AI & Environmental Impact

01 · Run it yourself — per-query energy is hardware arithmetic

You do not have to estimate what a query costs. Open-weight frontier-class models — Llama-3, Qwen-72B, DeepSeek-V3, Mistral — download from HuggingFace and run on servers you can rent by the hour or buy outright. The energy per query is then just physics:

Wh/query = (GPUs in node × rated TDP watts × seconds per query) ÷ (queries served concurrently) ÷ 3600 × PUE

Every term is knowable. GPU TDP is a published NVIDIA spec. The GPU count is fixed by the model's weight memory (params × 2 bytes for bf16, ÷2 for fp8/int8, ÷4 for 4-bit). Seconds-per-query and concurrency you read straight off nvidia-smi while vLLM serves. Here is that math for four real open models:

GPU TDP (A100-SXM 400 W, H100/H200-SXM 700 W) — NVIDIA published spec MEASURED · weight memory + throughput → Wh — standard vLLM serving arithmetic DERIVED · assumes ~500-token response, PUE ~1.1–1.2

02 · You are not the reason they build the data center

Here's a piece of the causal picture that individual-responsibility framing usually misses. H100s are not manufactured in response to demand you can place. The real bottleneck on accelerator supply is TSMC's CoWoS advanced-packaging line, and it's booked out by a handful of hyperscalers — Microsoft, Meta, Google, Amazon, xAI, Oracle — placing multi-billion-dollar orders on multi-year horizons. The data center gets built, the GPUs get bought, and the racks get powered on the basis of capital-allocation cycles and capacity planning, committed long before — and regardless of — whether you personally ask anything.

And once the racks are spinning, your query is inframarginal: it drops into batch capacity that is already running (§01). The dominant footprint of AI is the amortized fixed cost of the buildout — chip fab, construction, and the baseline power to keep committed capacity available — not the tiny dynamic increment of one more prompt filling a batch slot that was going to spin anyway. So attributing the buildout to any one person's queries points at the wrong node in the causal graph: the lever is capex and the vendor-pipeline forecasting that inflates it (§04), decided in boardrooms — not the individual deciding whether to hit enter.

Here is the chain that actually produces a running AI data center. Follow where the constraint is — and where you enter it:

The scale of what's committed before any of us hits enter:

Big-4 (Amazon, Alphabet, Meta, Microsoft) combined capex, planned/actual VERIFIED · CoWoS ~90% share + "sold out … into 2026" (TSMC C.C. Wei) VERIFIED · B200 208B transistors VERIFIED · interconnection queues 5–10 yr VERIFIED

"All the hyperscalers report that their markets are supply-constrained, rather than demand-constrained." — industry capex analysis, 2026. The thing holding AI back isn't too few users asking; it's too little packaging capacity and power. That is the opposite of a story where your individual query is the driver.

03 · "Data centers" ≠ "AI" — stop conflating them

Here is the sleight almost every headline runs on. The famous anchor — "data centers are ~1–1.5% of global electricity" — is all data centers: streaming, cloud storage, email, corporate IT, and crypto. AI is a slice of that slice. How big a slice is genuinely unresolved: nobody meters AI separately. No operator discloses AI vs non-AI electricity — Google's last such disclosure was 2022 — and the IEA publishes AI-focused data centres only as a growth rate, never as a standalone total. The best available estimate put AI at 10–15% of data-centre energy in 2023, and the IEA says AI-focused data centres grew 50% in 2025 against 17% for data centres overall. Roll that forward and AI lands around 15–30% of data-centre electricity in 2025. Do the multiplication:

Whenever you see "data centers will double by 2026 to ~1,000 TWh," note what's inside: the IEA bundles data centers + AI + crypto into one figure and provides no clean AI-only carve-out. Crypto alone was ~110 TWh in 2022 (Bitcoin ~120–200 TWh by itself) — larger than the AI-specific slice at the time. Reporting the whole bundle as "AI's appetite" annexes a crypto load the size of Thailand's national grid.

04 · Where the operational numbers actually land (and why the scary ones are shaky)

Disambiguated honestly: data-center electricity is growing fast — the IEA projects it "more than doubles by 2030 to around 945 TWh" (≈ Japan's total), and in the US data centers may drive nearly half of demand growth to 2030. That's a real grid-planning and interconnection problem. But the truly alarming numbers are 2030 extrapolations built on vendor sales pipelines, not measurements.

