The green AI myth got its clearest real-world test on 27 January 2025, the day DeepSeek’s R1 model shook the AI industry by matching leading Western reasoning models at a reported fraction of the training cost. Nvidia’s market value fell sharply within hours.
Microsoft chief executive Satya Nadella posted a one-line response on social media: “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.” He was invoking a 160-year-old idea from economist William Stanley Jevons, who observed, as NPR’s Planet Money explained at the time, that more efficient coal-fired steam engines led to more coal consumption, not less, because efficiency made the technology cheaper to run and therefore more widely used.
Eighteen months on, Nadella’s instinct has held up better than most industry predictions from that week. The green AI myth, the idea that making models more efficient will shrink AI’s overall environmental footprint, keeps colliding with a different reality: every efficiency gain in model design has so far been consumed, and then some, by the next wave of demand it makes possible. Understanding why the green AI myth persists, despite the data, is the starting point for reading any AI company’s environmental claims correctly.
What the 2026 Data Actually Shows
Testing the green AI myth against a full year and a half of disclosure data is now possible in a way it was not in early 2025, and the picture that emerges is consistent rather than ambiguous. Global data centres consumed an estimated 448 terawatt-hours of electricity in 2025, according to a comprehensive report published in June 2026 by the United Nations University Institute for Water, Environment and Health. The same report projects that figure could reach 945 terawatt-hours by 2030, with AI accounting for around 40 percent of that growth, alongside an associated water footprint of 9.3 trillion litres and a land footprint exceeding 14,500 square kilometres for the physical infrastructure involved.
The report’s authors were explicit about the mechanism behind the growth: inference, the ongoing work of running an already-trained model to answer everyday prompts, now accounts for 80 to 90 percent of total AI energy use, not the one-time cost of training it. The report puts real numbers on how unevenly that energy cost falls across tasks. A typical conversational chat query uses around 200 times the energy of basic text classification. Generating a single AI image uses roughly 1,450 times that baseline, and a single short AI-generated video can consume as much electricity as 200,000 spam-email classifications.
“A lot of people think that the environmental footprint of AI reduces, as technology improves and processes become more efficient,” said Professor Kaveh Madani, the report’s lead investigator. “But that is only a partial picture of the overall problem. More efficient and affordable AI and energy mean more consumption of AI, making the overall footprint far bigger than what we save through efficiency gains.”
What This Means for You
If your organisation is weighing an AI vendor’s environmental claims, or if you are personally deciding which model to reach for by default, the green AI myth is worth naming out loud before you take an efficiency statistic at face value. A vendor reporting a large drop in energy per query is very likely telling the truth about that specific metric. It is a different claim entirely from saying the company’s total AI energy consumption has fallen, and the two get blurred constantly in marketing material. Ask which of the two is actually being claimed before repeating either one.
The DeepSeek episode itself is a useful test case for the green AI myth in practice. A genuinely cheaper, more efficient model did not shrink the market for AI compute. It triggered a wave of new entrants building on the same efficiency techniques, a round of price cuts across the industry, and a documented surge in usage as AI became viable for tasks that had not previously justified the cost. Efficiency opened the market wider rather than making the existing market smaller.
The Green AI Myth in One Number: 33x Efficiency, Bigger Footprint Anyway
Google published its own detailed methodology for measuring Gemini’s environmental footprint in an August 2025 Google Cloud Blog post, reporting that the median Gemini Apps text prompt used just 0.24 watt-hours of energy, roughly the equivalent of watching nine seconds of television. Google said it had cut energy use per prompt by a factor of 33 and carbon footprint by a factor of 44 over the preceding twelve months. Those figures are credible and independently documented, produced through a full-stack measurement methodology covering the entire serving infrastructure rather than just the AI chip itself.
None of that changes the total trajectory, because the efficiency gain and the demand growth are not separate stories. They are the same story, and this is the green AI myth in its purest form. A 33-fold drop in the cost of a single query is precisely the kind of change that, per Jevons’s original logic, makes people run far more queries, feed longer documents into the model, generate more images, and build the query into workflows that did not exist when it was more expensive to run.
The per-query number gets smaller. The number of queries gets very much larger. Multiplying the two together is what determines total consumption, and on the evidence assembled so far, the second number is winning. That is the green AI myth’s central failure mode: a true efficiency statistic, presented on its own, implies a conclusion about total impact that the same company’s own growth numbers usually contradict.
