Writing

How much does AI really pollute? The math, burger in hand

The viral numbers about energy-hungry artificial intelligence are an order of magnitude out of date. The real ones tell a different story: uncomfortable for meat at the individual level, uncomfortable for data centers at the system level.

Daniele Bianchini · June 2026 · English · ~15 minutes

Work in progress — this piece is still being revised. Researched and written with the help of Claude Fable 5 (June 2026). Leggilo in italiano →

I make video games for a living, and I keep AI agents working for most of the day: they write code, analyze sales data, grind through millions of tokens while I do something else. In every consumption table you’ll read in this article, I sit in the worst cell. So when I asked myself how much AI really pollutes, it wasn’t a rhetorical question: I wanted to know whether my daily work is an environmental problem.

At the same time, I haven’t eaten meat in years. The comparison you’re about to read is therefore not a defense of my steak: the steak left my plate a long time ago. I have conflicts of interest in both directions, and that’s exactly why the only thing I trust here is the numbers.

So I went looking for them. Not the viral ones: the published ones. The technical reports from Google and Mistral, the projections of the International Energy Agency, the largest meta-analysis ever conducted on the footprint of food. And I picked a concrete benchmark, something almost all of us do without thinking: eating beef.

Two spoilers, just to set the orders of magnitude. To match the water footprint of a single hamburger, you would have to use a chatbot every day for more than six centuries. And to match the CO2 of the beef an average American eats in a year, I would have to send about 237,000 requests a day. Every day, for twelve months.

But beware: this is not an article that absolves AI. It’s an article that declares its lenses. And through the right lens, AI has a real problem too. It’s just not the one everyone talks about.

Why almost every comparison you’ve read is broken

Two flaws spoil most of the debate.

The first: the viral numbers are old. The estimate still circulating today (“one ChatGPT request uses ten times a Google search, about 3 watt-hours”) dates back to 2023. In August 2025 Google published the first detailed technical report from a major provider on per-request costs in production, and the number is 0.24 watt-hours: twelve times less [1]. Not only that: energy per prompt dropped 33-fold in a single year, between May 2024 and May 2025 [1]. Quoting 2023 estimates to describe the AI of 2026 is like reviewing a smartphone with the spec sheet from three generations ago.

The second flaw is sneakier: mixing scales. “AI consumes as much as a small country” and “one query consumes as much as one second of a microwave” are both true statements, and whoever wants to convince you of something picks the one that suits them. The single action and the global system give different answers, and both are correct in their own domain. Any article that doesn’t tell you which lens it’s using is steering you, knowingly or not.

This piece uses both lenses, one at a time. All the assumptions are collected in the methodology note at the bottom, with calculations you can redo by hand.

Lens 1: the single action

What one query weighs

Gemini’s median text prompt consumes 0.24 watt-hours of electricity, emits 0.03 grams of CO2 equivalent and consumes 0.26 milliliters of water [1]. The figure includes not just the chips doing the computing, but also CPUs, memory, idle machines, cooling and power conversion. OpenAI declares a similar value for ChatGPT: 0.34 Wh per request [2]. Translated into everyday experience: one second of a microwave oven.

Honesty requires a caveat: that’s the median. A deep-research session, a long reasoning chain or an agent working for hours sits one or two orders of magnitude above it. That’s why, in the calculations that follow, I also use a “heavy query” at 5 grams of CO2 — 165 times the median — as a safety margin working against my own thesis.

What one serving of beef weighs

The reference is the meta-analysis by Poore and Nemecek published in Science in 2018, built on 38,000 farms across 119 countries [3]. Beef from beef herds (the steak, the beef burger) emits on average about 100 kg of CO2 equivalent per kilogram; beef from dairy herds about a third of that. A 100-gram serving is therefore worth about 1,000 grams of CO2e.

The division

1,000 divided by 0.03: one serving of beef equals about 33,000 median chatbot requests. Even counting only heavy agent queries, it takes 200 of them to match one hamburger.

