lists the old 2.9 Wh figure as outdated.

The most complete public measurement is the Google measurement of delivering AI at scale: a median Gemini text prompt at 0.24 Wh, 0.03 grams of CO2 equivalent, and 0.26 milliliters of water.

That count includes idle machines and facility overhead, and excludes the network, your device, and training. Google also reported the energy per median prompt fell about 33 times in the twelve months to May 2025 on its systems.

Hold two caveats with that figure: it is company measured, not independently audited, and outside researchers have argued its accounting boundary is incomplete.

The Task Ladder: Text Is the Bottom Rung

Prompts are not one thing. The TokWatt inference energy database harmonized 134 published estimates to one accounting boundary and published reference values by task.

AI task	Reference energy	Roughly equals, in text prompts
Short chat answer	0.34 Wh	1
Reasoning answer	4.7 Wh	About 14
Short video clip	124 Wh	About 360
One delegated coding agent task	150 Wh	About 440

The ladder carries the honest personal guidance this topic allows: text is cheap, reasoning costs more, and generated video and long agent runs are the heavy end by orders of magnitude. Anyone rationing their footprint should ration at the top of the ladder, not the bottom.

The same database found the five independent estimates for default production chat systems agree within a factor of 1.8, while the raw published literature spans a factor of 100.

Ledger Two: Water

Direct and Indirect, Withdrawal and Consumption

Water accounting defeats most coverage before it starts, so the distinctions come first. Direct water is consumed on site, mostly evaporated in cooling. Indirect water is consumed at the power plants generating the electricity.

Indirect is usually the larger share, often 80 percent or more.

Withdrawal is water taken and largely returned. Consumption is water evaporated or otherwise not returned. Headlines routinely quote withdrawal with consumption language. The careful explainer of AI water accounting keeps these terms straight, and this section follows its discipline.

Location rules this ledger more than any other. The same computation in a cool, wet region on a clean grid can carry a small fraction of the water footprint of the identical run in a hot, dry basin on fossil power.

The Bottle of Water Claim, Checked

The claim that a chatbot prompt drinks a bottle of water traces to two real sources, neither of which says what the headline says.

The first is the Making AI Less Thirsty study, which estimated that GPT-3 consumes a 500 milliliter bottle for roughly every 10 to 50 medium length responses, depending on when and where it runs, under an assumed 4 Wh per request that later work revised sharply downward.

The second origin is a 2024 Washington Post collaboration with researcher Shaolei Ren, which estimated 519 milliliters for a 100 word email written by GPT-4 at an average US data center, counting direct cooling water and the indirect water of electricity generation together.

Ren’s own later estimate for a comparable prompt, reported secondhand in a 2026 account, sits near 15 milliliters including offsite water. Secondhand means a reported update, not a published revision, and it prints labeled as one.

Training Water and Basin Level Stakes

The same study’s most quoted training figure: training GPT-3 in Microsoft’s US data centers could directly consume about 700,000 liters of freshwater, rising to about 5.4 million liters with electricity generation water included.

The study excluded failed runs and tuning searches, so its authors framed it as a floor for that era’s methods.

Its forward projection carries the policy weight: global AI water withdrawal of 4.2 to 6.6 billion cubic meters by 2027 under its scenarios, of which a fraction, 0.38 to 0.60 billion, would be consumed.

The basin point survives every recalculation: national percentages are small, but water is consumed in specific watersheds, some of them stressed, and a facility that is trivial nationally can be decisive locally.

Ledger Three: Materials and E-Waste

AI runs on physical hardware with short frontline lives: accelerators are replaced on a few year cycle as each generation outclasses the last, and the displaced equipment becomes electronic waste.

The modeling benchmark is the 2024 study in Nature Computational Science by Wang and colleagues: generative AI could add between 1.2 and 5.0 million tonnes of cumulative e-waste across 2020 to 2030 under its scenarios.

In the higher scenarios that reaches up to 12 percent of projected global e-waste in 2030. For scale, the Global E-waste Monitor put all e-waste at 62 million tonnes in 2022, heading toward 82 million by 2030.

