In August 2025, Google published the most-cited corporate disclosure on AI's environmental impact: a research paper and blog post by Amin Vahdat (SVP…

In August 2025, Google published the most-cited corporate disclosure on AI's environmental impact: a research paper and blog post by Amin Vahdat (SVP, Chief Technologist for AI and Infrastructure) and Jeff Dean (Chief Scientist, Google DeepMind and Google Research) measuring the energy, water, and carbon cost of a single Gemini text prompt. Their headline numbers: 0.24 watt-hours of energy, 0.03 grams of CO₂-equivalent, and 0.26 millilitres of water per median prompt. Google's framing called this "less than nine seconds of watching TV." They also reported that per-prompt energy had dropped 33× over the prior twelve months, and total carbon 44×. > GOBLIN FACTS — median is a choice. Google's Gemini paper foregrounded a median text prompt at 0.24 Wh and 0.26 mL of water. Median inference is useful; it is not the full footprint of training, hardware, image/video generation, or aggregate demand. Five months earlier, in January 2025, three researchers (Alexandra Sasha Luccioni of Hugging Face, Montréal; Emma Strubell of Carnegie Mellon; and Kate Crawford of Microsoft Research / USC) had submitted a paper to the 2025 ACM Conference on Fairness, Accountability, and Transparency titled "From Efficiency Gains to Rebound Effects: The Problem of Jevons' paradox in AI's Polarized Environmental Debate." The paper made the argument that per-query efficiency improvements in AI are systematically reinvested into more AI use, larger models, broader deployment, and induced demand in other sectors. The net effect on AI's total environmental footprint is increase, not decrease. Both papers are technically rigorous. Both are written by credible authors. They are not contradicting each other. They are measuring different things, at different scales, and a clear read of AI's environmental impact means holding both at once. This chapter does that work. You'll leave with: a working model of training versus inference environmental costs; a clear understanding of what Google's measurement actually measures and what it doesn't; Jevons' paradox as a portable analytical tool you can apply to any efficiency claim; the Canadian-specific carbon arithmetic of provincial grid choices; the Tomlinson et al. "AI emits less than humans" finding read carefully rather than dismissed; and the structural finding, from Luccioni et al.'s citation of Owen et al. (2023), that connects AI's environmental footprint to Indigenous-lands extraction in a way that puts Chapters 8, 9, 11, and 13 on the same underlying geography. ---

Training versus inference — the most important environmental distinction

If you remember one technical distinction from this chapter, make it the difference between training and inference. The environmental conversation routinely collapses them, and the collapsing produces both the alarmist-overstatements and the corporate-understatements that dominate public discussion.

Training is the (very expensive, very energy-intensive) one-time process of building an AI model from data. For a large language model (GPT-4, Claude, Gemini, Cohere's Command), training involves running specialized computing hardware (mostly NVIDIA GPUs, as we covered in Chapter 6) for weeks to months, processing trillions of tokens of text, while consuming continuous electricity on the order of tens of megawatts for the duration of the training run. The total electricity bill for training a single frontier large language model has been estimated at tens of millions of dollars worth of energy, depending on the model size and the duration. Training GPT-4 reportedly cost OpenAI on the order of US$100 million in compute. Training Gemini Ultra likely cost Google a comparable amount.

Inference is what happens every time a user types a prompt and gets a response. Each individual inference call is small: Google's measurement of 0.24 watt-hours per median Gemini text prompt is in the same order of magnitude as charging a smartphone for a few minutes. But inference happens at massive scale. ChatGPT receives roughly a billion queries per day. Google Gemini, Claude, and other commercial services together likely receive several billion daily queries. At billion-per-day scale, even tiny per-query numbers aggregate to substantial total energy.

Here is the arithmetic, with conservative assumptions. A billion median-equivalent prompts per day at 0.24 Wh each equals 240 megawatt-hours per day — about 87,600 megawatt-hours per year for one billion daily prompts. The major model providers serve aggregate billions of daily queries combined. Inference at scale is in the same order of magnitude as training for any major model, and growing faster, because deployment is growing faster than new training runs.

