The Environmental Cost of AI Nobody Talks About

It is not the big queries. It is the thoughtless ones.

When Sam Altman admitted that users saying “please” and “thank you” to ChatGPT costs OpenAI tens of millions of dollars in electricity, most people treated it as a curiosity. A quirky data point. A reason to feel either guilty or charmed, depending on your disposition.

The actual takeaway is harder to sit with. If politeness tokens cost tens of millions, what does the rest of it cost? The AI-generated meme that got four likes. The chatbot query asking for a word that means “happy”. The company-wide GenAI rollout that produced nothing measurable after eighteen months of compute. These are not hypothetical edge cases. There is now research quantifying exactly this kind of waste, and the numbers are uncomfortable reading.

The “10x Google” number is wrong. The problem is not.

The most repeated figure in AI environmental discourse is that a ChatGPT query uses ten times the electricity of a Google search. It traces to a 2023 paper in Joule by Alex de Vries, who extrapolated from a 2009 Google energy figure and a remark by an Alphabet executive. It was a reasonable estimate for 2023 hardware. It is not a reasonable figure now.

Google’s own 2025 disclosure put a typical Gemini text query at approximately 0.24 watt-hours. Epoch AI’s independent analysis arrived at roughly 0.3 Wh for GPT-4o. These are comparable to a few seconds of an LED bulb, or about nine seconds of watching television. For text queries, the individual footprint is genuinely small.

The “10x Google” correction matters because the conversation keeps landing in the wrong place. Scolding individuals for their text queries is a distraction. The real environmental story has three parts, and per-query text energy is not one of them.

What you actually need to watch: images, video, and raw scale

The energy cost of generative AI is not flat. It varies by several orders of magnitude depending on what you are generating.

Sasha Luccioni and colleagues at Hugging Face and Carnegie Mellon published a benchmarking study in 2023 that quantified this gap directly. Generating 1,000 images with Stable Diffusion XL produced roughly 1,600 grams of CO2, equivalent to driving a gas car about 4.1 miles. Generating 1,000 text responses with the same compute time produced a fraction of that. A single image generation task consumes about as much energy as fully charging a smartphone. A thousand text generations consume about 16% of a phone charge.

Video is in a different category entirely. A 2025 investigation by MIT Technology Review and Hugging Face, using the University of Michigan’s ML.Energy benchmark lab, found that generating a five-second video with the open CogVideoX model used approximately 3.4 million joules. That is comparable to running a microwave for over an hour. A later study estimated that video diffusion models are roughly 30 times costlier than image generation and around 2,000 times costlier than text. Sora-class commercial video generation may approach one kilowatt-hour per five-second clip.

This is not an argument against all generative image or video work. It is an argument against the throwaway version. The AI-generated thumbnail that replaced a five-minute Canva task. The bulk social media content farm running on diffusion models. The novelty filters shipping with every major messaging app.

Luccioni’s finding that most people did not think about: using a general-purpose generative model for tasks like text classification consumes around 30 times more energy than using a task-specific fine-tuned model, even when controlling for parameter count. The decision to reach for the largest available model because it is convenient is not neutral. It has a real cost.

43% of your AI queries probably did not need AI

The sharpest empirical work on unnecessary AI use comes from a 2026 University of Calgary study titled “Sustainable AI Assistance Through Digital Sobriety.” The researchers analysed the public LMSYS-Chat-1M dataset, coding queries as either factoid (answerable by conventional search with no reasoning required), complex (genuinely needs the model), or convenience.

Their finding: 43.2% of sampled queries were factoid requests. Nearly half. Information retrieval was the single largest source of avoidable use, accounting for 76% of factoid queries in the information-retrieval category.

The researchers modelled the energy implications. Using 0.3 Wh per AI query versus 0.03 Wh per search, redirecting factoid queries to search could reduce associated energy use by roughly 43%. If convenience queries are included, the estimate rises to 58.7%.

There are caveats worth stating plainly. The sample was 148 manually coded queries from a single dataset. The paper is a preprint. The 58.7% figure is speculative. But the directional finding is consistent with what Luccioni and de Vries have argued separately: a large portion of generative AI use is doing work that cheaper, lower-energy tools could handle adequately.

Luccioni put it plainly in a CNN interview in June 2025: “We don’t need generative AI in web search. Nobody asked for AI chatbots in messaging apps or on social media. This race to stuff them into every single existing technology is truly infuriating, since it comes with real consequences to our planet.”

The water problem almost nobody reads past the headline

The figure that circulates most is 519 millilitres of water per 100-word ChatGPT response, from a 2023 University of California Riverside study by Shaolei Ren and colleagues. Sam Altman disputed this publicly, citing 0.3 millilitres per query.

Both numbers are real. They measure different things. Ren’s 519 ml includes indirect water consumed at electricity-generating plants, plus on-site evaporative cooling. Altman’s 0.3 ml counts only direct on-site cooling. The gap is not a factual dispute. It is an accounting boundary dispute, and it matters which boundary you accept.