Audited: nothing. Not one AI-specific energy figure anywhere is third-party assured. The hyperscalers do publish assurance letters — but they cover total-company emissions, one aggregate energy line and water, and explicitly disclaim the rest of the report, which is precisely where the AI numbers sit. Microsoft's per-datacentre electricity table is in the unassured section; NVIDIA's assured Scope 3 stops one category short of the GPUs it sells. No mandatory regime forces an AI carve-out either — not CSRD, not California SB 253, not the EU data-centre scheme. When someone quotes you a precise figure for AI's electricity, no auditor has stood behind it.

And there is a structural reason it stays this way. Meters, industry codes and interconnection contracts have no workload dimension. Utilities segment by customer, statistics offices by industry code, grid operators by connection size — none of which knows whether a rack is training a model or serving email. So the metered national figures that do exist are for data centres, and every AI-specific number is necessarily derived outside official statistics. That is a permanent feature of how the counting works, not a gap someone will close next year.

Goldman's viral "+165% by 2030" is a compound extrapolation off 2023, and energy economist Jonathan Koomey names the flaw: "usually what they're assuming is NVIDIA's business plan is our growth forecast." The precedent is the dot-com scare — analysts claimed computing would hit 13% of all electricity; the real number came in at 3%. And the boxes are not consistent even within one institution: two 2024 IEA reports give 2022 figures that differ by roughly half, because one bundles crypto in and the other leaves it out. That is a scope difference, not an error — but it means you cannot compare two data-centre numbers until you know what each one counts. Treat the doubling/165%/21% figures as upper-bound scenarios with a stated assumption stack, not facts. The "21% of electricity by 2030" line was never primary-sourced — it's a projection, not a measurement, so it isn't a data point at all; treat it as one scenario's assumption, not evidence.

05 · Operator economics: batching & PUE are the utilization levers

It's easy to miss that electricity is the operator's dominant operating cost — so the incentive to raise utilization points the same way as the environment. Two independent levers, both about not wasting watts:

Both directions, honestly. The 33× is one workload (text prompt), software-driven, and Google says it does "not represent … all prompts nor … future performance"; the 44× carbon bundles grid decarbonization with compute. The 0.24 vs 0.10 Wh disclosure is a vendor publishing the less flattering number — conceding most public estimates undercount true at-scale load by ~2.4×. But Google is an interested party, and independent critics (e.g. de Vries-Gao, via Data Center Dynamics) push back on the boundary: it's a median (heavier long-context prompts cost more), the water counts only on-site cooling (not the water to generate the electricity), the carbon uses "market-based" accounting (which clean-energy purchases flatter), and it isn't peer-reviewed. The internal math is sound; the scope is chosen.

06 · Small / local / MoE bend the curve — but Jevons bends it back

Capacity forecasts hard-code 2025's model size, cost, and cloud-centric inference — every one of which is dropping fast. DeepSeek-V3 hit frontier-competitive quality with only 37B of its 671B params active per token (MoE), so per-token compute — and energy — is closer to a ~40B dense model despite the big memory footprint. Its final training run was 2.788M H800-GPU-hours ≈ $5.576M. Epoch measures fixed-performance inference price falling a median ~50×/yr (200×/yr since 2024), and frontier quality now lands on a sub-$2,500 consumer GPU within 6–12 months — pushing real inference off the cloud for models up to ~40B params.

07 · Water: measurable, engineerable, and in proportion

Same discipline as energy — count the actual liters, and set them in proportion. Per-query water is not one number; it's a boundary choice. Two credible 2025 disclosures sit ~170× apart because one counts on-site cooling only and the other counts full life-cycle (incl. electricity generation).

And in proportion: globally, agriculture is 69% of freshwater withdrawals; industry (which merely contains data-center cooling) is 19%; municipal 12%. In the US, thermoelectric power-plant cooling (133 Bgal/day) and irrigation (118 Bgal/day) so dominate that data centers don't register as a national USGS category. The viral "500 ml per chat" traces to one 2023 GPT-3 estimate (Li et al.) whose own authors give a 5× range ("10–50 responses") and stress location-dependence. Critically, evaporative cooling is an engineering choice, not a law of physics: Microsoft's closed-loop chip cooling recirculates water and claims to eliminate it entirely (>125M L/yr avoided per site). The honest ask is WUE disclosure and closed-loop retrofits in genuinely stressed basins — not treating inference as a top-tier water villain when alfalfa and once-through power cooling dwarf it.