ALIGNMENT — which number is this? Before trusting any AI energy or water figure, ask two things: training or inference, and one query or the whole fleet? A reassuring per-prompt number and an alarming national-grid number can both be honest — they are measuring different things. A lot of environmental arguments are two people holding opposite ends of that stick.

This matters for two specific reasons.

First: when corporate sustainability reports focus on per-query efficiency improvements (as Google's August 2025 paper does), they are reporting the metric that most easily improves through technical optimization. The per-query metric will continue to drop. The total inference energy use is growing despite the per-query metric dropping. The per-query win is real and the aggregate rise is real, and the guide asks readers to keep both in view rather than letting one cancel the other.

Second: when academic critiques focus on aggregate AI environmental footprint, they are correct that the total is growing, and they are sometimes vulnerable to the corporate counter-argument that "we are working hard on efficiency" because the per-query metric is, technically, improving. The argument that lands is not "Google is lying" or "AI is unsalvageably wasteful." The argument that lands is Jevons' paradox. The rest of the chapter is built around it.

---

Jevons' paradox, properly understood

In 1865, the English economist William Stanley Jevons published The Coal Question, in which he made an observation that has been variously confirmed across industries ever since: improvements in the efficiency with which a resource is used often increase, rather than decrease, the total consumption of that resource. Jevons was specifically observing that James Watt's steam engine, despite being dramatically more efficient than its predecessors, had led to more total coal consumption in Britain, not less. The reason: efficiency made coal-powered work cheaper per unit, which made coal-powered work attractive in more applications, which expanded the market for coal-powered work, which increased total coal demand faster than efficiency could reduce it.

This is now called Jevons' paradox in the formal economics literature, the rebound effect in environmental economics, and induced demand in transportation planning. The same structural phenomenon shows up across resource economies. Adding highway lanes doesn't reduce congestion; it induces more driving. More efficient lighting hasn't reduced total electricity used for lighting, because it expanded where lighting goes. And more efficient agriculture hasn't reduced total farmland: it expanded production instead.

EXAMPLE — the efficient furnace, the bigger house. When furnaces got twice as efficient, people didn't halve their heating bills — they heated bigger houses. That is Jevons' paradox, and it is why "each AI query uses less energy now" can be completely true while AI's total energy use keeps climbing. Efficiency reads as an invitation to use more.

🧌 GOBLIN CHECK — Jevons' paradox, goblin edition: making a thing cheaper per use has never once, in the recorded history of industrial civilization, made people use less of it in total. Watt's engine was the efficient one. Britain did not — you may have noticed — burn less coal. When a company tells you its per-prompt number went down, the goblin's question is not "is that true?" It's "and how many more prompts are there now?"

Luccioni, Strubell, and Crawford's 2025 paper applies this framework specifically to AI. Their argument, boiled down:

Per-query AI efficiency is improving. The total environmental footprint of AI is growing faster than efficiency is improving. The growth is being driven by exactly the mechanism Jevons identified: cheaper-per-query AI is being deployed in far more contexts, at far larger scale, with much more inference per use, plus induced demand in adjacent sectors.

The paper distinguishes eight categories of indirect effects: different ways that AI efficiency gains translate into expanded total consumption rather than reduced total consumption. The full taxonomy is technical, but the key categories worth carrying with you:

Substitution effects. AI-generated images replace photographs (which had a smaller resource footprint per image but were used in smaller quantities). Each AI-generated image is more efficient than a Hollywood photoshoot, but the number of AI-generated images per day is vastly higher than the number of photographs that were being produced before.

Scale effects. Models keep getting bigger. The per-parameter efficiency is improving, but total parameters per major model has grown roughly 1000× since GPT-2 in 2019.

Direct economic rebound. Cheaper AI per query means AI is integrated into more workflows. Each integration is small; the integrations are now in the hundreds of millions.

Indirect economic rebound. Apple Intelligence runs only on the most recent iPhones. The AI features create economic pressure to upgrade hardware (with the embodied carbon of new device manufacturing) faster than would otherwise occur.

