Sustainable AI: How to measure and cut AI emissions

What is sustainable AI?

Sustainable AI is the practice of designing, choosing and running artificial intelligence so that its environmental impact is known and kept as low as the work allows. It covers the electricity used to run models, the carbon emitted to generate that electricity, the water used to cool data centres, and the emissions from making the hardware.

Sustainable AI has two meanings. The first is the sustainability of AI itself: how much energy, water and carbon AI systems use, and how to reduce it. That is what this page covers. The second is AI for sustainability: using artificial intelligence and machine learning to help meet sustainability goals, such as flood forecasting, biodiversity monitoring, resource management, balancing renewable energy on the grid or improving energy efficiency in buildings. Work like this is often framed around the UN Sustainable Development Goals. Both matter, and AI for sustainable development only adds up if the environmental sustainability of AI itself is under control.

You will also see the term green AI. Green AI usually refers to research on making AI models themselves more efficient, such as smaller architectures and cheaper training. Sustainable AI is wider. It applies to any organisation that uses AI, including the many that never train a model and only call one through an API, a cloud service or a SaaS tool.

For those organisations, the footprint is driven by everyday choices: which model a team picks, how many tokens each request uses, and where the work runs. Those choices are also what sets the AI bill, which is why cost and carbon tend to move together.

Why sustainable AI matters now

AI energy consumption is one of the fastest-growing sources of electricity demand in the world of technology. The International Energy Agency expects electricity use by data centres worldwide to roughly double to around 950 TWh by 2030, with AI-focused data centres growing fastest. The IEA also expects data centre emissions to double, to around 350 million tonnes by 2035, so growth in AI usage feeds into carbon emissions and climate change unless the efficiency of each request improves.

For businesses, the pressure comes from three directions. AI spend is rising fast, sustainability teams have to report the greenhouse gas emissions it causes, and customers and investors increasingly ask how AI fits with the company's sustainability goals. A clear view of AI's environmental footprint answers all three, turns usage data into insights teams can act on, and usually points to savings at the same time.

Why the environmental impact of AI is hard to see

Most organisations use AI through several routes at once: cloud AI services on AWS, Azure and GCP, direct API access to model providers, and AI tools built into SaaS products. Each route reports usage in its own way, usually as tokens or spend. Very few report energy, carbon or water, and where providers do publish figures, the methods differ, so the numbers cannot be added up.

The impact itself has three parts:

• Energy and carbon at run time: Electricity used by GPUs, CPUs and the data centre, multiplied by the carbon intensity of the local grid. See how much energy AI uses.
• Water: Used to cool data centres and to generate their electricity. See how much water AI uses.
• Embodied emissions: The carbon from making GPUs and servers, spread over their working life. This is part of the carbon footprint of AI.

Step 1: Map where your AI usage happens

You cannot reduce what you cannot see, so start with an inventory. List every provider, tool, chatbot and AI agent your teams use, and find where the usage and billing data for each one is kept.

Typical sources are:

• Cloud billing exports, such as the AWS Cost and Usage Report, for Amazon Bedrock, Azure OpenAI and Vertex AI
• API usage logs from model providers such as OpenAI, Anthropic and Google
• Admin reports from per-seat tools such as ChatGPT Enterprise, Claude Team or Gemini in Google Workspace

The useful level of detail is tokens by model and token type (input, output, cached). Spend alone is not enough, because the same spend on two different models can mean very different energy use. Your FinOps team often holds most of this data already.

Step 2: Measure energy, carbon and water per token

To measure the energy and emissions of AI model usage in the cloud, multiply your token counts by per-token factors for each model, then convert energy into carbon and water for the region where the work ran.

A reliable estimate includes:

• Energy per token for the specific model, token type and processing tier (real-time, batch, priority)
• Data centre overhead, through PUE
• A share of the energy used to train the model
• Embodied emissions of the server hardware
• Grid carbon intensity for the region and, where possible, the hour

Few providers report energy per request to their customers, so most figures here are estimates. Good estimates say so, and show their uncertainty as a range rather than a single number. Starting from energy also gives a more consistent number than starting from carbon, since carbon and water can then be worked out for each region.

Step 3: Compare AI models before you deploy

The cheapest moment to cut AI emissions is before a model goes into production. Engineering teams can compare the environmental impact of different models in three ways:

• Per-token energy for the actual task: Run a sample of real prompts through each candidate model and record input and output tokens. A model that writes longer answers can use more energy overall, even if it costs less per token.
• Quality per unit of energy: Check whether a smaller model, such as a “mini”, “Flash” or “Haiku” tier, meets the quality bar. For classification, extraction and short answers it often does.
• Cost and carbon side by side: Providers price by the token, and the energy ranking often follows the price ranking: GPT-6 Astra uses about 49 times the energy of GPT-6 Luna per output token, at around 100 times the list price. Putting both in one view makes the trade-off clear.

The AI energy calculator estimates a workload's footprint on a chosen model, so you can try the same workload on different models before you commit.

Step 4: Cut AI emissions without cutting output

Most of these changes lower the AI bill as well as the footprint, so sustainability and cost efficiency pull in the same direction:

• Right-size the model: Keep the largest models for work that needs them and route simpler tasks to smaller ones.
• Trim tokens: Ask for shorter answers and structured output where a long response adds nothing.
• Cache what repeats: Reusing a cached system prompt or shared context avoids processing the same tokens again.
• Batch what can wait: Batch APIs serve requests together, which is cheaper and uses less energy per request.
• Choose cleaner regions: Where latency and data rules allow, run workloads in regions with lower-carbon grids.
• Make it visible: Put energy, carbon and water next to AI spend in the reports teams already use.

