← AI investigation
Act II · the actual footprint

How big is AI's actual environmental footprint?

Most coverage flattens this into one number, which produces both alarmist exaggeration (“ChatGPT is destroying the planet”) and dismissive hand-waving (“it's nothing”). The honest picture requires three lenses.

Read in 30 seconds
Per query · small

~0.24 Wh per Gemini text prompt

Same hierarchy as a Google search

Heavy chatbot user's annual footprint ≈ 1.1 kg CO₂ — vs ~624 kg/yr for one steak/week

"AI" is many things

5 orders of magnitude per query

Phone NPU vs cloud reasoning

BERT classifier 0.0001 Wh · Llama 8B on Mac 0.04 Wh · cloud reasoning 7 Wh

Real concern · system

$680B 2026 capex vs $40–60B revenue

Risk socialized to non-shareholders

PJM auction +833% YoY · 80 GW gas turbines to 2029 · 14-yr ratepayer contracts

Per query, AI is small (green). 'AI' is not one thing — five orders of magnitude across the model spectrum (violet). The real concern is the financial/grid system being built on top of demand that hasn't fully materialized (amber).

00 · The three-lens thesis

Three claims, only one of which is being publicly debated.

(1) Per-query AI is in the same hierarchy as a Google search. (2) 'AI' is not one thing — the spectrum spans 5 orders of magnitude. (3) The real concern is system-level: capex, grid mortgages, financialization of risk.

The popular narrative collapses these layers. A heavy chatbot user's annual footprint, the marginal energy of a phone-resident model, and the$680B-per-year infrastructure mortgage are different conversations with different policy implications. This page separates them.

The numbers below are anchored on Google's August 2025 methodology paper — the most rigorous public per-prompt disclosure to date — alongside the IEA Energy and AI report, Hugging Face's AI Energy Score, and primary hyperscaler ESG filings.

01 · Per query — fairly compared

What does AI usage actually look like next to other things you do?

Slide your AI usage and watch where it lands on the same log-axis ladder as Netflix, AC, beef, and flights. For a typical chatbot user it sits well below a phone charge.

Your AI footprint vs. everything else (log scale)
30
At 30 text prompts/day you use 7.2 Wh per day on AI. Compare with 77.0 Wh for one Netflix-hour, or 28.40 kWh for an average US household per day.
AIOther digitalHouseholdTransportFood (CO₂-eq)
↔ scroll horizontally
1.0 Wh10.0 Wh100.0 Wh1.00 kWh10.00 kWh100 kWh1000 kWhOne Gemini / ChatGPT text prompt0.24 WhOne Google search0.30 WhOne AI-generated image1.5 WhYOUR daily AI usage7.2 Wh1 hour Spotify10.0 WhSmartphone full charge19.0 WhBoil one kettle (1 mug)19.0 Wh1 hour TikTok40.0 Wh1 hour Netflix HD77.0 Wh1 hour YouTube HD120.0 WhMicrowave 10 min200.0 Wh1 hour PS5 gaming210.0 Wh1 mile in an EV350.0 Wh1 hour central AC3.50 kWhAverage US household daily total28.40 kWh1 kg beef (CO₂-equivalent)150 kWh1 transatlantic flight (per pax, CO₂-eq)6000 kWhFor an average chatbot user, AI is roughly the magnitude of a few Google searches per day.
Log scale spans 8 orders of magnitude. AI bars (violet) cluster in the small end. Even a heavy reasoning-model user (~100 reasoning prompts/day = ~0.7 kWh) sits below an hour of central AC. Weekly steak ≈ ~150 kWh CO₂-equivalent; one flight ≈ 6,000 kWh CO₂-equivalent. That said: per-query small does not mean system-level small — see the next figure.
Median Gemini text prompt: 0.24 Wh / 0.03 g CO₂e / 0.26 mL water (Google's full-stack methodology, May 2025 data).
The widely-cited 'ChatGPT uses a bottle of water per query' figure is the worst case, routinely misquoted.
02 · Many models, not one

"An AI query" is a category error.

A fine-tuned BERT classifier on a laptop and a frontier reasoning model in the cloud differ by ~5 orders of magnitude per task. Treating 'AI energy' as one number hides where the actual cost lives.

The model spectrum — five orders of magnitude per query
SpecializedOn-deviceLocal workstationCloud frontier (text)Cloud frontier (reasoning)0.10 mWh1.0 mWh10.0 mWh100.0 mWh1.00 Wh10.00 WhBERT classifier (110M)on laptop CPU0.10 mWhWhisper Tiny (1 min audio)on any phone1.0 mWhApple Intelligence (~3B)on iPhone Neural Engine2.0 mWhLlama 3.2 3Bon Snapdragon X NPU4.0 mWhPhi Silica (~3.8B)on Snapdragon X NPU4.8 mWhLlama 3.1 8B (Q4)on M-series Mac40.0 mWhLlama 3.1 70B (Q4)on M4 Max MacBook240.0 mWhMedian Gemini text prompton Google TPU + DC overhead240.0 mWhGPT-4o-class chat queryon cloud GPU300.0 mWhLlama 3.1 70B (Q4)on RTX 4090480.0 mWhFrontier reasoning (o-class)on cloud GPU + thinking budget7.00 Wh“AI energy” treated as one number conflates a phone NPU and a cloud reasoning model.
The same word — “AI” — covers everything from a 110M-parameter BERT classifier on a laptop (~0.1 mWh) to a frontier reasoning model with thinking budget (~7 Wh). That is roughly the energy ratio between switching on an LED for a second and boiling half a kettle. Treating “AI energy” as one number conflates the two. Local 8B-class inference on hardware you already own is ~6× lower than a cloud Gemini median query; phone NPU inference is ~100× lower. The energy-intensity-per-useful-operation is also collapsing fast — Google reports 33× per-prompt energy reduction in 12 months (May 2024 → May 2025).
Local Llama 8B on a MacBook is ~6× lower energy than a cloud Gemini median query — and the laptop has no PUE overhead.
Specialized small models beat general-purpose LLMs by ~30× for the same task.
03 · System level — where the bill comes due

Per-query small does not mean system-level small.