Economy-wide effects. AI is being deployed as a general-purpose technology, like the steam engine. The total impact will be felt across the entire economy, not just in the AI sector itself.

Induction effects. Targeted advertising is one of AI's most lucrative applications — the worldwide market for AI in marketing alone was valued at roughly US$36 billion in 2024, per a Statista estimate cited by Luccioni and colleagues. Targeted advertising induces consumption that would not otherwise occur. The carbon footprint of the induced consumption is enormous and is rarely counted as AI's footprint.

Time rebound. AI saves users time on cognitive tasks. That saved time is often spent on activities with higher carbon footprints (more travel, more shopping, more energy-using leisure).

This is the framework Chapter 1 introduced and the chapter that's been promising it ever since. It is the most important single analytical tool the environmental chapter offers, because it explains why the per-query efficiency story and the aggregate-growth story can both be true, and why "Google is becoming more efficient" cannot, on its own, be the argument for "AI's environmental impact is decreasing."

---

What Google's August 2025 paper actually shows

With Jevons' paradox in mind, the Google paper becomes more legible, both for what it does claim and for what it doesn't.

What Google measured. Energy consumption of a median Gemini text prompt across Google's data-centre fleet, in May 2025. The measurement includes the "comprehensive" methodology that Google explicitly contrasts with the simpler "active-machine-only" methodology used in most prior public estimates. The comprehensive methodology counts not only the energy used by the AI chip during inference, but also idle machine overhead, CPU/RAM usage, data centre cooling overhead (the PUE multiplier), and water consumption for cooling. Per the paper:

"Comprehensive methodology: 0.24 Wh / 0.03 gCO₂e / 0.26 mL water" "Active-machine-only methodology: 0.10 Wh"

The difference is significant. The comprehensive figure is roughly 2.4× higher than the simpler measurement. Google's framing of this is itself a finding worth flagging — they note that most prior public estimates of AI inference energy have used the simpler methodology, and therefore have underestimated the real footprint by roughly 2.4×. That is an unusually candid corporate admission, and the guide credits it.

The trajectory claim. Google reports that energy per median Gemini prompt dropped 33× between May 2024 and May 2025, and total carbon dropped 44× over the same period. The drivers attributed: full-stack efficiency improvements across model architecture (Transformer optimizations, Mixture-of-Experts, hybrid reasoning), algorithms (quantization, speculative decoding, distillation), custom TPU hardware (the Ironwood TPU claimed at 30× efficiency over first-generation TPU), idle-machine management, ML compiler stack improvements, and Google's fleet-wide PUE of 1.09 versus the global average of approximately 1.55. If accurate, these are real engineering achievements.

What the paper doesn't measure, by explicit scope choice.

Training is excluded. Embodied carbon (the carbon footprint of chip and server manufacturing, including the Owen et al. extraction footprint we'll return to) is excluded. End-of-life e-waste is excluded. Image generation is excluded. Video generation is excluded. Complex coding and agentic workloads are excluded. The measurement is median text prompts only. Heavy users, long prompts, image generation, video generation, and agent tasks can consume orders of magnitude more than the median. The 0.24 Wh figure is the friendliest measurable subset of inference, reported accurately within that scope.

Footnote 2 of the Google paper addresses verification directly: "The data and claims have not been verified by an independent third-party." Every number in the paper is corporate self-disclosure. The methodology is published and is technically defensible. The numbers may well be accurate within their stated scope. They are not independently verified, and Google explicitly says so.

On framing: "less than nine seconds of watching TV" is a deliberately friendly comparison. The choice of median (the friendliest statistic over mean or maximum), of inference (excluding training), of text (excluding image and video), and of TV-watching (a familiar low-stakes activity) is strategic. Honest technical content and strategic framing are not mutually exclusive; this disclosure is both.

Reading it whole: Google's August 2025 paper is the most methodologically careful inference-environmental-disclosure currently published by any major AI provider. The methodology is better than the field's prior standard, and Google deserves credit for that. The framing chosen to surround the methodology is strategic in ways that flatter the result. The scope is real but limited. The numbers cannot currently be verified by anyone outside Google. None of these readings cancels the others, and the guide's job is to give readers the tools to read corporate environmental disclosures with all of them in view.