08 · A sense of proportion

None of this means the impact is zero, or that caring is foolish. It means the sizes are worth getting right before we assign guilt. Three proportions, all built from numbers already on this page.

Water is not "deleted from the universe"

The mental image behind the panic — a bottle of water consumed by your prompt and gone forever — is physically wrong. Cooling withdraws water; most is returned, and the fraction that evaporates re-enters the water cycle as precipitation. Water is not destroyed; it is borrowed, and some is relocated. The real, legitimate concern is local and temporal scarcity — a specific aquifer in a specific drought-stressed basin at a specific time — not a cosmic ledger where the universe's water is ticking down. Withdrawal ≠ consumption ≠ destruction, and collapsing all three is how "0.26 ml of cooling" becomes "you killed a bottle of water."

The conversation about it is bigger than the query

Both of these are measured numbers already cited on this page. A median Gemini text prompt emits ~0.03 gCO₂e (Google's metered per-prompt disclosure, alongside its 0.24 Wh / 0.26 ml). Streaming 30 minutes of video emits ~18 gCO₂e (IEA streaming fact-check, §11). So:

A 30-minute video explaining AI's water crisis costs roughly 600 prompts worth of carbon; the time spent reading and writing about it online adds more. That's not an argument against making the video or having the conversation — both are worth it. It's a reminder that ordinary digital actions, an AI query among them, are all individually tiny, so singling one out for guilt doesn't track the actual sizes.

Where the leverage actually is: alfalfa

AI's entire global water footprint is a fraction of a percent of agriculture's (§07). In the drought-stressed US West, alfalfa — much of it grown as cattle feed for export — is among the single thirstiest crops. Trimming a small percentage of alfalfa irrigation would free more freshwater than the entire AI sector uses. VERIFIED: in the Colorado River Basin, cattle-feed crops (alfalfa + hay) consume ~46–55% of the water — nearly two-thirds of all agricultural use (peer-reviewed; LA Times / E&E) the specific "1% of alfalfa > all of AI" ratio is illustrative, not a fetched figure. Again, the point isn't that AI's water use is nothing — it's that if we're serious about freshwater, the highest-leverage targets are alfalfa, thermoelectric cooling, and the capex-driven buildout (§02), and they deserve more of our attention than any individual's chatbot use.

09 · What nobody counts, part 1: waste — and the system already wastes most of it

Start with the context that reframes the whole argument: the US energy system already throws away most of the energy it produces. The long-documented Lawrence Livermore energy flow chart shows roughly two-thirds of US primary energy is "rejected" — lost as waste heat in generation and conversion before it does any useful work — with another ~5% of electricity lost in transmission and distribution. AI's marginal draw lands inside a system already running at maybe a third of its theoretical efficiency.

Combine that with §02: the buildout is capex-driven and inelastic to your individual query, and it sits atop enormous pre-existing waste. Focusing on an individual's prompts optimizes a rounding error while the system rejects two-thirds of everything it makes. The real lever is systemic efficiency — which, as it happens, is one of the things AI is measurably good at (§10). Then there's the AI-specific waste the per-query and operational-TWh math both skip:

• Embodied carbon & hardware churn. Manufacturing a GPU — TSMC fab energy, HBM stacks, CoWoS packaging — carries a large embodied footprint that no per-query number includes. Accelerators churn on a ~2–3-year cycle (A100 → H100 → H200 → B200); the operational-energy story quietly assumes the hardware just exists.
• E-waste. Decommissioned accelerators, servers, and networking gear — most not recycled, some containing materials whose extraction had its own upstream water and energy cost.
• Speculative overbuild & stranded capacity. The vendor-pipeline forecasting that inflates demand estimates (§04) drives capex ahead of real load. Data centers built on speculation sit underutilized — idle GPUs still draw power and still embody all that manufacturing carbon for little useful work.
• Discarded and redundant training. DeepSeek's own paper says its $5.576M figure excludes "prior research and ablation experiments." Failed runs, models trained and never shipped, and five labs each training a near-identical frontier model are duplicated footprint that never shows up in a per-query stat.
• Wasted inference. Agentic retry loops, needless long-context reprocessing, and reaching for a frontier model where a 7B — or no model — would do. The efficiency gains of §06 are partly eaten here.

