Note: I am not trying to defend AI-Generated art. I am merely trying to disprove myths surrounding its power and water use. It's best to stay reasonable about things you dislike.

_update 20.06.2026: removed the note that training is the biggest drainer. its not; inference is. i also added sources, adjusted wording and added resource usage comparison between models. _

Water use

while it's true that AI uses a considerable amount of water, universally it doesnt have a huge impact. it is mainly prelevant smaller (~ a few thousand citizens) towns and cities where AI data centers exist, although AI's impact is also rather small. In major US states like Phoenix, Arizona and Texas, the water usage is very negligible, where Agriculture consumes around 80% of local water alone. AI Datacenters use ~0.8% - A single golf course uses over 1500 metric ton every single day. More than AI,

On average, a 30 message long AI chat uses around 7,8ml (0,26ml per prompt). that's one single drop. Doomscrolling on tiktok or twitter for an hour, watching Stranger things on Netflix or 50 Cases Of Wii Lost Media on YouTube uses much more than that; around 500ml per hour. though this only counts direct cooling, not the water used in power generation.

AI Companies are starting to implement something called Two-Phase Immersion-Cooling; the fluid actually boils when it touches hot components. the vapor rises, hits a condenser at the top of the tank, turns back into liquid and „rains“ back down. this significantly reduces waterusage. Not all datacenters use it yet due to its complexity.
Another choice is Direct to Chip; this is the most common transition for existing data centers. Cold plates sit directly on the GPUs/CPUs, with liquid circulating through internal micro-channels to pull heat away.

Mostly, the reason why AI water use is still so high is because many existing datacenters still use legacy cooling. Companies are putting crazier chips into old datacenters. You also need power to run the AI, and since majority of the US energygrid STILL uses coal and nuclear, the water use is indirectly higher.

Although, in 2025, Microsoft and Google signed huge deals to use Small Modular Reactors (SMRs, basically nuclear energy) to power data centers. this decouples them from the country's grid.

Power use

It's honestly comparable to water use. Generating one image usses as much power as charging your phone; around 0,002kWh. For contrast, playing a game on a high-end PC for an hour uses 0,6kWh or 450 watt. An AI datacenter may use a little more, but only in short bursts. Gaming's power draw is long and consistent. You'd need to generate 300 images just to match the energy use of some guy in norway playing cyberpunk or someone in maryland playing final fantasy.

Generating a 5 second video takes 1KWh, however. And it's estimated that for training OpenAI's GPT-4 model, it took 50GWh and 100m$, enough to power a city for days.

So, if you were to get everyone who uses AI for research and move them back to Google, they'd have to crawl through at least 10-20 different websites to find an answer that isn't SEO-spam or a Reddit thread from 2012. Between the server power for those 20 sites and the energy your own PC wastes keeping all those tabs open and opening new ones in Firefox, you're actually burning more power than one AI prompt. But uBlock origin does help, yeah.

And, between 2020 to 2025, AI chips and models have been getting 30x more efficient at doing the same shit.

How models compare

Each consumer-available LLM Model uses different amount of energy. At the time of writing (20.06.2026), the most inefficient model might be Claude sonnet 3.7. The most efficient is Gemini 3.5. Here is a comparison of 3rd party estimates:

AI Model	Energy per Prompt	Water per Prompt	Equivalent To
Google Gemini 3.5	0.24 Wh	0.26 mL	9 seconds of watching TV
OpenAI GPT-4o	0.3 Wh	0.32 mL	9.5 seconds of watching TV
OpenAI GPT-5.5	0.34 Wh	~1.0 mL	Powering a laptop for 2 minuutes
Anthropic Claude 3.7 Sonnet	0.84 Wh	~1.5 mL	4 minutes of an LED bulb

For the reasoning LLMs:

AI Model	Energy per Prompt	Water per Prompt	Equivalent To
OpenAI GPT-5/o3	18-40Wh	~60 mL	4 hours of an LED bulb
Anthropic Claude 3.7 (Extended)	3.49 Wh	~5.0 mL	17 minutes of an LED bulb
DeepSeek-R1	23.82 Wh	~35.0 mL	2 hours of an LED bulb

Gemini uses Google's own TPUs tailored explicitly for their AIs, as opposed to buying Nvidia GPUs, which are general-purpose AI chips. Google's TPUs are specialized for one job only, which is why energy usage is drastically lower.

Please note that the Claude estimate is entirely mine. There's no public data anywhere online. I estimated it based on how it does more — can execute commands, has its own filesystem to play with, vastly more reasoning, longer response time — and scaled it based on GPT. Please do not take it as a fact.

idk what to call this section

You mightve seen graphs on social media about how much AI uses.

Yes. This is blatant disinformation. The graph above (with a source from 2019) compares training of AI to air travel with ONE passanger from NYC to San Francisco. This is not at all comparable…
They compare the training of language models with mundance, incomparable things to make the graph scary. Add annual usage of other industries and AI becomes tiny. The 626,000 lbs figure traces back to one specific worst-case Transformer training run with architecture search — not a typical model, and not anything resembling what „training an AI model“ means in 2026 Also, the other graphs are literally useless.

conclusion

genAI isn't inherently more harmful to the environment than all the other tech we use every day; a car produces more co2 than a few daily gemini messages, and walnut production uses more water than ai every day. it is also stupid as fuck to blame individuals for climate change. it's the system that's at fault for everything. don't blame individuals for anything.

corps replacing people with that shit can go bankrupt. and people using it as art replacement should ~~kill themselves~~ kindly stop.

Sources