AI Green
Scorecard
Rating the world's largest AI companies on carbon emissions, water usage, and transparency.
Explore the dataScroll
The reality
Every query has a cost.
Every model, a footprint.
Data centres consumed 415 TWh of electricity in 2024 — more than the UK — and are on track to double by 2030. Yet most AI companies refuse to disclose their individual impact.
415 TWh
consumed in 2024
60%
powered by fossil fuels
4 of 11
companies publish no data
The Scale of the Problem
The numbers
don't lie.
consumed by global data centres in 2024
More than the entire UK uses in a year.
projected by 2030 — up 128%
Adding Japan's entire electricity grid to the internet.
of new data centre power from fossil fuels
Despite record clean energy investment.
more energy: Claude Opus vs Google Search
One advanced AI query powers a bulb for 10 minutes.
In Context
AI vs the world's electricity grids.
Annual electricity consumption in terawatt-hours (TWh)
Sources: IEA Energy and AI Report 2025 · IEA World Energy Statistics 2024
Energy Per Query
Not all AI is equal.
How much electricity each request really costs
0.0035 Wh
Send an email
baseline
0.30 Wh
Google Search
85× email
2.9 Wh
ChatGPT query
829× email
10 Wh
GPT-4 query
2,857× email
46 Wh
AI image gen
13,143× email
200 Wh
Video generation
57,143× email
To put it simply: one ChatGPT query uses about 10× more electricity than a Google Search. One AI-generated image uses more energy than leaving a lightbulb on for three hours.
Sources: Goldman Sachs Research 2025 · IEA Energy and AI 2025 · Luccioni et al. (2023)
The Rankings
Company Scorecards
11 AI companies scored across six environmental dimensions. Missing data means Not Rated.
Mistral AI
Set a new industry standard by publishing a peer-reviewed lifecycle analysis (LCA) with Carbone 4 and ADEME — a first among pure-play AI companies. Mistral Large 2: 20,400 tonnes CO₂ lifecycle, 281K m³ water. Per query: 1.14g CO₂ and 45mL water. Hosts on French data centres running on 90%+ low-carbon electricity (nuclear + hydro). Building own facility to lock in nuclear energy access. Total emissions just 20,400 tonnes — a fraction of US-based peers.
Massive AI infrastructure investment for LLaMA models. Net Scope 1 & 2 are remarkably low (~50K tonnes). 100% renewable matched with 29 GW contracted. Industry-leading WUE of 0.18 L/kWh. Restored 1.6 billion gallons of water in 2024. Targeting net zero across entire value chain by 2030. All data centres LEED Gold certified or higher.
Committed to being carbon negative by 2030. Scope 1 & 2 down 29.9% from baseline, but Scope 3 up 26% due to supply chain expansion. Matched 100% electricity with renewables. Contracted 34 GW clean energy across 24 countries. Secured 22M tonnes of carbon removal deals. New zero-water cooling design saves ~125M litres/year per facility.
Down 60% emissions from 2015 peak. 100% renewable for all global facilities. Scope 1 & 2 are negligible (58K tonnes). Suppliers sourced 18 GW clean energy, avoiding 21.8M tonnes CO₂. Uses 99% recycled rare earth and cobalt. Data centre footprint is relatively small at 2.3 TWh. On track for carbon neutral across entire footprint by 2030.
Operates one of the world's largest data centre networks. Data centre electricity doubled since 2020 to 30.8 TWh. Emissions up 51% from 2019 baseline despite aggressive clean energy investment. Signed 8 GW of clean energy contracts and pursuing nuclear SMRs with Kairos Power.
Powers nearly all AI training hardware. Scope 3 is 96.6% of total at 6.9M tonnes (supply chain). Achieved 100% renewable energy for offices and data centres. Blackwell GPUs are 50x more energy efficient than CPUs for LLM inference. However, Greenpeace ranked NVIDIA last on AI supply chain decarbonisation, giving it an F for transparency.
World's largest cloud infrastructure and largest corporate renewable energy buyer for 5 consecutive years. 600+ renewable projects in 28 countries. Total company emissions remain the highest at 68.25M tonnes due to massive logistics and supply chain. AWS WUE of 0.18 L/kWh is industry-leading. Investing $20B in Pennsylvania data centres near nuclear plants.
Anthropic
AI safety-focused company behind Claude. No standalone sustainability report published. Runs on Google Cloud and AWS infrastructure. Third-party estimates: Claude 3 Opus uses ~4.05 Wh per query (~1.80g CO₂), while Claude 3 Haiku uses ~0.22 Wh (~0.10g CO₂). Has pledged to cover 100% of grid upgrade costs and invest in water-efficient cooling.
🔒
No public data available
Zero transparency disclosed
OpenAI
Microsoft partnership
Creator of ChatGPT with 700M weekly users. No sustainability report published. GPT-4 training estimated at 12,000-15,000 tonnes CO₂ using ~50 GWh. ChatGPT inference uses ~340+ MWh/day. The $500B Stargate project plans 5+ data centre sites with ~7-10 GW capacity and SMR nuclear reactors. DitchCarbon rates them 23/100.