---

The growth side of the ledger — what the aggregate numbers actually show

If per-query efficiency is improving and total environmental footprint is growing, the question is: by how much, and in what direction?

The Luccioni et al. paper cites the following independently-measured findings, which the guide carries forward as the empirical foundation of the aggregate-growth case:

Microsoft's global water consumption increased 34% between 2021 and 2022, topping 1.7 billion gallons. Google's increased 20% over the same period. These are corporate disclosures from Microsoft and Google, in their own annual sustainability reports. They reflect total water consumed, not per-query.

Google's greenhouse gas emissions have increased 48% since 2019, per Google's own 2024 Environmental Report, which explicitly attributes the increase to "increases in data centre energy consumption." Microsoft's GHG emissions have increased 29.1% since 2020. Baidu's increased 32.6% over a 2021 baseline. These are not advocacy numbers. They are the companies' own disclosed emissions, trending upward at substantial rates, during the same period when per-query efficiency was improving.

One academic estimate (O'Brien, 2024) suggests real greenhouse gas footprints of major tech-company data centres may exceed reported values by over 600%, due to accounting choices around renewable energy certificates (RECs), market-based versus location-based accounting, and the gap between contracted clean energy and actual grid energy consumed. This is one estimate, not the consensus figure, but it's in the peer-reviewed academic literature and deserves consideration alongside the corporate self-disclosure.

NVIDIA shipped an estimated 3.76 million data-centre GPUs in 2023, roughly a million more than the year before, despite all the efficiency improvements at the chip and software level. The total compute installed is growing faster than the per-unit efficiency is improving.

The Microsoft-ExxonMobil AI deal (announced in 2019) involves Microsoft providing cloud and AI services to ExxonMobil's Permian Basin oil operations, a partnership expected to expand production by as much as 50,000 oil-equivalent barrels per day by 2025. A coalition of Microsoft employees calculated, in an internal memo reported by Grist and cited in the Luccioni paper, that the emissions enabled by that production increase could amount to roughly 640% of Microsoft's 2021 carbon-removal target, and these emissions are not included in Microsoft's carbon accounting, because they are the customer's emissions, not Microsoft's. This is the kind of structural accounting question Jevons' paradox forces into view. The carbon footprint of AI's use is often not counted as the carbon footprint of AI's production, and that boundary choice changes the answer dramatically.

Northern Virginia's data-centre corridor, anchored by Loudoun County's "Data Center Alley," is the core of a fleet consuming roughly a quarter of all electricity used in the state of Virginia (Dominion Energy reported 24% of its 2023 sales went to data centres), with more than 40 million square feet of data centres in Loudoun alone as of mid-2024, per Virginia's legislative audit commission. This is the operational scale that Canadian growth projections are pointing toward.

The IEA's global projection puts a number on the trajectory. The International Energy Agency estimates that data centres used approximately 415 TWh of electricity globally in 2024, and projects that they could reach approximately 945 TWh by 2030 in the IEA's base case. That's a roughly 130% increase over six years, against a backdrop of efficiency improvements that the Jevons-paradox framework predicts will be reinvested into more compute rather than reducing aggregate demand. Canada's share of this growth is not separately projected by the IEA, but the provincial allocation data in Chapter 6 (Ontario's 13% of new demand, Alberta's 16 GW of requested load, BC's competitive allocation, Quebec's new tariff) points in the same direction.

Put together: corporate environmental disclosures show major tech company emissions and water consumption growing 20–50% over the past few years, the academic literature suggests these self-reported numbers may understate the real footprint by 600%+, total installed AI compute is growing faster than efficiency is improving, and entire regional grids are now structured around data-centre demand. The aggregate environmental footprint of AI is growing at a rate that per-query efficiency improvements are not currently offsetting. That is the read Luccioni et al.'s Jevons-paradox framework explains.

