AI & Environmental Sustainability

Novice · Governance, Ethics & Equity track · ~45 min hands-on + readings and quiz

← Back to the Governance, Ethics & Equity track

AI has a physical footprint — understand it and make proportionate choices.

What you’ll be able to do

  • Describe the energy and water cost of AI compute, with a sense of the actual range (not a single made-up number)
  • Explain why public estimates for AI’s footprint disagree so much — and what that does and doesn’t tell you
  • Right-size model use to the task
  • Explain why an AI agent can be cost-competitive with a person and still use far more electricity for the same result
  • Weigh benefit against environmental cost in your work, and in conversations with students or colleagues who are anxious about it

Overview

AI runs on data centers that consume electricity and water at scale. The footprint is real and growing, and it’s an environmental-health issue squarely in DSPH’s wing. Awareness doesn’t mean abstention — it means proportion.

It’s also a topic where the public numbers are a mess. Depending on the source, “how much energy does one AI query use” gets answered anywhere from about a tenth of a watt-hour to several watt-hours — a 30x spread — because sources disagree on what to count (chip only vs. the whole data center, a short answer vs. a long one, whether to fold in the one-time cost of training the model). The honest answer is a range with sources, not a single confident number. That’s also the more useful thing to model for students: this is a case where “I don’t know exactly, but here’s the defensible range and why it’s a range” is the correct scholarly answer, not a cop-out.

Right-sizing is the practical lever: use the smallest model that does the job, avoid wasteful re-runs, and reserve heavy compute for tasks that warrant it.

Right-sizing matters most for agentic work, where the model runs for hours and launches its own experiments rather than answering one question. In July 2026 METR proposed a metric it calls the expenditure horizon: the dollar amount at which an AI agent’s progress on a task equals what a human researcher achieves for the same spend. On a hard ML-engineering benchmark, the best model tested crossed at about $3,300 — roughly comparable to a person. The energy accounts are not comparable at that point. A dollar spent on GPU time buys on the order of 15x more electricity than a dollar spent on someone’s hours, so a task where AI and a person cost the same can still differ by an order of magnitude in kilowatt-hours. Cost parity and energy parity are different thresholds, and the metrics the field is building measure the first one.

That cuts both ways, and the explorer shows both directions: a single chatbot query uses far less energy than doing the same small task by hand, while a five-day autonomous agent run on 32 GPUs uses far more than the researcher it replaces. Task type decides the answer.

🔎 Open the AI Environmental Footprint Explorer

The explorer above puts specific AI tasks (a quick chatbot query, a long reasoning query, an image, a 5-second video) on the same scale as things you already have intuition for — a Google search, an hour of streaming, a load of laundry, a mile of driving — and against the “manual” way of doing the same task. Every figure shows its low/mid/high range and its source. Activity 3 below walks you through it.

Practice activities

Activity 1 · Novice — Right-size a task

Time ~12 min · Tools ChatGPT Edu (model picker)

Goal. Use the smallest model that does the job.

Setup. Use this abstract (approximately 200 words) for the right-sizing comparison:

Built environment and pedestrian injury severity: A multi-level analysis of 12,400 crashes in Philadelphia, 2017–2023. We examined the association between built-environment features and pedestrian crash severity using a multilevel logistic regression framework. Crash records from PennDOT were geocoded and linked to road-segment characteristics (lane count, speed limit, presence of sidewalks, crosswalk markings, and lighting) and census-tract-level variables (population density, median income, percent non-white). Severe injuries (KSI: killed or seriously injured) were modeled as the outcome. Roads with ≥4 lanes had 2.3 times the odds of KSI compared to 2-lane roads (95% CI: 1.8–2.9). Absence of marked crosswalks was associated with 1.7 times the odds of KSI (95% CI: 1.3–2.2). Tracts in the lowest income quartile had 40% higher KSI odds after adjusting for road features. A small-model approach (GPT-4o mini or equivalent) is adequate for summarizing and reformatting this kind of text; a frontier model adds little value for straightforward summarization tasks but consumes substantially more compute.

Steps.

  1. Run this prompt on a smaller/faster model (e.g., GPT-4o mini or the non-default option in ChatGPT Edu):

    Summarize this abstract in 3 bullet points for a non-technical audience. Keep every number accurate.

  2. Run the same prompt on the largest available model (e.g., GPT-5.5).

  3. Compare using this checklist:

    Criterion Small model Large model
    All numbers accurate?
    Readable for non-expert?
    Captures the 3 key findings?
    Any added insight the small model missed?