And on an annual basis? An average American eats about 26 kg of beef a year at retail weight [12], roughly 2.6 tonnes of CO2e. Here’s the comparison with four AI usage profiles, from the most common to the most extreme:

The “extreme agentic dev” scenario is, I repeat, mine: agents working all day, every day. Even so, my annual AI footprint stays twelve times below the beef alone of an average American. For a normal user the ratio is thousands to one. And it holds on this side of the Atlantic too: in Italy, actual beef consumption is about 10.5 kg per person per year — about one tonne of CO2e — still five times my worst-case scenario.

At the level of the individual person, the conclusion is not ambiguous: diet dominates the chatbot by orders of magnitude.

Carbon is the easy part

If the comparison stopped at CO2, I would be underselling the gap. It’s on the other dimensions that beef carries its most extreme disadvantage.

Water. A kilo of beef has a total water footprint of about 15,000 liters [4]; a query consumes 0.26 milliliters. That makes 58 million requests per kilo of beef. An analysis by SemiAnalysis calculated that the “blue” water alone (the water drawn from aquifers and rivers) of a single hamburger equals 668 years of daily chatbot use [4].

Here, though, comes the article’s most important caveat in favor of AI’s critics: data center water is a real local problem. The IEA estimates global data center water consumption at around 560 billion liters in 2023, possibly doubling by 2030 [5]. It’s little compared to agriculture, but it concentrates in a few areas, sometimes already under water stress, and evaporation in cooling towers takes fresh water away from the surrounding community. The technologies to reduce it exist (immersion or direct-to-chip cooling cuts the requirement by up to 95%), but they have to be demanded. The point is not that data centers don’t consume water: it’s that per unit of service the comparison is lopsided, and that the problem is managed with local rules, not by giving up queries.

Land. Perhaps the starkest asymmetry of all. Producing 100 grams of protein from beef requires a median 104 square meters of land; plant proteins require 1 to 7 [3]. An American’s annual beef occupies between 4,000 and 5,000 square meters — nearly half a hectare per person, every year. Globally, livestock uses 83% of the planet’s farmland while providing only 18% of calories and 37% of protein [6]; according to the same authors, a diet without animal products would free up 76% of the world’s agricultural land, an area the size of the United States, China, the European Union and Australia combined. A data center’s land footprint, per user served, is measured in square centimeters.

Deforestation. Cattle ranching is the single largest driver of tropical deforestation: about 41% globally, 72% in Brazil [7], up to 80% of deforestation in Amazonian countries. Two points of honesty. First: part of these emissions is already included in the 100 kg/kg factor, so it must not be counted twice. Second: data centers have no direct links to deforestation (at most, indirect impacts through the mineral and hardware supply chain). But the fact remains that this cost column weighs almost entirely on one side of the comparison.

“Sure, but the training”

The classic objection: fine for the single query, but training the models costs a fortune. Let’s see.

The number that has circulated for years — roughly 500 tonnes of CO2e to train GPT-3 — is from 2020 and is by now more meme than data point. Recent models cost much more: training Llama 3.1 405B is declared by Meta at about 11,400 tonnes [13], and the full life-cycle analysis of Mistral Large 2 (training, plus 18 months of usage, plus hardware manufacturing) comes to about 20,400 tonnes [8].

They sound like enormous figures, until you do the division. ChatGPT processes about 2.5 billion prompts a day [8], over 900 billion a year. Spreading Llama’s 11,400 tonnes over a year of traffic at that scale yields 0.012 grams of CO2e per request: less than half the cost of the inference itself. For models used by hundreds of millions of people, training is the minor item in the per-request footprint.

While we’re at it, let’s update the “5 cars” meme too: that comparison came from a 2019 experiment of about 284 tonnes, not from GPT-3. With a realistic life cycle of 57 tonnes per car, GPT-3 is worth about 9 car-lifetimes and Llama 405B about 200. Correct numbers should be used even when they’re less convenient for your own thesis.

Lens 2: the system — where the story flips

Let’s switch lenses and look at the planetary balance sheets.