---

The Canadian-specific carbon arithmetic

Canada's electricity grid is structurally unusual among major AI infrastructure jurisdictions: approximately 82% of Canadian electricity generation in 2023 came from non-emitting sources, dominated by hydropower (~60%), nuclear (~15%), and wind/solar (~7%). The remaining 18% is fossil-fuel-based, concentrated heavily in Alberta and Saskatchewan.

This produces a province-by-province carbon arithmetic for AI infrastructure that the AI for All strategy does not engage explicitly, but which Canadian readers should hold clearly.

A data centre in Quebec or Manitoba running on the provincial grid operates with roughly 30 grams of CO₂-equivalent per kilowatt-hour. This is among the lowest carbon intensities of any major data-centre jurisdiction globally. The Montréal cluster (520 MW of data centre load, Canada's largest) operates on Hydro-Québec's hydropower-dominated grid.

A data centre in British Columbia operates similarly low (~12-30 g CO₂e/kWh depending on the year), thanks to BC Hydro's hydropower-dominated mix.

A data centre in Ontario operates at roughly 50-100 g CO₂e/kWh depending on the year — primarily nuclear and hydropower, with some gas peaking.

A data centre in Alberta operates at roughly 400-500 g CO₂e/kWh — primarily natural gas, with declining coal and growing wind/solar. This is 10-15× higher per kilowatt-hour than Quebec or British Columbia.

A data centre using behind-the-meter dedicated natural gas generation (the architecture proposed for Wonder Valley in Treaty 8 territory, and increasingly used in US Texas and other gas-rich jurisdictions) operates at roughly 400-500 g CO₂e/kWh regardless of which province it's in, because it's bypassing the provincial grid mix entirely and using dedicated natural gas combustion.

The carbon implication of provincial siting choices. A 100 megawatt AI data centre operating continuously for one year consumes about 876,000 megawatt-hours of electricity. In Quebec at 30 g/kWh, that's about 26,000 tonnes CO₂-equivalent annually. In Alberta at 450 g/kWh, the same data centre produces about 394,000 tonnes annually: fifteen times the carbon. A gigawatt-scale facility (the Wonder Valley target) at Alberta intensity is on the order of 3-4 million tonnes of CO₂-equivalent annually from electricity alone, excluding embodied carbon, water-supply emissions, and supply-chain footprint.

The AI for All strategy does not commit to provincial siting requirements that would direct AI compute toward lower-carbon provinces. Whether it should is contested policy. Alberta's provincial government would push back hard against any federal direction of where Alberta-based industries can site. But the carbon arithmetic is what it is. Where Canadian AI compute is built directly determines its carbon intensity, and the current trajectory is to build substantial new capacity in Alberta (AWS Canada West in Calgary, Wonder Valley in Sturgeon Lake, and other gas-grid or behind-the-meter projects).

A separate Canadian-specific finding worth foregrounding. Natural Resources Canada's own technical guide notes that Canada's cold climate is a structural advantage for free-air cooling: "in cooler and dryer climates such as in many parts of Canada," substantial portions of the year can be cooled with outside air rather than evaporative water cooling or mechanical refrigeration. The per-facility water consumption for Canadian data centres is, by global standards, modest: a real environmental advantage. But the water reporting rate (22% of Canadian facilities publicly report) means this advantage is more inferred than measured. We'll come back to the transparency point in Chapter 18.

Here is a structural point the Canadian environmental conversation has largely sidestepped. The framing that Quebec and BC offer effectively unlimited low-carbon power for AI siting is not what the underlying hydrology supports. The Canada Energy Regulator documents that Quebec's electricity exports fell from 25.9 TWh in 2019 to 13.3 TWh in 2023 amid lower precipitation and changing domestic demand patterns. British Columbia has increased imports from the US Pacific Northwest for similar reasons over the same period. Site C strengthens BC's long-term supply but does not eliminate the climate-driven variability in hydropower output. The implication: "clean" and "surplus" are not the same thing, and the provincial low-carbon advantage that makes Quebec and BC attractive for AI siting is real but bounded by hydrology that varies year-to-year in ways natural-gas or nuclear generation does not. New compute clusters that depend on hydropower-dominant grids implicitly depend on continued favourable precipitation. Prudence argues for tying new compute clusters to incremental clean generation, transmission reinforcement, or firm low-carbon capacity rather than assuming legacy headroom. AI for All does not engage this constraint directly.