  4. Note where the smaller model is good enough — and where, if anywhere, the larger model genuinely added value.

Expected result. For a straightforward summarization task like this, the small model should perform comparably. You should be able to identify that the large model was not necessary.

Check your work. Did the big model actually add value here, or just cost more energy and money? For summarization, the answer is usually no.

Common pitfalls. Defaulting to the biggest model for everything is wasteful. The energy difference is real and larger than most people guess: measured comparisons put a frontier model at roughly 50x (order-of-magnitude range 10–100x) the energy of a small model on the same simple query.

Stretch (optional). List your five most common AI tasks and estimate which need a frontier model vs. a smaller one. You may find that 70–80% of your daily use doesn’t require the largest model.

Activity 2 · Novice — Connect AI's footprint to environmental health

Time ~13 min · Tools a reading + ChatGPT Edu

Goal. Frame AI’s resource use as a DSPH-relevant issue.

Setup. Read one source on data-center energy or water use (in your DSPH Readings folder).

Steps.

  1. Note the key figures (energy, water).

  2. Draft a framing:

    Summarize the environmental-health and equity concerns of data-center siting in two sentences for a public audience.

  3. Verify the figures against the reading.

Expected result. A short, accurate framing connecting AI to environmental health.

Check your work. Confirm the numbers — don’t repeat an AI-stated figure unchecked.

Common pitfalls. This is a real tradeoff, not a reason to abstain; aim for proportion.

Stretch (optional). Draft a one-line “responsible use” norm for your team.

Activity 3 · Novice — Put AI's footprint on a scale you know (and find its limits)

Time ~20 min · Tools the AI Environmental Footprint Explorer

Goal. Practice reasoning with ranges and sourced uncertainty instead of a single scary or reassuring number — and know where the honest answer is “we don’t know.”

Setup. Open the AI Environmental Footprint Explorer in a new tab.

Steps.

  1. In Module A, pick “Typical frontier chatbot query.” Note the mid-range energy figure, then switch the toggle to Low and High. Write down all three numbers and the everyday activities they land closest to.

  2. Still in Module A, switch to “One 5-second AI video clip.” Notice the confidence flag on this one — read why it’s marked low-confidence before you use this number in front of a class or in writing.

  3. In Module A, switch the Single task / Annual use toggle to Annual use. Set the usage-rate slider to “Typical,” then to “Heavy,” for a task you personally use often. Read the per-capita callout underneath the chart — write down the percentage of your annual electricity (and water, if shown) that one year of this usage represents. Then read the caveat about the “Typical” preset averaging every weekly user, including people who barely touch it — does that change how you’d describe your own usage relative to the average?

  4. In Module B, compare an AI task against 20 minutes of manual laptop use for the same task. Read the four caveats underneath it. Pick the one caveat you find most underappreciated in public conversation about AI and water/energy, and write one sentence on why.

  5. In “When the agent costs more than the person,” read the three tiles and the chart. The two bars are the electricity used at the same $3,300 of spend, where METR measured an agent’s progress matching a human researcher’s. Write down the ratio and the range. Then read caveats 2 and 3 together: why does this chart point the opposite direction from the one just above it, and what does that mean for how you’d advise a student deciding whether to hand a task to an agent?

  6. In Module C, note the right-sizing multiplier and the efficiency-over-time stat — then read the Jevons-paradox caveat next to it. In your own words: how can AI get 33x more efficient per query and Google’s total data-center electricity and water use still rise in the same period? (Both are true — this is the point.)

  7. Read Module E (“Why do the public numbers disagree so much?”) and Module F (“What we genuinely don’t know”). Pick one item from Module F and think about how you’d answer a student who asked you that question directly.

Expected result. You can state a defensible range (not a single number) for at least two AI tasks, explain one reason estimates diverge, and name one thing that genuinely isn’t known yet.

Check your work. If you find yourself quoting a single AI-footprint number with confidence in class or a paper, go back and check whether you have the range and the source — the range and the “why it varies” are the actual teachable content here, more than any one figure.

Common pitfalls. The tempting failure mode is picking whichever number (highest or lowest) supports the conclusion you already had. Use the mid-range as your default and cite the range when precision matters.

Stretch (optional). Using Module D’s aggregate figures, draft the two-sentence framing from Activity 2 again — does having the aggregate context (share of US electricity, direct vs. indirect water) change how you’d word it?

Activity 4 · Advanced — Estimate AI's electricity use from real data, and find out why your answer is wrong

Time ~35 min · Tools R or Stata · Data Epoch AI, AI Data Centers (CC-BY)

Goal. Produce your own bottom-up estimate of AI data-center electricity use, compare it with two published estimates, and diagnose the gap. The diagnosis is the point — this activity is designed so that your first answer is wrong by roughly a factor of three, for reasons that generalize to almost any capacity-based estimate you will ever make.