The world’s livestock sector emits 6.2 billion tonnes of CO2e a year, about 12% of the anthropogenic total; cattle alone, between meat and dairy, account for 3.8 billion tonnes, about 7% [9]. Data centers consume 415 TWh, 1.5% of the world’s electricity [10]; in terms of emissions we’re between 0.5 and 1.5% depending on the calculation boundaries. Be wary of anyone citing 3.7%: that figure covers the entire ICT sector, devices and networks included, not data centers. Today, in emissions, cattle weigh about 14 times data centers.

But the present is not the point. The IEA projects data centers at about 945 TWh in 2030: 15% growth per year, four times faster than the rest of the world’s electricity consumption [10]. Livestock grows about 1% a year, and in rich countries meat consumption is falling. The two curves point in opposite directions: the meat problem is enormous but standing still; the data center problem is small but running.

And for now, it runs on gas. In the 2024 electricity mix of American data centers, natural gas weighs over 40%, renewables 24%, nuclear 20%, coal 15% [11]. The reason is prosaic: connecting a gas plant to the grid costs about a tenth of a solar or wind farm [11]. There it is, the legitimate concern about AI: not your chat, but an infrastructure race that in the short term is powered by fossil fuels, in one of the few sectors in the world with growing emissions.

The moving target (and the cow you can’t decarbonize)

Here lies the asymmetry that decides the comparison in the long run, and it’s the concept I take home from all this research.

AI’s carbon per query is a target moving downward, pushed by two independent engines. The first is the grid cleaning itself up: in the IEA’s base scenario, renewables cover about half of new data center demand by 2030 [10], and the big tech groups are signing nuclear deals, including the reactivation of decommissioned plants. The second is model efficiency: 33-fold in twelve months, as we’ve seen. On top of that, AI actively works on the power grid itself: according to the IEA it can cut blackout duration by 30-50% through automatic fault detection, and unlock up to 175 GW of existing transmission capacity through sensors and smart management [10] — more than the entire projected data center load increase to 2030.

Beef’s carbon per kilo, on the other hand, has a biological floor: enteric methane is a byproduct of ruminant digestion. It can be mitigated with feed additives and genetic selection, not zeroed out. You can’t decarbonize a cow the way you decarbonize a power grid.

Honesty toward the other side as well: there is a serious scientific debate, around the GWP* metric, according to which — for a herd of stable size — methane, being a short-lived gas, warms less than the standard metric used in these calculations suggests. It’s a fair point, and livestock’s defenders are right to raise it. But its limits remain true too: herd growth adds net warming, and the CO2 from deforestation and feed cultivation is not biogenic and does not recycle.

In short: the two footprints do not share the same destiny over time. AI’s can only go down, per unit of service. Meat’s, where it stays, stays.

What mass usage is unknowingly funding

There’s one last gear on the AI side that deserves to enter the picture, and I know it well for professional reasons.

The GPUs that train today’s frontier models weren’t born for science: they were born for video games. For twenty years, millions of players — buying graphics cards to run their favorite titles better — unknowingly funded the development of parallel computing. In 2012 AlexNet, the neural network that inaugurated the modern era of deep learning, was trained on two gamer video cards. “Frivolous” mass demand built the infrastructure today’s research runs on.

With AI it’s happening again, at a larger scale. The billions of daily requests [8] fund the frontier models, and those models are already accelerating research in documentable ways: AlphaFold solved protein structure prediction and earned the 2024 Nobel Prize in Chemistry, GNoME proposed hundreds of thousands of new stable materials, next-generation weather models beat traditional systems at a fraction of the compute cost, and plasma control in experimental fusion reactors has already been demonstrated with reinforcement learning agents. Ray Kurzweil would call it the law of accelerating returns; I prefer the list of concrete cases.

Careful, though, with how you handle this argument, because it’s the easiest one to use badly. It’s not an accounting wash (“today’s queries are offset by tomorrow’s science”): it’s option value, real but conditional, depending on where the computing capacity gets directed. You can say that mass demand is funding a general-purpose technology with already measurable scientific returns. You cannot say that this zeroes out today’s environmental cost. The first claim is a fact; the second is marketing.