A second Canadian finding comes from cross-sector comparison, and it cuts the same way. A 100 MW AI campus consumes near zero to 0.18 million m³/year with dry/closed-loop cooling, or 1.6 to 2.5 million m³/year with evaporative cooling. Set against Canada's big water-using sectors (manufacturing alone withdrew 4,046 million m³ in 2021 per Statistics Canada / ECCC, and Ontario's thermal electricity generation withdrew 18,575 million m³), that is a rounding error. The scale you pick for the comparison decides whether AI water use looks small or large. At the national scale it is small; at the watershed, municipality, or summer-peak scale, a single AI campus can be the dominant new water user — particularly if it draws potable water for evaporative cooling. That is the structural reason water-stress screening belongs at the front of any siting decision, not in a box-checking exercise. In water-stressed local basins, evaporative cooling on potable water deserves far more scrutiny than the national-scale numbers alone would suggest.

One methodological caveat for Chapter 18's transparency analysis. Indirect water consumption from electricity generation depends substantially on how hydropower reservoir evaporation is treated, and in Canada's hydro-heavy grids that is a live choice: standard frameworks book hydropower as low-water generation, but reservoir surface evaporation can be substantial in absolute terms, and no standardized Canadian accounting method exists for it. Canada's hydro-grid water story is more complicated than the simple "low-water generation" framing suggests — a methodological question the federal AI strategy has not engaged.

---

The contrarian finding — Tomlinson et al. (2024)

Treating the environmental conversation seriously requires engaging the strongest contrarian empirical claim, not dismissing it. In 2024, a research paper by Tomlinson and colleagues (published in Scientific Reports, peer-reviewed) made the following claim: AI systems emit between 130 and 1500 times less CO₂-equivalent per page of generated text than human writers do for the same task.

The paper's argument is straightforward. A human writer producing a page of text consumes food (the embodied carbon of which is non-trivial), uses a computer for hours, sits in a heated and lit space, and produces the page over a duration of perhaps hours. An AI generates a page of text in seconds, consuming a fraction of a watt-hour. Per-page, the comparison favours AI dramatically.

The Luccioni paper engages this finding explicitly. Their critique:

First, the Tomlinson comparison excludes social impacts. If AI-generated text displaces human-produced text in markets where humans depended on that work for income, the displaced human writers don't stop having a carbon footprint — they have one without income, with all the downstream costs that produces.

Second, the comparison excludes rebound effects. If AI makes per-page production radically cheaper, the total quantity of text produced increases dramatically. The number of pages written per day, globally, was relatively bounded by human writing capacity before AI. With AI, the bound is removed. Total text-generation footprint may rise even as per-page footprint falls.

Third, the comparison excludes the broader question of what AI-generated text is being used for. If much of it is induced advertising content, the carbon footprint of the induced consumption needs to be counted somewhere.

Fourth, the per-page metric is the friendliest possible metric for the AI side. A different metric (per-task-completed, per-decision-made, per-dollar-of-economic-activity) might produce different results.

Reading Tomlinson et al. carefully. The paper is methodologically real. Its per-page conclusion is defensible within its scope. The scope is narrow. The broader claims sometimes drawn from it (that "AI is more sustainable than human work") do not follow from the data the paper actually presents. Tomlinson is the strongest peer-reviewed contrarian finding, and the guide treats it as such — engaging it carefully rather than dismissing it, while pointing out the boundary issues that Luccioni and others have raised in the academic exchange that followed.

This is one of the places where the guide's "contested questions stay contested" methodology matters most. The Tomlinson finding is real. The Luccioni critique of it is also real. Reading them side by side and noticing that the disagreement is genuinely about scope choices rather than about whether the data is fake is the practice the chapter is teaching.

