What AI costs the planet.
Explainers, company deep-dives, measurements and incidents. Every post sourced, every estimate labelled.
Explainers
The units, boundaries and jargon behind every AI footprint number, in plain words.
How much electricity does one AI prompt really use?
Google measured 0.24 Wh for a median Gemini prompt; OpenAI says 0.34 Wh for ChatGPT. Here is what those numbers include, what they leave out, and why a long document can cost a hundred times more.
Why two honest water numbers can differ 170 times
Google says a prompt uses 0.26 mL of water. Mistral says a response uses 45 mL. Neither is wrong. They draw the boundary in different places, and the boundary is the whole story.
415 TWh today, 945 by 2030: the IEA numbers behind the data-centre panic
The International Energy Agency puts global data-centre electricity at 415 TWh in 2024, about 1.5% of world demand, and projects 945 TWh by 2030. Here is what sits inside those figures and why the range around them is wide.
Data centres used 4.4% of US electricity in 2023. 12% by 2028?
Berkeley Lab's report for the Department of Energy is the most cited US figure: 176 TWh in 2023 and a 2028 range from 325 to 580 TWh. What explains a range that wide, and what would each end mean for the grid.
The Dalles fought 13 months to hide how much water Google used. It was 29% of the city's.
When The Oregonian asked how much water Google's data centres in The Dalles consumed, the city sued the newspaper, with Google paying the legal bills. The settlement revealed 355 million gallons in 2021, nearly triple 2017.
Google's 10.9 billion gallons: the water story inside the AI boom
Google's freshwater consumption rose 34% in 2025 to 10.9 billion gallons, more than double 2021, with data centres responsible for most of the increase. The company replenished about 78% and targets 120% by 2030.
Materials & hardware
Chips, copper, concrete, diesel and e-waste: the footprint before the first prompt.
The 164 kg of CO₂ inside every H100
NVIDIA's product carbon footprint puts an eight-GPU H100 baseboard at about 1,312 kg of CO₂e, roughly 164 kg per GPU, with memory the largest share. Multiply by a fleet and the manufacturing footprint rivals a year of operation.
17 kg per GPU: the e-waste wave of 2030
A Nature Computational Science study projects generative AI will produce up to 2.5 million tonnes of e-waste a year by 2030, a thousandfold increase on 2023. The index turns that into a figure per company: tonnes of servers that will be retired within five years.
Why Mistral ranks first, and what it still does not say
Mistral tops the index at 75 because it is the only frontier lab with a peer-reviewed life-cycle assessment and it serves from one of the world's cleanest grids. It also publishes no company inventory, no WUE and no dated targets.
Google: the best disclosure in the index, attached to the worst trend
Google measures its AI footprint better than anyone: 0.24 Wh, 0.03 g and 0.26 mL per median prompt, fleet WUE, hourly carbon-free share, electricity by year. The same reports show emissions up 18%, electricity up 37% and water up 34% in 2025.
Gemini's 0.24 Wh: what Google measured, and what it left out
Google's August 2025 paper is the only fleet-wide, methodology-backed per-prompt measurement any AI company has published. It is also a median, self-reported, and silent on training. Here is how to read it.
ChatGPT: 0.34 Wh and no methodology
OpenAI's only per-query figure is a sentence in a blog post. Epoch AI's independent estimate for GPT-4o lands close to it for short prompts and far above it for long ones. The index records both, and the difference between a number and a disclosure.
xAI's 35 turbines against a permit for 15: the Memphis record
The Shelby County Health Department permitted 15 gas turbines at Colossus as backup power. Aerial imagery counted 35 on site, 24 of them running. What the records show, what each side says, and why the index treats it as the heaviest conduct deduction on the site.
Colossus 2: 46 gas units against a permit for 41, within two months
xAI's second campus in Southaven, Mississippi, was permitted for 41 turbines in March 2026. Records obtained by SELC showed 19 more installed between late March and early May, taking the site to 46 units and more than 500 MW. Elon Musk says the cluster will reach 2 GW.
EU AI Act Annex XI: energy must be documented, not disclosed
From August 2025, providers of general-purpose AI models in Europe must keep technical documentation that includes the model's energy consumption, known or estimated. They do not have to publish it. Here is what the rule requires, what it allows, and what it leaves out.
The GPAI Code of Practice: what signing it means for energy reporting
The European Commission's General-Purpose AI Code of Practice offers model providers a voluntary route to compliance with the AI Act. Its transparency chapter includes a model documentation form with an energy field. What the form asks, who has signed, and why it is not a disclosure regime.
How to read an index score
A company's score is a weighted sum of six pillars, each 0–100, with a confidence grade attached. Here is what the number means, why a 70 with a D grade is not a 70 with an A grade, and the three bands the index uses.
How we estimate what companies won't tell us
161 of the 210 key figures in the index are estimates or derivations, because 27 of 30 companies publish nothing per prompt and many publish nothing at all. Every estimate shows its working. Here are the rules, the factors and the uncertainty.