The asymmetry that actually matters

There’s an argument circulating on the pro-AI side that I advise you not to use: “AI’s emissions are an investment in the future, meat’s are consumption at a dead loss.” That’s a value judgment dressed up as accounting: food produces nutrition and pleasure, like almost every human consumption. A careful reader takes it apart in two lines.

The defensible asymmetry is a different one: substitutability. Beef has functional substitutes (legumes, poultry, other proteins) with a footprint per gram of protein 10 to 50 times lower: most of its emissions are avoidable today, at equal nutritional value, by changing what’s on the plate. A useful cognitive task, on the other hand, still has to get done: the alternative to AI is not “nothing” — it’s human time plus a computer that’s switched on.

But watch out for the symmetric error, the pro-AI one: comparisons like “AI does in 30 seconds what costs a human 4 office hours, saving 90% of the energy” are broken at the root, because the office and the worker don’t switch off while the AI works (that’s the critique that sank a well-known 2024 study on AI and human writers). The honest version of the calculation is marginal — it counts only the additional consumption — and gives AI a 5-20x advantage on the single task. An advantage that holds only as long as the total volume of tasks doesn’t explode. Because the Jevons paradox — efficiency lowering costs and multiplying demand — is not a footnote to this story: it’s exactly what is happening. Queries 33 times more efficient, total consumption doubling. Both things at once, without contradiction.

The plastic straw of 2026

A few years ago the world declared war on plastic straws. They were visible, symbolic, perfect for moralizing; they weighed a minuscule fraction of the plastic ending up in the oceans. For a couple of years they absorbed a disproportionate share of collective environmental attention, while the flows that really mattered stayed comfortable and unobserved.

The AI query risks becoming the straw of 2026: the new, visible gesture to concentrate blame on, two or three orders of magnitude below the behaviors our attention prefers not to look at. Guilt is a scarce resource: spending it on 0.03 grams while there are a thousand on the plate is, before being unfair, a bad allocation.

I want to be precise on one point, because it’s easy to slide into moralism here. I don’t call this asymmetry hypocrisy: hypocrisy presupposes knowing the numbers, and almost nobody knows them, since the public debate tells them backwards. That’s why this article exists. But once you’ve seen the numbers, the asymmetry stops being ignorance and becomes a choice. Getting informed, staying aware and adjusting your habits a little, without extremism and without sermons: that doesn’t seem like an unreasonable ask.

Four things I’m not saying

To keep this article from being read as what it is not:

1. I’m not saying AI has zero impact. The data center race is real, it currently runs largely on gas, and it’s one of the few sectors with growing emissions while the rest of the economy tries to decarbonize.
2. I’m not saying you should stop eating meat. I haven’t eaten it in years, but that remains a personal choice and this is not an article of dietary proselytizing. I’m saying that anyone worried about their footprint should know where the orders of magnitude lie, and today the public debate tells them backwards.
3. I’m not saying data center water doesn’t matter. It’s a serious local problem, deserving mandatory transparency on consumption and strict rules in water-stressed areas.
4. I’m not saying AI will save the climate. Its enabling value is real and documented — I wrote about it above — but conditional: today most inference serves chat, code and advertising, not AlphaFold. Presenting it as automatic compensation would be as dishonest as the viral numbers of 2023.

The two answers, stated plainly

If the question is personal (“is my AI use an environmental problem?”), the answer is no, by enormous margins: your diet dominates your chatbot by two, three, four orders of magnitude on every measurable dimension — carbon, water, land. Skipping one burger a week “buys” more than any digital abstinence you could practice. I will keep running my agents without guilt.

If the question is systemic (“is AI an environmental problem?”), the honest answer is: not yet, but the trajectory needs watching. One percentage point of global emissions by 2030 is not a catastrophe; +15% a year powered by gas, however, is the kind of curve that deserves public policy, mandatory transparency on consumption, and pressure for the new demand to be born clean.