The same study carries the hopeful figure: circular strategies, longer service lives, refurbishment, reuse, and recycling, could cut that waste stream by 16 to 86 percent, an operations decision rather than a research breakthrough.

Upstream sits mining and manufacturing: critical minerals, energy intensive chip fabrication, and embodied carbon that operational figures exclude. Disclosure here is the thinnest of any ledger, itself a finding worth stating.

Ledger Four: Place

Electricity and water are consumed somewhere specific, and AI demand is unusually concentrated. The IEA notes that data centers cluster in particular places, which makes them harder for a grid to absorb than the same demand spread evenly.

The US and China together account for nearly 80 percent of projected growth to 2030.

Clustering produces the local stories national averages cannot settle: grid connection queues, and communities negotiating over water, land, and reported backup generation concerns. Together they establish that the place ledger is real even where global percentages look small.

Place is also where the footprint is most governable: siting, disclosure before permits, and demand flexibility are local decisions with measurable effects on every other ledger.

The Numbers Moved: An Estimate Archaeology

This field’s figures have a half life of about a year, and the honest way to handle that is to show the strata. Two famous numbers, excavated:

The 3 Wh per query figure (2023) assumed A100 generation hardware, 2,000 token outputs, and maximum server power draw.

Within two years, independent modeling put a typical query near 0.3 Wh, the platform’s own chief executive said 0.34 Wh, and a measured median landed at 0.24 Wh. The old figure still circulates as if current.

The 500 milliliter bottle (2023 to 2024) belonged to GPT-3 era arithmetic and to one specific GPT-4 email estimate with full indirect water counted. Its own research lineage has since produced far smaller per prompt numbers, reported above with their secondhand flag intact.

The lesson is not that early researchers were dishonest; their assumptions were labeled in their papers. The failure happened in transmission, where the labels were stripped and the numbers repeated as timeless.

The Other Column: What AI Saves, Honestly Sized

Measured Cases

AI also works on the environment directly, and its strongest entries are measured, not promised.

The canonical case is Google DeepMind’s optimization of Google’s own data center cooling, which cut cooling energy by up to 40 percent in the company’s reporting.

Grid forecasting is the second solid shelf: machine learning predictions of wind and solar output reduce the errors that force grids to burn backup fuel. Weather modeling is a third, with national centers running AI models at a small fraction of legacy compute in reported strands.

Add methane detection, materials discovery, and logistics optimization, and the column is real. It is also, today, smaller in documented tonnage than the footprint’s growth, which is why the next subsection needs care.

Scenario Potentials, and Why They Do Not Net

The large numbers live one shelf up, in scenario studies. The IEA’s applications case suggests widespread AI adoption across energy, transport, and buildings could yield about 1,400 megatonnes of CO2 reductions by 2035.

A Grantham Institute study in npj Climate Action modeled AI applications in power, food, and transport cutting 3.2 to 5.4 billion tonnes of CO2 equivalent a year by 2035.

Both are serious studies. Neither is a measurement. Their figures must not be subtracted from measured footprints to declare AI carbon negative, a netting move this page refuses on method grounds.

The honest summary: AI’s climate applications could plausibly outweigh its footprint at maturity, and today they do not yet, on documented evidence. The gap between those sentences is the deployment agenda, not a debating point.

The Rebound: Efficiency and Growth at the Same Time

Google’s 33 fold per prompt efficiency gain in one year and the IEA’s doubling demand curve are both true, and together they describe the rebound effect.

Cheaper answers produce more answers: more users, longer reasoning, new modalities like video, and agent workflows that fire hundreds of prompts per task.

Efficiency is therefore necessary and insufficient: it bends the per unit curve while the total curve follows adoption. Serious policy works on the total: grid carbon intensity, siting, disclosure, and hardware lifetimes.