Why this dataset. Epoch AI tracks individual AI data centers — 78 sites, roughly 12.1 GW of current IT power capacity — built up from satellite imagery, construction permits, and company statements, and released under a Creative Commons Attribution license. It is the only open, facility-level, AI-specific dataset available. Note what it does not contain: any estimate of energy consumed, or any utilization assumption. Epoch stops at capacity deliberately, because the utilization assumption is where the disagreement actually lives.

Setup. Download the starter script for your language and run it. Both scripts print the column names first — check them against what the code expects before trusting any output, since Epoch revises the schema as the database grows.

  • R: epoch_ai_capacity.R
  • Stata: epoch_ai_capacity.do

Steps.

  1. Run the naive estimate. The script computes: tracked IT power ÷ Epoch’s stated coverage fraction × PUE × utilization × 8,760 hours. Record what you get (it should land around 280–440 TWh/year).

  2. Compare it with two published estimates for 2025.

    Source Estimate (2025) Method
    IEA, Key Questions on Energy and AI (2026) ~155 TWh Top-down; whole AI-focused facilities
    de Vries-Gao, Joule (2025) 46–82 TWh Bottom-up from chip-packaging supply; AI hardware only

    Both are estimates for 2025. Your capacity snapshot is from 2026. Hold that thought — it is the answer to step 3.

    Note that the two published figures differ from each other by a factor of three before you even add yours.

  3. Write down, before reading further, why you think your number is too high. Be specific — name a quantity and a direction.

  4. Now fix the two things that matter. First the time base: an August 2026 capacity snapshot describes 2026, not 2025, and in a fleet doubling annually the energy is the integral of the growth curve, not the endpoint times 8,760. Second the load factor: published fleet-wide effective load runs 0.4–0.6, not 0.6–0.8. The scripts do both, and report a 2026 estimate and a 2025 back-cast separately.

  5. Rank the four corrections by how much each one moves your answer. Change one parameter at a time from the naive values and record the effect. Then check your ranking against Module “Capacity is not consumption” in the Footprint Explorer.

  6. State the boundary problem. In one or two sentences, explain why the IEA figure being larger than de Vries-Gao’s does not mean either is wrong.

Expected result. A defensible band of roughly 120–250 TWh for 2026 and 55–130 TWh for 2025, a ranked list of what went wrong in the naive version, and a written statement of why your 2025 estimate should sit below the IEA’s 155 TWh rather than matching it.

Check your work. If your ranking puts “planned capacity is included in the total” at the top, re-read the Epoch documentation — sites under construction contribute zero to the current-capacity total. The dominant error is the time base: you annualized a snapshot taken today against a calendar year that has already ended, for a fleet that grows about 2× a year.

Common pitfalls.

  • Dividing a power figure by a compute-based coverage fraction. Epoch’s 27% is a share of delivered H100-equivalents, not a share of megawatts, and it carries a different date than the capacity snapshot. Both facts matter, and they push in opposite directions.
  • Treating nameplate capacity as average draw. Published fleet-wide effective load factors run 0.4–0.6; LBNL models inference fleets nearer 0.2.
  • Reporting a point estimate. Every input here is a range. Your output is a range.

Stretch (optional). The facility-level data has a column for location. Join it to your state or county of interest and ask a different question — not “how much electricity,” but “who lives near this, and what does the local grid look like?” That is the question the aggregate national figures cannot answer, and it is closer to what public health actually needs.

Citation. Epoch AI, “AI Data Centers.” Published online at epoch.ai. Retrieved from https://epoch.ai/data/ai-data-centers

Check your readiness

Answer these, then check — your score suggests whether to dive in or skim the readings first.

Useful resources

How these materials were built

The AI Environmental Footprint Explorer and the activities on this page were built by D. Alex Quistberg with Claude (Anthropic). Claude did the research synthesis, the code, and the testing; every figure was traced to a primary source and checked before it went in, and the editorial judgments — which numbers to trust, which to leave out, how to frame the disagreements — are the author’s.

Several figures were corrected during that process after checking against primary sources, including a widely repeated per-image estimate that turned out to be about fifteen times too high. That is why the changelog exists, and why nothing in the Explorer is presented without its source.

Worth naming on a page about AI’s costs: building this used AI, and that use had a footprint. It is a reasonable example of the right-sizing Activity 1 asks you to practise — and of the disclosure this track argues for.