They are two problems of different natures: one small and running fast, but it knows how to decarbonize; one enormous and standing still, and it cannot. The comparison holds without rhetoric. Just use the 2025 numbers instead of the 2023 ones, declare your lens, and acknowledge the costs on both sides.

Which, come to think of it, would be a good honesty test for any debate — not just this one.

Methodology

All the calculations in this article can be redone by hand with these assumptions, chosen where possible to work against the main thesis:

• Energy and carbon per query: 0.24 Wh and 0.03 gCO2e (Gemini median, Google data [1]); for the annual scenarios, carbon is recalculated at 250 gCO2e/kWh, a more realistic grid intensity, double the one implicit in Google’s data (~125 g/kWh).
• “Heavy” query: 5 gCO2e, about 165 times the median, to represent reasoning, deep research and agentic workloads.
• Beef: 100 kg CO2e/kg, global average for beef herds (Poore & Nemecek 2018 [3]); dairy-origin beef drops to ~33. Full range across producers: 9–105 kg/kg.
• US consumption: 26 kg/year at retail weight (USDA [12]); availability is ~38 kg. Actual Italian consumption: ~10.5 kg (half the apparent figure, loss-adjusted method).
• Water: 15,000 L/kg is the total water footprint (green + blue + grey); blue water alone is much lower, but SemiAnalysis’s per-burger comparison [4] uses exactly that, and the result still comes to six centuries.
• Amortized training: 11,390 t (Llama 3.1 405B [13]) divided by a year of traffic at ChatGPT scale (~900 billion prompts [8]).
• Car-lifetime: 57 tonnes of CO2e per American car, manufacturing and fuel included.

Sources

1. Google (2025), technical report on the environmental impact of Gemini inference: 0.24 Wh, 0.03 gCO2e, 0.26 mL per median prompt; 33x improvement in 12 months. blog.google
2. OpenAI / S. Altman (2025), average consumption per ChatGPT query ~0.34 Wh. adnkronos.com
3. Poore, J. & Nemecek, T. (2018), “Reducing food’s environmental impacts through producers and consumers”, Science 360(6392). Data via Our World in Data. ourworldindata.org
4. SemiAnalysis (2026), “From Tokens to Burgers: A Water Footprint Face-Off”; beef water footprint ~15,000 L/kg (Mekonnen & Hoekstra). semianalysis.com
5. IEA (2025), “Energy and AI”: data center water consumption ~560 billion liters (2023), possibly doubling by 2030. iea.org
6. Oxford LEAP (2018): livestock = 83% of farmland, 18% of calories, 37% of protein; plant-based diet = -76% agricultural land. leap.ox.ac.uk
7. Sentient Media / Greenpeace (2022): cattle ranching = 41% of tropical deforestation, 72% in Brazil; Yale Global Forest Atlas: ~80% in Amazonian countries. sentientmedia.org
8. Mistral AI, ADEME, Carbone 4 (2025), LCA of Mistral Large 2: 20.4 ktCO2e; ChatGPT ~2.5 billion prompts/day (Chatterji et al. 2025). Summary in arXiv 2512.11863. arxiv.org
9. FAO (2023), “Pathways towards lower emissions” (GLEAM): livestock 6.2 GtCO2e/year (~12%); cattle 3.8 Gt (62% of the livestock total). fao.org
10. IEA (2025-2026), “Energy and AI”: data centers 415 TWh (2024) toward ~945 TWh (2030), +15%/year; renewables ~50% of growth; AI unlocks up to 175 GW of transmission and cuts blackouts by 30-50%. iea.org
11. Pew Research / IEA (2025): US data center mix 2024 (gas >40%, renewables ~24%, nuclear ~20%, coal ~15%); American Action Forum: gas connection costs ~10x lower. pewresearch.org
12. USDA ERS / World Population Review: US beef consumption ~57.6 lb (~26 kg) retail, ~38 kg availability; Italy ~21 kg apparent (Ismea). worldpopulationreview.com
13. Meta (2024), Llama 3.1 model card: 11,390 tCO2e location-based, 30.84M GPU-hours. huggingface.co