The other side of the ledger: AI for the climate

The footprint is one half of the environmental account. The other half is the claim that AI will help fix the very problem its data centres worsen, and it earns the same scrutiny in both directions. Some of the benefit is demonstrated, and some of it is Canadian. Hydro-Québec put AI load-forecasting into production in 2023 and credits it with reading a 2024 heatwave its older models misjudged; Montréal's GHGSat uses AI to spot methane plumes from orbit; the Vancouver startup SenseNet's sensor-and-AI network flagged 217 fires in a single British Columbia pilot; and AI weather models now beat the gold-standard European forecaster on its own benchmarks. These are real. Most of them are also narrow, "traditional" machine learning rather than the generative kind doing the heavy energy consumption, which matters when the benefits of one are used to justify the footprint of the other.

The hype is real too. A widely repeated claim that AI could cut 5 to 10 percent of global emissions traces back to a 2021 consulting blog post extrapolating from client anecdotes. A 2026 analysis of 154 industry climate-benefit statements found roughly three-quarters unproven and a quarter citing no evidence at all; it is an advocacy report, so weigh it as the skeptics' number, but the pattern it describes is the bait-and-switch this chapter keeps naming, low-footprint forecasting tools cited to launder a high-footprint generative build-out.

The most calibrated read is the International Energy Agency's assessment, which lands almost exactly where this chapter does: fears that AI will wreck the climate are overstated, and so are the hopes that it will save it. The IEA estimates AI could cut emissions equal to roughly 5 percent of energy-sector emissions by 2035, larger than data centres' own footprint but far short of what the climate needs, and adds that there is currently no momentum to guarantee even that. Whatever efficiency AI delivers still meets Jevons at the door.

---

The structural finding — extraction geography

The Luccioni paper cites a 2023 study by Owen and colleagues, published in Nature Sustainability, that the guide has flagged in Chapter 1 and will return to in Chapter 9. The finding bears repeating with the chapter's full environmental context:

54% of the world's energy-transition-mineral projects (extracting cobalt, lithium, coltan, gallium, copper, tungsten, germanium, and rare earths, the same mineral families computing hardware depends on) are located on or near Indigenous peoples' lands; 69% once peasant lands are included; and 62% of the projects on those lands are in high-water-risk locations, against 53% of projects globally.

The technology-critical materials list is not abstract. Cobalt is in lithium-ion batteries used for uninterruptible power supply at data centres, and in the battery storage that pairs with renewable-energy commitments. Lithium is in the same battery systems. Coltan refines into tantalum, used in capacitors throughout server hardware. Gallium and germanium are in advanced semiconductor manufacturing. Copper is in essentially every electrical connection. Tungsten is in semiconductor manufacturing equipment. Rare earths are in the magnets used throughout precision computing hardware.

Every NVIDIA H100 GPU has its physical substance assembled from these mineral families. Every server rack, every data-centre power infrastructure component, every networking switch. The roughly 3.76 million data-centre GPUs NVIDIA shipped in 2023 each draw, at the materials level, on supply chains in which a majority of extraction projects sit on or near Indigenous and peasant lands. Project counts aren't material volumes; the study doesn't let us say what share of any single chip's substance came from those lands. But it does let us say the supply chain runs through that geography structurally, not incidentally.

This is the structural finding that connects the environmental conversation to the sovereignty conversation in a way that cannot be politely separated. The same Owen et al. number appears in Chapter 9 doing different work (sovereignty), in Chapter 11 doing different work (intellectual and material property), and in Chapter 15 doing different work (ethics). The chapters of this guide are not independent topics; they are different views onto the same underlying geography.

The environmental footprint of AI is not only what happens in the data centre. It is also what happens at the extraction site. The Canadian conversation about AI's environmental impact, when it focuses primarily on Canadian data-centre carbon intensity (which is lower than US comparators thanks to the hydropower grid), risks missing this layer. The compute substrate Canadian AI uses was extracted somewhere. The environmental and human costs of that extraction are borne by communities that are not part of the AI for All consultation pillars.