None of this has a clean, disclosed number — the conversation fixates on the one quantity with a good meter (operational Wh) and skips the ones that may dominate lifetime impact.

10 · What nobody counts, part 2: net — and here the numbers finally exist

Every figure on this page — and in nearly all the coverage — is gross: energy and water going in. The net (gross draw minus what AI displaces or enables) is almost never computed. But it isn't unknowable — the IEA's 2025 "Energy and AI" report put real numbers on the credit side, and they dwarf the debit.

The IEA net ledger

The debit is real but bounded: data-centre electricity emissions grow from 180 Mt CO₂ today to 300 Mt (Base Case) or up to 500 Mt (Lift-Off) by 2035. The credit the IEA models is far larger — emissions reductions from broad application of existing AI solutions could equal ~5% of energy-related emissions in 2035. In the report's own words, the data-centre increase "could potentially be offset by emissions reductions enabled by AI if adoption is widespread."

The load-bearing caveat, out loud. The credit is potential and conditional — it assumes widespread deployment of efficiency tools that mostly haven't scaled — while the data-centre growth is near-certain. The IEA says "could potentially," not "will." This is a plausible net-positive path, not a measured net.

Where the savings come from — IEA "widespread adoption" scenario

Already deployed, already measured

Not all of it is projection. DeepMind's control system cut the energy used for cooling Google's data centres by 40% (≈15% of total facility PUE overhead); a later autonomous version reached ~30% cooling savings. And GenCast produces a 15-day weather ensemble in 8 minutes on a single TPU, versus "hours on a supercomputer with tens of thousands of processors" for the physics-based equivalent — a large compute reduction on a task run constantly.

Read the credit side carefully — sort each figure by what it actually measures

The honest failure mode here is the one this page criticizes everywhere else: quoting a flashy percentage as if it were an emissions cut when it measures something different. Sorted by evidence, the credit side looks like this:

And every one of these sits against the Jevons rebound: peer-reviewed work warns that "just because an AI model becomes more efficient … the inverse effect [more total consumption] is highly plausible." A real credit side — read honestly, tier by tier — is still not a green halo.

The honest statement, updated The sign of AI's net effect still isn't formally closed — no one has completed the full ledger (gross draw + embodied + waste − displaced − enabled, net of Jevons rebound). But "we have no idea" is no longer honest either. The credit side now has real, sourced figures plausibly an order of magnitude larger than the data-centre debit — if the efficiency tools actually scale. The defensible claim was never "AI is green." It's that the gross number everyone quotes is the small, easy half of a ledger whose other half is finally being measured.

11 · Why the discourse is broken — and the harms that are actually real

Why is a ~0.2%-of-electricity load reported like an apocalypse? Because the alarming numbers are thinly sourced and corrections don't travel. The "training one model = five cars' lifetimes" headline (MIT Technology Review's phrasing, not the paper's) came from an atypical neural-architecture-search run: Strubell et al. 2019 reported 626,155 lbs CO₂e for that search — ÷ 126,000 lbs for a car's lifetime ≈ 5 — while the same paper's normal BERT training was just 1,438 lbs. Google's Patterson et al. then recalculated the search on the actual hardware and found the original 88× too high (their verbatim "reduce by 88X"; ~18.7× for an average operator, 88× for an efficient one like Google — Strubell had assumed the search ran on full-size tasks when it used small proxy tasks). The 2019 "Netflix is boiling the planet" panic was likewise corrected by the IEA ~90× lower. Same structure as the dot-com "a lump of coal per book ordered" scare — off by orders of magnitude, every time.