Verdict by Claim: A Scorecard

Famous claim	Verdict	The checked version
Each prompt uses a bottle of water	Outdated as stated	500 mL covered 10 to 50 GPT-3 responses under 2023 assumptions; later estimates run from about 15 mL down to a measured 0.26 mL
AI uses ten times a Google search	Outdated	It descends from the 3 Wh era; current chat estimates cluster near 0.24 to 0.34 Wh
AI will eat the grid	Exaggerated, directionally serious	About 1.5 percent of world electricity in 2024, base case just under 3 percent by 2030, but the growth rate leads all major loads
Training is the main problem	Mostly wrong now	Inference is 80 to 90 percent of AI compute; training is concentrated and visible, use is diffuse and dominant
AI’s savings cancel its footprint	Not on current evidence	Scenario potentials reach gigatonnes by 2035; documented savings today are real but smaller, and netting the two is a method error
AI is worse than bitcoin	Depends entirely on method	The two are accounted in different traditions with different boundaries; any one line verdict is hiding its assumptions

What Would Actually Reduce AI’s Footprint

The levers rank by the ledgers they move. Grid carbon intensity where the compute sits is the master lever: the same workload’s emissions swing severalfold between grids, so clean power matched to real hours and places matters more than distant offsets.

Second, siting and cooling design govern the water and place ledgers: clean grid, water abundant locations, and dry or liquid cooling where basins are stressed. Third, hardware lifetimes and circular supply chains address the waste ledger.

Fourth, disclosure: the field runs on estimates because providers are not required to publish measurements. Standardized, audited reporting would replace the archaeology sections of pages like this one with meter readings.

What You Can Do, Scaled Honestly

Individual action belongs here at its true scale. Your text prompts are a trivial slice of your footprint on current measurements; the top of the task ladder is where personal choices register, in video, heavy reasoning, and long agent runs.

Practical version: use the smallest tool that does the job, and keep video and agent use deliberate rather than ambient. Then put your real weight where the ledgers move: permitting, siting, and disclosure debates where facilities are built.

For the business version of the same question, GrandLoopy on how AI customer support is changing small online stores looks at how AI’s demands land on the small companies adopting it.

The GrandLoopy AI and technology coverage follows the wider technology picture as these figures are revised, and the GrandLoopy guides hub collects the newest GrandLoopy guides across news, fashion, health, and business.

The Glossary in One Pass

PUE: power usage effectiveness, the ratio of a facility’s total energy to the energy reaching its computers. It is the overhead multiplier hiding inside every per prompt figure.

Withdrawal versus consumption: withdrawal is water taken and mostly returned; consumption is water evaporated or otherwise not returned. AI headlines frequently quote the first with the emotional force of the second.

Inference: the everyday work of answering prompts, now 80 to 90 percent of AI compute.

Rebound effect: efficiency gains consumed by growth in use.

How to Read Any AI Footprint Claim: Five Questions

New figures land monthly, and most disputes dissolve under five questions asked in order.

1. Which unit: per task, or total? A per prompt figure cannot refute a grid total, and a grid total cannot shame a prompt. Claims that swap units mid argument are advocacy.
2. What year, and what hardware? A 2023 estimate built on A100 chips describes a retired world. Current claims should ride on current silicon or say why they do not.
3. What is counted? Chips only, the whole facility, the network, your device, training, manufacturing: each boundary is defensible, and none may be silently swapped for another.
4. Where? Grid carbon intensity and basin stress move the same workload severalfold. A global average is a starting point, never a local answer.
5. Measured, modeled, or projected? A production measurement, a bottom up model, and a 2035 scenario are three different kinds of number, and honest writing keeps their labels on.

The Timeline in One Pass

The debate’s strata, dated. 2019: early academic work prices transformer training in tonnes and starts the field. 2023: the 3 Wh estimate and the water paper set the viral terms, on the hardware of their moment.

2024: the LBNL report gives the US its official range, the Washington Post collaboration prices a GPT-4 email at 519 milliliters, and the Wang study models generative AI e-waste for the decade.

2025: the IEA publishes Energy and AI, Google publishes the first full production measurement of a median prompt, and OpenAI’s chief executive states 0.34 Wh without a method. 2026: the IEA update lifts the totals, and TokWatt harmonizes the per task literature.

The pattern explains itself: measurements keep arriving smaller per unit and larger in total, and each camp quotes its half.