Google’s 2025 Environmental Report disclosed 8.1 billion gallons of water consumed across its data centres and offices in 2024. That number is real regardless of where you draw the boundary. Data centres in drought-affected regions are already producing local conflicts: in Santiago, Chile; in Aragón, Spain, where protests carried signs reading “Your cloud is drying my river”; in Newton County, Georgia, where a Meta facility used roughly 500,000 gallons per day, equalling approximately 10% of the county’s total consumption.

The aggregate water story tracks the energy story. The individual query is negligible. The aggregate buildout is not.

The system-level picture, by the numbers

The International Energy Agency’s 2025 Energy and AI report projected global data-centre electricity consumption at roughly 415 TWh in 2024. It projects this reaching approximately 945 TWh by 2030, slightly more than Japan’s total electricity consumption today. AI is identified as the primary driver of that growth. AI-focused data-centre demand grew 50% in 2025 alone and is projected to triple by 2030.

Goldman Sachs projected global data-centre power demand increasing 165% between 2023 and 2030, with AI’s share of that rising from around 14% to over a quarter. Lawrence Berkeley National Lab, looking specifically at the United States, projected American data centres consuming between 325 and 580 terawatt-hours by 2028. The upper end would represent roughly 12% of total US electricity use.

There are second-order effects already visible. The PJM Interconnection, the largest wholesale electricity market in the United States, saw capacity prices rise 75.5% year-on-year in 2026. PJM’s own market monitor attributed 63% of that increase to data-centre load growth, and described the impacts as “very large and not reversible.” Consumer electricity bills in data-centre-dense regions are rising, paid in part to subsidise compute that, in many cases, is producing nothing measurable.

The ROI problem makes it worse

MIT’s Project NANDA published a report in July 2025 analysing over 300 enterprise GenAI deployments across 52 executive interviews and 153 leader surveys. The headline finding: 95% of integrated GenAI pilots produced no measurable profit-and-loss impact. Five percent extracted significant value. The rest spent compute, energy, and money on nothing the business could point to.

This does not mean AI cannot deliver value. It means most organisations deploying it currently are not capturing that value. The environmental cost of failed or directionless AI deployment is real energy spent on real infrastructure for demonstrations, proofs of concept, and executive dashboards that went nowhere.

The researchers noted a “shadow AI economy” where unsanctioned consumer-tool use by employees may deliver uncounted productivity gains. That is plausible. It does not resolve the fundamental question of whether large-scale model deployment for vague organisational goals has been worth its physical costs.

What the researchers actually recommend

The proposals that appear across the credible literature are specific and consistent.

Luccioni advocates for what the French call “sobriété numérique,” digital sobriety: a principle that you should use the smallest, most task-specific model that can do the job. A 70-billion-parameter general model to classify whether a review is positive is not digital sobriety. A fine-tuned small model for the same task, using 30 times less energy, is.

The University of Calgary authors recommend routing: before reaching for generative AI, ask whether the task genuinely requires reasoning. If it is a factoid, use search. If it is a classification task, consider whether a smaller model or conventional tool is adequate.

At the infrastructure level, researchers point to carbon-aware scheduling, direct-to-chip cooling, and renewable procurement. On the policy side, mandatory energy and water disclosure at the facility and model level would help. California’s 2023 climate laws and the EU’s CSRD move in this direction. A federal Clean Cloud Act was proposed in the United States in 2025. Whether any of this moves fast enough to match the buildout pace is a separate and more uncertain question.

The part worth sitting with

The Jevons Paradox applies here as much as anywhere. When something becomes more efficient, people use more of it. AI text generation is genuinely becoming cheaper and less energy-intensive per query. That does not mean total consumption is falling. The evidence so far suggests the opposite.

Efficiency gains and demand growth are running in the same direction, not in opposition. The IEA projects AI-specific data-centre demand tripling by 2030 despite hardware efficiency improvements that would have been extraordinary by any historical standard. The concern is not that AI is irredeemably harmful. The concern is that the current deployment pattern, reflexive, unaudited, generative by default, is consuming a resource budget that nobody formally approved and that lands disproportionately on grid users, local water systems, and decarbonisation timelines.

The individual who stops using ChatGPT as a search engine for factoids will not move the needle on global electricity demand. But the organisational habit of deploying the largest available model for every task, and the product habit of embedding generative AI into every surface whether it adds value or not, those are tractable decisions. They are being made right now by people who have not done this accounting.

Sources: IEA Energy and AI Report (2025, 2026); Luccioni, Jernite & Strubell, “Power Hungry Processing” (Hugging Face/CMU, 2023); de Vries, “The growing energy footprint of artificial intelligence,” Joule (2023); Jennings, Deb & de Souza Santos, “Sustainable AI Assistance Through Digital Sobriety,” University of Calgary (2026 preprint); MIT Project NANDA, “The GenAI Divide” (2025); Ren et al., “Making AI Less Thirsty,” UC Riverside (2023); Goldman Sachs Data Center Power Demand Analysis; Lawrence Berkeley National Laboratory Data Center Report (2024); Google Environmental Report (2025); PJM Monitoring Analytics (2026).

Anjula Weeranayake @ TekDruid (https://tekdruid.com/)