Where the Emissions Are Actually Going
Microsoft’s own environmental sustainability report, released in July 2026, put a concrete figure on that dynamic at company scale. Total emissions rose 25 percent year over year in its 2025 fiscal year, reaching 20.29 million metric tons of carbon dioxide equivalent, up from 16.21 million tons the year before. The company attributed the rise primarily to data centre expansion, alongside an accounting change in which it stopped counting certain unbundled renewable energy certificates not tied to a specific new energy project, a shift that pushed its Scope 2 emissions share from 2 percent to 13 percent of its total footprint in a single year.
Microsoft’s own leadership described the tension plainly in the report’s foreword: AI is “driving demand for energy, water, land and materials,” while “sustainability solutions are not scaling fast enough to meet demand.” That is not a story about one company failing to keep its promises. It is closer to the mechanism the UNU-INWEH researchers describe at the level of the whole industry, and it is the green AI myth playing out in a single company’s own disclosure: real efficiency work is happening at the model level, at the same time as absolute infrastructure build-out grows even faster, because the efficiency gains are precisely what make that build-out commercially viable at scale in the first place.
Why Some Rebound Effects Are Real and Others Are Overstated
The honest caveat, well documented in energy economics, is that Jevons-style rebound effects are not always as total as the paradox implies. Economists who have studied fuel-efficient cars and appliances generally find that rebound effects are real but partial: a more efficient refrigerator or vehicle still tends to reduce total energy consumption over its lifetime, even after accounting for the fact that cheaper running costs encourage somewhat more use.
A full Jevons paradox, where efficiency gains are not just offset but overwhelmed by demand growth, requires demand to be highly responsive to falling costs. That is precisely the condition that appears to hold for AI inference at the moment, given how many new use cases become viable each time a query gets cheaper, but it is not a law of nature, and it does not apply equally to every efficient technology.
That distinction matters because it is also the one LiveAIWire’s reporting on machine greenwashing in the AI industry found companies routinely blur. A specific, audited efficiency figure for a named product is a genuinely different claim from a general assertion that AI’s environmental trajectory is improving, and the industry’s public communications have tended to lead with the first while implying the second. The green AI myth survives on exactly that blur, and it does not survive contact with a company’s own total emissions disclosure.
The Other Half of the Ledger
None of this means AI’s net effect on energy systems is straightforwardly negative. AI-driven grid optimisation and renewable forecasting are delivering quantified benefits that the International Energy Agency has documented independently, including as much as 175 gigawatts of additional transmission capacity unlocked in existing grids and up to 110 billion dollars in annual power plant savings by 2035.
The full accounting of both sides of that ledger, covered in LiveAIWire’s broader look at AI’s environmental cost and benefit, matters precisely because the green AI myth is not that AI has no environmental upside. It is the narrower, more specific claim that efficiency improvements in model design are, by themselves, enough to shrink AI’s total footprint. On the 2026 evidence, they are not, because they are also the mechanism driving the growth they are supposed to offset.
What Would Actually Change the Trajectory
The policy tools that have addressed rebound effects in other energy-intensive industries apply to AI with modest adaptation. Disclosure rules that separate AI-specific energy, water and carbon use from a company’s general computing footprint would make the rebound visible in company accounts rather than buried in aggregate figures, closing exactly the gap that lets an efficiency statistic stand in for a total-impact claim.
Guardrails on default output length, resolution and model selection, choosing the smallest model genuinely capable of a task rather than defaulting to the most capable one available, target the demand side of the equation directly rather than relying on efficiency gains that the rebound effect keeps absorbing. Carbon and water pricing that reflects the actual environmental cost of a query would change the economics that currently reward scale over restraint, and would make the green AI myth much harder to repeat with a straight face.
None of that requires new technology. It requires treating AI’s environmental footprint the way other resource-intensive industries have eventually been required to: measured at the level of the specific product and workload, disclosed consistently, and priced so that the true cost of a query is visible to whoever is deciding how many of them to run.
Until that happens, the honest reading of the 2026 data undercuts the green AI myth directly: efficiency gains in AI model design are real, and they have so far been a reliable predictor of how much bigger AI’s total footprint is about to get, not how much smaller. The green AI myth will keep resurfacing with every new efficient model release, right up until the disclosure and pricing gaps that let it survive are actually closed.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.