🔒
No public data available
Zero transparency disclosed
Cohere
Enterprise AI company focused on language models. No sustainability report, no CDP disclosure, no public emissions data. Complete environmental opacity as of March 2026.
🔒
No public data available
Zero transparency disclosed
xAI
Elon Musk
Operates the Colossus supercomputer in Memphis (200,000+ GPUs). Draws 150-250 MW currently, expanding to 2 GW — 40% of Memphis's average daily consumption. 35 gas turbines emit 1,200-2,000 tonnes NOx/year. Plans to use up to 1M gallons of water daily. Located in a predominantly Black, low-income community. No sustainability report. No environmental disclosures.
🔒
No public data available
Zero transparency disclosed
How We Score
Methodology
6 scoring dimensions
Carbon Emissions
Scope 1/2/3 reporting & year-over-year reduction
Renewable Energy
Verified renewable percentage of operations
AI-Specific Reporting
Per-query energy/CO₂, AI/ML workloads disclosed
Water Usage
WUE score, lifecycle water accounting
Power Efficiency
PUE — compute delivered per unit of power drawn
Transparency
CDP, third-party audits, reporting regularity
At a Glance
How they compare
Scored across six environmental dimensions. The gap between leaders and laggards is significant.
The Hidden Cost
Energy per query
A single Claude 3 Opus query uses 169× the energy of a Google search. Model choice is a genuine environmental decision.
Claude 3 Opus
1.8g CO₂
Mistral Large 2
1.14g CO₂
GPT-4o
0.13g CO₂
Gemini
0.1g CO₂
Claude Haiku
0.09g CO₂
Google Search
0.01g CO₂
Sources: CarbonCredits.com · Sustainability Magazine · Mistral AI LCA (Carbone 4/ADEME). Estimates vary by prompt length.
What This Actually Means
Put it in perspective.
Abstract terawatt-hours don't feel real. Here's what AI energy use looks like in everyday terms.
1 ChatGPT query
= A lightbulb on for 10 min
~0.3 Wh · 0.13g CO₂
1 Claude Opus query
= Boiling a kettle
~4.05 Wh · 1.80g CO₂
1 image generated
= Charging your phone 2%
~2.9 Wh · 1.3g CO₂
Training GPT-4
= 500 return flights London–NYC
~50 GWh · 12,000 tonnes CO₂
ChatGPT per day
= 37,000 US homes powered
340+ MWh · 700M weekly users
Google's AI data centres
= Half of Ireland's electricity
30.8 TWh in 2024
Real Questions
What
should I
actually do?
Do I actually need the biggest model?
For most tasks — no. Claude Haiku uses 18× less energy than Opus with similar quality for simple queries.
Is my company's AI provider transparent?
OpenAI, Cohere, and xAI publish zero environmental data. Mistral and Google are leading on disclosure.
Do carbon credits mean a company is clean?
Not necessarily. Microsoft's Scope 3 emissions rose 26% while claiming 100% renewable energy.
What difference can I actually make?
Collectively, model choice matters. 700M ChatGPT users switching to lighter models would save millions of tonnes annually.
A Deeper Question
Why this matters.
The Lord God took the man and put him in the Garden of Eden to work it and take care of it.
Genesis 2:15
Long before the first data centre hummed to life, a mandate was given: tend the garden. This ancient call to stewardship is not opposed to technology — it invites us to wield our tools with wisdom.
As Christians, churches, and organisations increasingly adopt AI, there is a quiet cost being overlooked. Every query draws on energy and water — resources that belong to a shared creation.
The data on this page is imperfect. Many companies refuse to disclose their impact at all. That silence, too, is a kind of answer.
Questions worth asking
- Does our AI provider publish an annual sustainability report?
- Are they genuinely renewable, or just buying offsets to claim credit?
- Could we achieve the same outcome with a smaller, efficient model?
- Are we using AI intentionally, or without reflecting on its cost?
This scorecard was inspired by a reader named Jacob — who asked whether AI companies were taking environmental stewardship seriously, especially for churches and Christians adopting these tools.
Take Action
What you can do.
Individual choices add up. Here's where to start.
01
Choose greener providers
Prefer companies with real sustainability reports, not just pledges. Meta, Microsoft, and Mistral lead on transparency. Avoid companies publishing nothing at all.
02
Right-size your AI use
Use the smallest model that does the job. Claude Haiku, GPT-4o mini, or Gemini Flash use a fraction of the energy of flagship models — often with comparable results.
03
Ask for transparency
Email your AI provider. Ask what their Scope 3 emissions are. Ask for a sustainability report. Collective demand drives corporate behaviour.
04
Offset and reduce together
Calculate your team's AI carbon footprint using tools like ML CO₂ Impact. Then reduce first, offset second — not the other way around.
Read the source reports