$680B in 2026 hyperscaler capex against $40–60B in generative AI revenue. The contracts being signed run 14–20 years — past the credible visibility horizon for AI demand. Equity takes the first loss; ratepayers, taxpayers, and pensioners take the rest.

Who pays — capital → infrastructure → loss surface
CAPITAL POURED ININFRASTRUCTURE MORTGAGEDWHO ABSORBS THE LOSSHyperscaler 2026 capex~$680B · 75% AI-attributedStargate$500B / 10 GW announcedSovereign wealthUAE MGX · Saudi PIF · ~$200B+VC mega-rounds~$84B in 2025Gas turbine backlog80 GW to 2029 · 5–7 yr leadNuclear PPAsTMI restart · Talen $18B · SMRsGrid capacity / PJMauction +833% YoYData centers in stressed regions2/3 of new builds since 2022Coal retirements delayedGA / IN / WY pausedEquity holders (first loss)if AI demand undershootsRatepayersVA: +$11.24/mo residentialTaxpayersDOE $1B TMI loan · $1.6B/yr abatementsPensioners (passive index)Mag 7 = 33.7% of S&P 500Equity holders take the first loss. Ratepayers, taxpayers, and pensioners absorb the rest.
The visual structure mirrors the doom and hype money flows (Act IV) — same shape, different accent. Where those traced influence, this one traces physical infrastructure mortgages and the people on the long-tail end of them. The contracts in the middle column run 14 years (Dominion GS-5), 17 years (Talen–Amazon Susquehanna), and 20 years (Microsoft–TMI) — past the credible visibility horizon for AI demand.
The PJM (mid-Atlantic) wholesale capacity auction cleared at +833% year-over-year. Data centers were 40% of that cost burden.
GE Vernova ended 2025 with an 80 GW gas-turbine backlog stretching to 2029, with hyperscaler volume agreements being negotiated out to 2035.
Bloomberg analysis (2025): more than two-thirds of new data centers built since 2022 are in water-stressed regions.
04 · Jevons paradox

Per-query efficiency is collapsing. Total demand is exploding faster.

Google reports 33× per-prompt energy reduction in 12 months. Microsoft's Nadella explicitly invoked Jevons after DeepSeek: cheaper inference creates more inference. Both can be true.

Algorithmic efficiency improves ~3× per year (Epoch AI: pre-training compute efficiency doubles every ~7.6 months). Add quantization, mixture-of-experts, and hardware generations — and you compound to the 33×/year Google has measured at the serving layer.

But cheaper inference enables agentic loops, always-on assistants, AI Overviews on every Google search, Copilot in every Office document, Meta AI in every WhatsApp thread. Activities that used to consume zero AI energy now consume some. The 2025 ACM FAccT paper From Efficiency Gains to Rebound Effects argues this is the dominant dynamic.

The cleanest framing: AI energy intensity per useful operation is collapsing fast, while total AI energy demand is exploding because cheap inference creates new demand. Both can be true, and conflating them — treating “an AI query” as one thing — is the category error driving most of the public confusion.

05 · What gets understated

Where the per-query numbers systematically miss.

Operational numbers (joules per query) leave out training amortization, embodied chip carbon, water in PPA-supplied power, induced demand, and local grid stress. None of these are catastrophic alone — together they're the gap between the headline figure and the honest one.

  1. 01

    Training amortization.

    GPT-4 training estimated at 50–60 GWh and 1,000–15,000 t CO₂. Amortized across billions of inferences this is small per query, but each new frontier release resets the clock.

  2. 02

    Embodied chip carbon.

    Manufacturing one H100 emits ~150 kg CO₂ (TSMC LCA estimates). A 100,000-H100 data center embeds ~15,000 t CO₂ before turning on.

  3. 03

    Carbon-accounting gap.

    Hyperscalers buy renewable PPAs but the local grid still burns gas; UCR's higher water numbers reflect the marginal grid, not the contractual one. Both are "true" depending on accounting frame.

  4. 04

    Induced demand.

    AI Overviews on every search, Copilot in every doc, AI in every WhatsApp thread — converts previously-zero-energy actions into LLM calls. Per-query gains can be wiped 100× by ubiquity.

  5. 05

    Local grid stress.

    ~1% of global electricity sounds small, but data centers cluster (Northern Virginia, Phoenix, Dublin, Loudoun County). Global averages hide the local problem.

See also
Sibling
  • Act I · Definitions
  • Act III · Tracking
Maintenance

Last verified against Google AI energy disclosure (Aug 2025), Hugging Face AI Energy Score, IEA Energy and AI report, BloombergNEF + Goldman Sachs research, PJM capacity auction filings, IEEFA energy infrastructure reports, and EPA / EIA / DOE official data on 2026-04-25.

Earliest expected staleness: hyperscaler ESG reports + IEA Energy & AI (annual updates) + Google AI energy disclosure (annual).

This investigation is treated as a living document. The claims marked stale in the receipt cards above are the ones to re-check first. How this is maintained →