---

The ledger, both sides

CHAPTER RECAP — you now have: - The training-vs-inference distinction that lets you read environmental claims at the right scale. - Jevons' paradox as the analytical frame that explains why per-query efficiency can be improving and total environmental footprint can be growing simultaneously. The eight categories of indirect effects as the portable toolkit you can apply to any AI efficiency claim. - Google's August 2025 paper read carefully — methodologically real within its scope, framing strategic, unverified by independent third party, scope chosen to flatter the result. - The aggregate-growth case anchored in corporate self-disclosure: Microsoft, Google, Baidu all reporting double-digit emissions increases; NVIDIA's continued GPU growth despite efficiency improvements; O'Brien's 600%+ underreporting estimate. - The Canadian-specific carbon arithmetic by province: Quebec/Manitoba/BC at the global low end, Alberta at 10-15× higher, behind-the-meter gas generation at similar levels regardless of province. The lack of AI for All siting requirements as a contested policy gap. - The Tomlinson et al. contrarian finding engaged honestly: real within its scope, with scope-boundary issues that the Luccioni critique surfaces. Both findings preserved as legitimate. - The Owen et al. structural finding that puts AI's environmental footprint on the same geography as Indigenous-lands extraction worldwide — the through-line that connects this chapter to Chapters 9, 11, and 13.

The next chapter (Chapter 9) takes the sovereignty conversation to its deepest layer: the FNIGC OCAP Principles, the Inuit Tapiriit Kanatami National Inuit Strategy on Research, the international CARE Principles via the Global Indigenous Data Alliance, and the three senses of sovereignty introduced in Chapter 1 (national, personal, Indigenous) tested against each other. The Owen et al. finding will appear again, this time doing sovereignty work rather than environmental work.

You can now read any corporate AI environmental disclosure with the conceptual equipment to ask the right questions: What's the methodology? What's in scope and what isn't? Is this a per-unit metric or an aggregate? Where does Jevons' paradox apply? What's the verification status? Where are the materials sourced? Until those questions get answered, the environmental claim (whatever direction it points) is incomplete.

---

Bias label for this chapter: critical-engaged analysis of corporate environmental self-disclosure paired with independent academic critique. Author lean: skeptical of unverified corporate environmental claims regardless of which direction they point; sympathetic to Jevons-paradox framing as the most defensible analytical scaffold for the field; willing to engage contrarian peer-reviewed findings (Tomlinson) on their merits rather than dismissing them; attentive to the structural connection between AI environmental impact and Indigenous-lands extraction. Corporate self-disclosure (Google, Microsoft, Meta sustainability reports) labelled and read accordingly. Independent academic critique (Luccioni, Strubell, Crawford; O'Brien; Owen et al.) treated as primary. Contrarian academic findings (Tomlinson et al.) treated as legitimate within scope. Government technical sources (NRCan) treated as authoritative within their scope.

Primary sources cited or relied on in this chapter: Google Cloud Blog "Measuring the environmental impact of AI inference" (Vahdat & Dean, August 21, 2025); Luccioni, Strubell & Crawford, "From Efficiency Gains to Rebound Effects: The Problem of Jevons' paradox in AI's Polarized Environmental Debate," FAccT 2025 / arXiv:2501.16548; Owen, Kemp, Lèbre, Svobodova & Pérez Murillo, "Energy transition minerals and their intersection with land-connected peoples" (Nature Sustainability, 2023); Tomlinson, Black, Patterson & Torrance, "The carbon emissions of writing and illustrating are lower for AI than for humans" (Scientific Reports, 2024); Natural Resources Canada, Best Practice Guide for Canadian Data Centres (November 2024); Google 2024 Environmental Report; Microsoft 2024 Environmental Sustainability Report; O'Brien (2024) on corporate AI emissions accounting; Wang et al. (2024) on AI e-waste projection; Li et al. (2023) on AI water consumption; International Energy Agency, World Energy Outlook 2024 data-centre electricity projections (415 TWh 2024 → 945 TWh projected 2030); Canada Energy Regulator provincial profiles and 2023 hydropower variability documentation; Statistics Canada and Environment and Climate Change Canada 2021 water-use data for manufacturing, mining, agriculture/irrigation, and commercial/institutional sectors. Detailed citations in the Sources appendix.

---