AI Carbon Footprint Calculator - For Businesses

Introduction

AI tools are now embedded in most business workflows, from drafting emails to reviewing contracts. But unlike other emission sources, the environmental cost of AI is opaque and varies hugely depending on model and usage.

To help, we've built a calculator to give you a practical estimate of the carbon, energy and water footprint of common AI use cases, so you can make more informed decisions about how your team uses these tools.

It's based on the latest published research, but in general information is sparse - so treat these numbers as indicative rather than precise! Our full methodology is published below.

Select a model, choose a use case, and set a frequency to see your results

Why should businesses care about the emissions of AI?

For most businesses, AI usage is a small but growing part of the overall carbon footprint. Right now, individual queries are low-carbon at the task level, but that picture can change when you look at it across a team, across a year, and against the backdrop of rapidly increasing AI adoption.

The real concern is what researchers call the rebound effect: as AI becomes cheaper and more capable, businesses use it for more tasks, at greater frequency, and in more complex ways. Total emissions grow even as each individual query becomes more efficient. Analysis from MIT highlights this as the central environmental risk - not the carbon cost of any single prompt, but the cumulative effect of embedding AI deeply across workflows.

There are also three specific reasons this matters for businesses already measuring their carbon footprint:

• ‍It sits in Scope 3. AI tool usage - whether through a paid subscription or API - is an indirect emission from a purchased service. As Scope 3 reporting requirements tighten (particularly under CSRD and for companies pursuing SBTi targets), AI usage is likely to attract scrutiny.‍
• The data gap is a problem. Unlike most software, AI providers don't surface per-query emissions data in their dashboards. That makes it uniquely difficult to quantify. Tools like this calculator exist precisely because the transparency isn't there yet - and we think that needs to change.‍
• Model choice matters more than most people realise. According to our methodology, the energy footprint of a task can vary by 5x or more depending on which model you use. Choosing an efficient model for routine tasks isn't just a cost decision; it's also a carbon one.

For a deeper look at the macro picture - including how AI compares to human-completed tasks and what we're calling on AI providers to do - read our full article: The Carbon Footprint of AI.

How does this calculator work? Our methodology

This calculator estimates the energy, carbon and water impact of AI inference - the act of running a query through a model - for common business use cases.

Energy per task

The energy calculation is calibrated to GPT-4o as a baseline, using an empirical figure of approximately 0.42 Wh per "standard" query (Hao et al., 2024). This accounts for the full system: AI accelerators, host CPUs, cooling infrastructure and idle capacity - not just the raw GPU compute.

Model usage can be measured in input tokens (the text you write in, for example) and output tokens (the response). Using EpochAI's analysis, we use 500 output tokens for a "standard" query, and an assumed figure of 100 input tokens - this provides the baseline. Then for each use case, we estimate a realistic number of input and output tokens. Output tokens are weighted four times more heavily than input tokens in the energy model, reflecting the higher compute cost of autoregressive generation versus input processing. This weighting is an assumption based on the published API pricing for each model (typically x4-5 for output tokens vs. input tokens), which is likely to reflect underlying energy usage.

Model multipliers

Different models consume different amounts of energy per token, depending on model size, hardware efficiency and how they're deployed. We express this as a multiplier relative to the GPT-4o baseline:

• GPT-4o mini and Claude 3.5 Haiku are both estimated at roughly 0.35–0.5× GPT-4o (smaller models with lower per-token compute requirements).
• Gemini 2.0 Flash is 0.6× GPT-4o, based on Google's own published data (Google, 2025), which puts the median Gemini prompt at 0.24 Wh.
• GPT-5 (auto) is estimated at 1.2×, reflecting its adaptive routing (most queries go to an efficient fast sub-model, with the deeper reasoning model reserved for complex tasks).
• Claude Sonnet 4.6 is estimated at 1.8×, based on third-party benchmark data (Devera, 2025). Anthropic has not published per-query energy figures.

Where figures are published by the model provider, we use them. Where they are not - which is most models - we use third-party benchmarks and clearly flag estimates as such. This lack of provider transparency is itself a sustainability issue.

Carbon intensity

Carbon emissions are calculated using the US average grid carbon intensity of ~350 g CO₂e/kWh, from the EPA's latest dataset. The calculator assumes US-based AWS data centre infrastructure, which is where the majority of major AI models are served. Grid intensity varies by region and continues to decline as grids decarbonise - this figure will be updated annually.

Water usage

Water consumption is estimated at 0.5 L/kWh. The major providers estimate 0.15-0.30 L/kWh for AI usage but the general industry average is higher at 1.8 L/kWh.

What this calculator doesn't capture

This tool is designed to be practical and accessible, not exhaustive. In particular, it does not capture emissions from model training - this is notoriously hard, because those "capital cost" style emissions need to be spread across the number of queries received by each model, which varies hugely between models and over time.

It also doesn't include hardware manufacturing emissions, network transmission, end-user device energy, or model fine-tuning.

Frequently asked questions

How much carbon does ChatGPT produce per query?

For GPT-4o - the default model for most paid ChatGPT users - a standard email draft produces roughly 0.15 g CO₂e per query. A more complex task like reviewing a 100-page report would produce around 3 g CO₂e. These figures use US grid carbon intensity and cover inference only. GPT-5, which uses adaptive reasoning, produces similar figures for simple tasks but significantly more for complex ones.

Which AI model has the lowest carbon footprint?

Among the models covered by this calculator, Google Gemini 2.0 Flash has the lowest published energy figure - 0.24 Wh per median prompt (Google, 2025) - which works out to roughly 0.6× the energy of GPT-4o for a comparable task. The caveat is that we can't be certain that these figures are truly comparable, as in each case they are based on a "standard" or "median" query and the assumptions behind what counts as such may be different. Claude 3.5 Haiku and GPT-4o mini are estimated to be similarly efficient. For routine business tasks that don't require the most capable model, these lighter options are meaningfully lower carbon.

Is AI bad for the environment?

At the individual task level, AI is often surprisingly low-carbon - sometimes lower than the human equivalent, particularly for text-based work. The bigger concern is the rebound effect: AI makes many tasks faster and cheaper, which tends to increase total volume of work done, and therefore total emissions. The environmental risk isn't the carbon cost of any single prompt; it's the aggregate effect of AI being embedded across millions of workflows at scale.

Should AI usage appear in my Scope 3 emissions report?

Paid AI subscriptions (ChatGPT Plus, Claude Pro, Google One AI) sit in Scope 3, Category 1 (purchased goods and services). API usage sits in the same category. In practice, most businesses currently have no way to quantify this precisely - AI providers don't publish per-query emissions, and usage data is rarely surfaced in billing dashboards. Until that changes, tools like this calculator can provide a reasonable estimate for disclosure purposes. If you need a defensible figure for a formal report, talk to us.

What's the difference between training emissions and inference emissions?

Training is the one-off (or periodic) process of trainingthe model on large datasets, which is extremely energy-intensive. Inference is what happens every time you run a query: this is what this calculator estimates. For models deployed at the scale of ChatGPT or Gemini, inference now accounts for the majority of lifetime emissions, because billions of queries cumulatively outweigh the training cost. For newer or less widely deployed models, training can dominate.

How do I include AI in my company's carbon footprint?

The simplest approach is to estimate your team's usage (queries per day, use cases, model used) and run those figures through a calculator like this one to get an annual kg CO₂e figure. That can then be added to your Scope 3 inventory. For a structured carbon footprint, Seedling's platform can help - get started here.

‍