Is generative AI bad for the environment?
Generative AI costs more per use than older AI because each word requires a full pass through billions of parameters — but a median Gemini text prompt is still only 0.24 watt-hours. It's the demand it created that matters: AI servers are growing 30% a year, versus 9% for conventional ones, and driving nearly half of data center growth.
Why — the first-principles explanation
Generative AI is environmentally distinct from the AI that came before it, and the reason is architectural. A spam filter or a recommendation model runs a small network once and returns a label. A generative model produces output one token at a time, and each token requires a full forward pass through the entire network — often tens or hundreds of billions of parameters, every one of which must be read from memory. A 500-word answer isn't one computation; it's roughly 700 sequential computations, each moving billions of numbers. That's why generation needs GPUs and HBM memory while a spam filter runs happily on a CPU, and why generative AI — not AI in general — is what set off the data center buildout.
But "more expensive than a spam filter" is not the same as "expensive." Google's production measurement puts its median Gemini text prompt at 0.24 Wh, 0.03 gCO2e, and 0.26 mL of on-site water — about nine seconds of television. The architectural penalty is real and the absolute number is still tiny, because chips got fast enough to make hundreds of billion-parameter passes cost less than boiling a teaspoon of water. Note the word text, though: that figure covers text prompts. Image and especially video generation require far more computation per output, and no comparable production measurement has been published for them, so anyone quoting a precise figure for AI video's footprint is estimating.
The honest case against generative AI is therefore not per-query, it's aggregate and structural. Generative models made compute worth buying at a scale nothing else justified. The IEA projects electricity for accelerated servers — the GPU hardware generative AI runs on — growing about 30% per year versus 9% for conventional servers, accounting for nearly half the net increase in data center electricity through 2030. Total data center power roughly doubles to about 945 TWh by 2030. Generative AI didn't make each computation dirty. It made an enormous number of computations worth doing, in a hurry, in a few places.
Two caveats keep this from being a verdict. First, training versus serving: training GPT-3 could directly evaporate about 700,000 liters of freshwater, but that's a one-time cost amortized across billions of queries — for a popular model, cumulative serving eventually dominates. Second, efficiency is moving very fast: Google's median prompt got 33x more energy-efficient in twelve months. Generative AI's footprint is a race between falling cost per query and rising query volume, and so far volume is winning. That's a statement about adoption, not about the technology being inherently dirty.
An example that makes it click
Compare a vending machine to a short-order cook. The vending machine has your snack sitting on a shelf — pressing the button just drops it. That's old-school AI: the answer is essentially precomputed, and retrieving it is cheap.
A generative model is the cook. Nothing is on the shelf. Every sentence gets made from scratch, one word at a time, and for each word the cook re-reads the entire hundred-billion-page cookbook. That sounds insane, and it is why generative AI needs specialized hardware. But here's the twist: the kitchen got so absurdly fast that cooking your meal from scratch now costs about nine seconds of TV. The problem was never the cost of your meal. It's that once cooking became this cheap, everyone ordered — and the town had to build a power plant for the restaurant.
Key facts
- A median Gemini text prompt uses 0.24 Wh, 0.03 gCO2e and 0.26 mL of on-site water — about nine seconds of television (Google technical paper, August 2025).
- Google's median text prompt energy fell 33x and carbon fell 44x over one twelve-month period ending 2025.
- Electricity for accelerated (AI/GPU) servers is projected to grow about 30% annually versus 9% for conventional servers, making up nearly half the net increase in data center power (IEA).
- Global data center electricity roughly doubles from about 415 TWh in 2024 to about 945 TWh by 2030 (IEA).
- Training GPT-3 in Microsoft's US data centers could directly evaporate about 700,000 liters of clean freshwater — a one-time cost, not per query (Li, Yang, Islam & Ren, arXiv 2304.03271).
- Google's published per-prompt figure covers text prompts only; no comparable production measurement exists publicly for image or video generation.
▶ The 60-second explainer (script)
Is generative AI bad for the environment? It's genuinely different from older AI — here's why, and why that still isn't the real problem. A spam filter runs a small network once and returns a label. A generative model builds its answer one word at a time, and every single word requires a full pass through the entire network — often a hundred billion parameters, all read from memory. A five-hundred-word answer isn't one computation. It's about seven hundred of them, back to back. That's why this needs GPUs, and why generative AI specifically set off the data center buildout. But more expensive than a spam filter isn't the same as expensive. Google measured its median text prompt at 0.24 watt-hours — nine seconds of TV. Chips got fast enough that hundreds of billion-parameter passes cost less than boiling a teaspoon of water. Note the word text, though — image and video generation cost far more, and nobody's published production numbers for those. So the real case isn't per-query. It's structural. Generative AI made compute worth buying at a scale nothing else justified. AI servers are growing thirty percent a year versus nine for conventional. Data center power doubles by 2030. Generative AI didn't make each computation dirty. It made an enormous number of them worth doing, fast, in very few places.
What authoritative sources say
People also ask
Does generating an AI image use more energy than text?
Substantially more per output, because image and video models run far heavier computation than a text response. Google's 0.24 Wh figure covers text prompts only, and no comparable production measurement for images or video has been published.
Is generative AI worse than regular AI for the environment?
Per use, yes — generation requires a full pass through the whole model for every token, while a classifier runs once. In absolute terms both are small; generative AI matters because of the buildout it triggered.
Does training or using generative AI cost more overall?
Training is a large one-time cost — GPT-3's could evaporate roughly 700,000 liters of water. But for widely used models, cumulative serving across billions of queries typically overtakes training over the model's lifetime.
Is generative AI getting greener?
Per query, dramatically — Google reported a 33x energy improvement in a single year. Total consumption still rises because adoption grows faster than efficiency improves, which is a story about demand, not dirty technology.