Frequently Asked Questions

1. Is AI bad for the environment?

Yes in specific, evidenced ways, and no in the viral ones.

AI drives fast growing electricity demand, water use in specific basins, and hardware waste. A single text prompt, on current measurements, is a very small cost.

The problem is scale and place, not your prompt.

2. How much electricity do data centers use, and how much is AI?

About 415 terawatt hours worldwide in 2024, roughly 1.5 percent of global electricity, on IEA figures.

AI is not the whole total, but it is named the main driver of the growth. AI focused facilities sat near 155 TWh in 2025 strands.

3. Will AI double data center electricity by 2030?

In the IEA base case, yes: from about 415 TWh in 2024 to about 945 TWh by 2030, just under 3 percent of world electricity.

It is a scenario, not a prophecy. The agency’s 2035 cases span roughly 700 to 1,700 TWh.

4. How much energy does one ChatGPT prompt use?

Current credible estimates cluster around 0.24 to 0.34 watt hours for an ordinary text prompt.

Google measured a median Gemini prompt at 0.24 Wh in production. OpenAI’s chief executive has said about 0.34 Wh for ChatGPT, with no methodology published.

5. Does a ChatGPT prompt really use a bottle of water?

Not as usually told.

The bottle figure came from GPT-3 era research where 500 milliliters covered roughly 10 to 50 responses, plus one 2024 estimate for a full GPT-4 email counting power plant water.

Later estimates run from about 15 milliliters to a measured 0.26.

6. How much water does AI use in total?

The most cited projection, from the Making AI Less Thirsty study, puts global AI water withdrawal at 4.2 to 6.6 billion cubic meters by 2027, a fraction of it consumed.

Training GPT-3 was estimated at about 700,000 liters direct.

7. Why did the per query energy estimate fall from 3 Wh to about 0.3 Wh?

Mostly because the 2023 estimate’s assumptions expired.

It modeled older A100 hardware, very long outputs, and maximum server power.

Newer hardware, realistic output lengths, and one production measurement brought the cluster near 0.3 Wh.

Microsoft now lists the old figure as outdated.

8. Is AI worse than bitcoin for the environment?

There is no method clean verdict.

The two are measured in different traditions with different boundaries, and both totals move yearly.

Bitcoin is a single purpose total.

AI is an estimated share of general data center demand.

Any one line answer hides its assumptions.

9. Does AI use more energy than a Google search?

The famous ten times figure descends from the outdated 3 Wh era.

With current chat estimates near 0.24 to 0.34 Wh, the gap over a conventional search is far smaller than the slogan claims.

Treat any exact multiple with suspicion.

10. What uses more energy, AI video generation or chatting?

Video, by orders of magnitude.

TokWatt reference values put a short chat answer at 0.34 Wh and a short video clip at about 124 Wh, roughly 360 chats.

A reasoning answer sits near 4.7 Wh, and a delegated agent task near 150 Wh.

11. Is training or everyday use the bigger problem?

Everyday use.

Reporting puts inference, the answering of prompts, at 80 to 90 percent of AI computing power now.

Training is concentrated and visible, which draws the headlines, but the traffic outweighs the construction over any model’s life.

12. How much CO2 does training a large model emit?

Providers rarely publish training energy, so figures are modeled.

Published estimates for recent frontier runs reach the scale of a small town’s annual electricity, in Epoch AI’s strand for one 2025 model.

Read them as scale illustrations, not meter readings.

13. What is AI’s total carbon footprint?

There is no single audited total.

The IEA projects data center electricity emissions rising from about 180 million tonnes toward 300 million by 2035, for all workloads.

Patterns modeled AI systems alone at roughly 33 to 80 million tonnes in 2025.

14. Does AI create a lot of e-waste?

On scenario modeling, yes.

A 2024 Nature Computational Science study put generative AI’s added e-waste at 1.2 to 5.0 million tonnes across 2020 to 2030, up to 12 percent of projected 2030 global e-waste in higher scenarios, driven by fast accelerator replacement.

15. Where does the water for AI cooling actually go?

Two places.