Step 5: Report AI emissions

AI usage bought from cloud and SaaS providers usually falls under Scope 3 emissions, mainly purchased goods and services. As AI spend grows, it is becoming a visible line in sustainability reports prepared under frameworks such as CSRD.

For reporting, the key is one method applied across every route, so figures from cloud AI, direct APIs and SaaS tools can be added together and compared year on year. Record which method and factor version you used, and keep the ranges alongside the central figures.

Read more about Scope 3 cloud emissions for CSRD.

Make sustainability part of AI governance

Sustainable AI works best when it sits inside the rules a business already has for AI. Many organisations have responsible AI principles covering privacy, fairness, transparency and other ethical questions. Adding environmental impact to that list makes sustainable practices part of how AI tools are approved and managed, not an afterthought.

• Set a default model for each type of task, with the largest models kept for work that needs them
• Ask AI vendors for energy and emissions data, and record the method behind any figures they give
• Track AI energy, carbon and water per team or product next to AI spend, so the people making decisions see them together
• Review the footprint when new AI applications or agents go live, not once a year

Transparency runs through all of it. Teams can only improve what they can see, and reports are only credible when the method behind the numbers is clear.

The future of sustainable AI

AI technology is getting more efficient. Smaller models, quantisation and newer GPUs all deliver more output for each unit of power, These innovations matter, but AI development and adoption are growing faster: the IEA still expects data centre electricity use and emissions around the world to keep rising.

That makes measurement the foundation of any credible plan for emissions reductions. Businesses that can see the energy, carbon and water behind each AI workload can set targets, test changes and show progress. Those that cannot are left with high energy consumption they cannot explain and climate commitments they cannot back up.

Regulation is moving the same way. Sustainability reporting rules such as CSRD already ask for Scope 3 emissions, and AI is a growing part of them. Getting AI emissions management in place now is easier than retrofitting it once the numbers are large.

How Greenpixie helps

GPX AI estimates the energy, carbon and water of AI across providers, SaaS tools and agents using one method. It works from any usage report that shows tokens by model and token type, and covers 500+ models.

• 32 open-weights LLMs benchmarked on H100 and B200 GPUs, extended to other models with predictive models
• Every value given as a median with an interquartile range
• Data centre overhead, training energy and embodied hardware emissions included
• Results cross-checked against independent research and metered energy data from enterprise clients

Carbon, water and embodied emission factors come from Greenpixie's cloud methodology. The methodology has been independently verified in accordance with ISO 14064-3, prepared with reference to ISO 14064-1, and designed to be in line with the GHG Protocol Corporate Standard. The AI per-token energy layer builds on these verified factors and is validated against independent research and metered energy data.

“This methodology provides actionable data that enables reductions in cost, electricity usage, CO2-eq emitted and water consumed in cloud and Software as a Service (SaaS).”

Frequently asked questions

Is there an AI that is sustainable?

No AI model has zero impact. Every model uses electricity, water and hardware to run. What makes AI more sustainable is how it is chosen and used: the smallest model that does the job well, fewer wasted tokens, batched and cached requests, cleaner grid regions, and a measured footprint so teams can see the effect of each choice. Smaller models from the major providers, such as the “mini”, “Flash” and “Haiku” tiers, use far less energy per token than the largest ones.

How can AI be sustainable?

AI becomes more sustainable when its footprint is measured and managed like its cost. Map where AI is used, estimate energy, carbon and water per model and token, compare models before deployment, cut waste through right-sizing, caching, batching and region choice, and report the results. The five steps above set this out in order.

How much energy does an AI query use?

A typical chat request to a large frontier model, with 1,000 input and 500 output tokens, uses about 1.6 Wh of electricity, based on Greenpixie's AI methodology (November 2026). That is about what a 10 W LED bulb uses in ten minutes. Smaller models use far less, so the model and the length of the answer matter most. See how much energy AI uses for model-by-model figures.

How can I measure the energy consumption and emissions of AI model usage in the cloud?

Export token usage by model and token type from your cloud bill or API logs, multiply by per-token energy factors for each model, then convert energy to carbon using the grid carbon intensity of the region where the work ran.

What is the most reliable way to estimate emissions from AI workloads on AWS, Azure or GCP?

Use one method across all three clouds, based on tokens by model rather than spend, and include data centre overhead, training energy and embodied hardware emissions. Prefer estimates that show a range, not just one number.

How do teams compare the environmental impact of AI models before deployment?

Run a sample of real prompts through each candidate model, record the tokens used, and compare energy per task alongside quality and cost. Smaller models often meet the bar for simpler tasks.

Do AI emissions count towards Scope 3?

For most organisations, yes. AI bought as a cloud or SaaS service usually sits in Scope 3, mainly purchased goods and services.

Sources

1. Greenpixie AI methodology
2. Horswill, Hunter, Clifford and Hall, Greenpixie's AI Token Methodology: Assessing the Energy, Water and CO2-eq Impact of AI Tokens for Open and Closed Weight Models, arXiv:2609.33965
3. Jegham et al., How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference, arXiv:2505.09598 (2025)
4. Greenpixie Cloud methodology
5. Li, Yang, Islam and Ren, Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models, Communications of the ACM 68(7), 2025
6. IEA (2026), Key Questions on Energy and AI