The harms that are concrete are local and infrastructural, not per-prompt:

• Localized grid strain & cost-shifting — a gigawatt campus on a regional grid can raise capacity costs for existing ratepayers and delay other interconnections. National averages hide this entirely.
• Air emissions & environmental justice — xAI's "Colossus" in South Memphis, beside the historically Black Boxtown neighborhood, ran up to 35 gas turbines (per SELC aerial imagery) largely without Clean Air Act permits (~1,200– 2,000 t/yr NOₓ). NAACP / SELC / Earthjustice litigated. The harm is air from on-site turbines, not water — the opposite of the viral framing.
• Water in already-stressed basins — sector proportion doesn't help the specific aquifer next to a specific facility in Arizona or Spain. Proportion is an argument about priorities, not a free pass on siting.
• Embodied & waste (§07) and the uncomputed net (§08) — the parts most likely to dominate, least likely to be measured.

Sources

• Luccioni et al., "Power Hungry Processing" — metered open-model energy, A100
• ML.Energy Leaderboard — live measured per-model energy on A100/H100
• NVIDIA A100 / H100 — published TDP specs
• vLLM — continuous batching (the utilization lever)
• DeepSeek-V3 technical report — 37B/671B MoE; $5.576M final run (excl. R&D)
• Epoch AI (2025) — ~0.3 Wh estimate; its "10% utilization" assumption
• Google (2025) — measured 0.24 Wh / 0.26 ml; 33×/44×; 0.10 vs 0.24
• Google efficiency — fleet PUE 1.09; >6× compute per unit electricity
• IEA "Energy and AI" (2025) — 945 TWh by 2030 (≈ Japan)
• IEA Electricity 2024 — the DC+AI+crypto bundle
• Goldman Sachs (Feb 2025) — AI 14%→27% of DC load mix in GW (not energy); +165% scenario
• IEA "Key Questions on Energy and AI" (2026) — 485 TWh in 2025; AI-focused DC +50% vs +17% all DC
• IEA-4E/EDNA (2025) — AI 10–15% of data-centre energy in 2023
• Koomey, Catalyst — "NVIDIA's business plan"; dot-com 13% vs 3%
• GiovanH — opacity, inference≪training
• FAO/AQUASTAT — agriculture 69% / industry 19% / municipal 12%
• LA Times / Richter et al. — cattle-feed crops ~46–55% of Colorado River water
• FoodPrint — beef ~1,800 gal/lb water footprint
• USGS — thermoelectric 133 + irrigation 118 Bgal/day
• Microsoft (2024) — closed-loop, >125M L/yr avoided
• Li et al. (2023) — "Making AI Less Thirsty" (the over-cited GPT-3 source)
• Mistral (2025) — peer-reviewed LCA, ~45 ml/response
• Patterson et al. (2021) — recalculated NAS emissions 88× lower
• Center for Data Innovation (Castro) — Strubell debunk; Google ML flat-share
• Bender & Hanna, The AI Con — framing critique
• Alexander Avila — "AI Wars" bibliography — compiled skeptical source set
• NAACP / SELC / Earthjustice — xAI Memphis turbines
• BrainBox AI (Cammeby's) — HVAC −15.8% (measured, single building)
• Centre for Net Zero / U. Chicago BFI — EV charging RCT, 42% peak / 0% total
• Atmos. Chem. Phys. (2023) — ML methane super-emitter detection
• DeepMind — wind forecasting +20% value
• Siemens (IEEE 2023) — cement kiln −2.6% specific heat
• Microsoft — AI battery-material screening (enabling)
• Big-Tech AI capex tracker — $226B→$410B→$725B (2024–26)
• TSMC CoWoS bottleneck — ~90% share, "sold out into 2026"
• Grid interconnection queues — 5–10 yr, ~20% of projects at risk
• Nvidia customer concentration — top-2 ≈ 36% of revenue
• IEA Global Energy Review 2025 — global electricity demand +1,100 TWh (+4.3%) in 2024
• Andy Masley — "The AI water issue is fake"
• Ars Technica — AI water "a drop in the bucket"
• Hannah Ritchie — ChatGPT carbon footprint, "very small"
• MIT Tech Review — "Stop worrying about your AI footprint"
• BBC / de Vries 2023 — "as much energy as the Netherlands"
• Euronews / Li et al. 2023 — "bottle of water" headline

Part of the AI Problems Index · see the Research Library and Risk Atlas. Hardware figures are published TDP × standard serving arithmetic — reproducible on rented or owned GPUs. Ranges, not points.