Direct water evaporates in cooling systems at the data center. Indirect water is consumed at power plants generating the electricity, usually the larger share, often 80 percent or more.

The grid behind a facility shapes its water footprint.

16. Why does it matter where a data center is built?

Because every ledger lands locally.

The same workload’s emissions swing with the grid’s carbon intensity, and its water stress depends on the basin. National percentages can look small while one town absorbs a facility the size of a small city.

17. Can renewable energy power AI’s growth?

Partly, and increasingly, but the honest answer is about matching.

Clean power helps most when generated in the same place and hours the compute runs. Distant annual offsets do less.

The IEA treats low carbon supply build out as this decade’s central race.

18. Is AI helping the environment too?

Yes, in measured cases.

AI optimization cut Google’s data center cooling energy by up to 40 percent in company reporting, and forecasting reduces the errors that force grids to burn backup fuel.

The documented column is real, and currently smaller than the growth.

19. How much could AI cut global emissions by 2035?

Scenario studies, not measurements.

The IEA’s applications case suggests about 1,400 megatonnes of CO2 reductions by 2035 from AI adoption in energy, transport, and buildings.

A Grantham study modeled 3.2 to 5.4 billion tonnes a year by 2035 across power, food, and transport.

20. Do AI’s savings cancel out its footprint?

Not on current evidence, and the arithmetic is a method error.

The savings figures are modeled potentials, and the footprint figures are measurements and estimates of what already runs.

Subtracting one kind from the other declares a winner neither study claimed.

21. What is the Jevons paradox in AI energy?

It is the rebound pattern in this field.

Efficiency makes each answer cheaper, and cheaper answers get used far more, for heavier tasks like video and agents.

Google’s per prompt energy fell about 33 times in a year while total demand headed toward doubling.

22. Why do estimates of AI’s footprint vary so much?

Because almost nothing is metered in public.

Providers are not required to publish measurements, so researchers model from hardware specs, assumed workloads, and chosen boundaries. Change the boundary, model size, or grid, and the figure moves severalfold.

Convergence is improving.

23. What is PUE and why does it change per prompt numbers?

Power usage effectiveness: a facility’s total energy divided by the energy reaching its computers.

A PUE of 1.2 means each watt of computing carries 0.2 watts of overhead.

It is why Google’s same prompt reads 0.10 Wh for chips alone and 0.24 Wh with the building.

24. What is the difference between water withdrawal and consumption for AI?

Withdrawal is water taken and largely returned to its source.

Consumption is water evaporated or otherwise not returned. AI headlines often quote withdrawal with consumption’s emotional force.

The scarcer figure is consumption, and it is usually the smaller of the two.

25. Are newer AI models more efficient?

Per task, clearly.

Google reported the energy per median Gemini prompt falling about 33 times in twelve months, and independent chat estimates converged near 0.3 Wh.

The catch is the rebound: total demand still rises because use grows faster than efficiency.

26. Which AI task is worst for the environment?

Generated video and long agent runs top the ladder.

Reference values run from 0.34 Wh for a chat answer to about 124 Wh for a short video and 150 Wh for one delegated coding agent task.

Heavy reasoning sits between.

Task choice matters more than prompt guilt.

27. Should I stop using chatbots for the environment?

On the evidence, no.

Text prompting is a trivial slice of a personal footprint on current measurements, and giving it up changes no ledger in this article.

Ration the heavy end instead, and spend civic energy on siting and disclosure debates.

28. What can one person actually do about AI’s footprint?

Three honest things.

Use the smallest tool that does the job, keep heavy generation deliberate, and prefer providers that publish measurements. Then treat the real arena as collective: grids, permits, and disclosure rules where facilities are built decide the totals.

29. What should companies disclose about AI energy use?

Energy, water, and carbon per service, measured, with boundaries stated.

The field runs on estimates because disclosure is voluntary, which is why researchers model and companies rebut.

Standardized audited reporting is the single reform researchers ask for most.

30. Is local AI on my own device greener than cloud AI?

It depends on utilization and hardware.

A modern data center runs efficient chips at high utilization with low overhead, but adds facility and network costs. A home device idles on older chips yet avoids the building.

For heavy steady work the cloud usually wins.

31. Does AI use drinking water?

Data centers cool mainly with water systems, and the water is fresh water of drinking quality in many designs, which is why the phrase sticks.

Per prompt quantities are small on current estimates.

The real question is basin level, at specific facilities.

32. Why are data centers clustered in a few places?

Physics and economics.

Facilities cluster near cheap reliable power, fiber, water, land, and existing grid capacity, concentrating in regions like Northern Virginia and parts of Ireland.

Clustering is efficient for operators and heavy for host grids planned over decades.

33. What is marginal emissions accounting and why do people argue about it?

Average accounting spreads a grid’s emissions across all users.

Marginal accounting asks which generator actually ramps up when new demand arrives, often a gas plant, which raises the footprint.

Companies prefer averages.

Researchers argue the marginal view is honest for new load.

34. How accurate are the viral claims about AI and the planet?

The famous ones are mostly outdated rather than invented.

The bottle and ten times claims began as labeled estimates, then lost their labels in transmission. The grid growth claims, by contrast, have held up and come from the IEA.

Date and method separate the fates.

35. Is AI’s footprint growing faster than its efficiency?

Today, yes, in total.

Per unit efficiency is improving remarkably, roughly 33 fold in a year on Google’s systems, while total demand heads toward doubling by 2030 in the IEA base case.

Adoption is growing faster than efficiency.

That is the rebound in one sentence.

36. What does the IEA actually say about AI and energy?

Cautiously, with ranges.

Its 2025 Energy and AI report put data centers at about 415 TWh in 2024, roughly 1.5 percent of world electricity, heading to about 945 TWh by 2030 in the base case, with AI the main growth driver.

37. Are AI data centers making electricity more expensive (reported strand)?

This is a reported strand, not a settled finding.

Coverage in several regions links rapid data center growth to pressure on household electricity prices through grid upgrade costs and capacity markets.

Economists dispute the size and direction by market.

38. What happens to old AI chips and servers?

Mostly into the general electronics waste stream today.

Servers and accelerators are refurbished, resold, parted out, or recycled at varying rates, with metals recovered and toxic fractions handled unevenly.

The Wang study modeled circular strategies cutting the added waste by 16 to 86 percent.

39. Could AI data centers help the grid instead of straining it?

In principle, yes.

A large flexible load that can shift work across hours is what a renewable heavy grid needs, and carbon aware scheduling already moves some workloads to cleaner hours.

Whether facilities are built for flexibility is a policy choice.

40. Is nuclear power part of AI’s energy future?

Yes, in the plans.

Large technology firms have signed nuclear agreements, from restarted plants to small modular reactor development, for firm low carbon power at data center scale.

Timelines are long and contested.

Treat announcements as intentions, not power already flowing.

41. What is green AI?

Green AI is the research and policy movement, named in a 2019 paper by Schwartz and colleagues, arguing that efficiency and environmental cost belong in how AI work is reported and rewarded, alongside accuracy.

Red AI is its counterpart label.

42. What would settle the AI environment debate?

Measurements replacing estimates.

If providers published audited energy, water, and carbon per service, with boundaries fixed by a standard, the archaeology sections of articles like this one would disappear.

Until then, honest knowledge is ranges with strands attached.

Conclusion

Is AI bad for the environment, answered at the length the evidence allows: it carries a real and growing cost, argued about in the wrong unit. The per prompt numbers that dominate the argument are the smallest and fastest improving part of the story.

The durable facts point the other way, toward scale and place. Total demand doubling on official base cases, water drawn in specific basins, hardware turning over in a few short years, and growth concentrated where grids and communities must absorb it.

Neither camp’s slogan survives the ledgers. The footprint is not a prompt level catastrophe, and the savings are not yet an offset. What would change the answer is measurable: cleaner grids where compute sits, disclosed measurements instead of estimates, and hardware that lives longer.

Keep the four ledgers separate, keep the dates on the numbers, and the next headline, from either direction, will be easy to grade. The GrandLoopy guides hub is where GrandLoopy files explainers like this one as the figures are revised.