The growing demand for resources to train and run language models presents new challenges for global digital infrastructure. A study by the Ökoinstitut predicts that by 2030, the energy consumption of AI data centers will increase so much that it will further jeopardize climate targets. The future also looks grim regarding the water consumption and e-waste generated by generative AI. Part of the problem is that major tech companies are often reluctant to provide information on the resource consumption of their new tools.
However, the resource demands of language models vary significantly depending on factors like the model’s size. To help users make informed decisions, the GenAI Impact team created EcoLogits, a tool that assists in selecting the right language model. We spoke with co-founder Caroline Jean-Pierre to learn more.
Measurement directly on the API or in the browser
EcoLogits is a freely available open source application that developers can integrate into their products. “If you, as a developer, send requests to OpenAI in your programmes, you can use EcoLogits [… to measure resource consumption …].” This is because EcoLogits directly accesses the APIs that companies such as OpenAI, Meta or Google offer as interfaces to their language models for its estimates.
What is an API?
The abbreviation API stands for Application Programming Interface and describes an access point to programmes and computer systems. Programmers of a new or existing programme can easily and securely access data from other programmes.
Using AI systems as an example, this means something like this: A developer of a language learning app wants to integrate an AI chatbot into their programme. And he can create a connection to systems such as ChatGPT, Perplexity or Google Gemini via an API.
EcoLogits can use the data transmitted via the API connection to generate estimates of resource consumption. As of July 2025, this is already possible for energy consumption, electricity consumption, greenhouse gas emissions and hardware resource consumption. GenAI Impact would like to add water consumption as a key figure in a future version.
“While we were developing EcoLogits, we also recognised an interest from the general public. That’s why we decided to also publish the EcoLogits Calculator”.
How big is the resource consumption of GenAI anyway?
Most people have heard by now that ChatGPT, for example, consumes more electricity than a Google search. However, increasing energy consumption is not the only problem with GenAI from an environmental perspective.
First of all, the energy consumption of AI data centres is particularly critical as they run around the clock. If data centre operators rely on green energy, they must also rely on fossil fuels to ensure continuous performance. The expert Ralph Hintemann attributes a large proportion of CO2 emissions in digitalisation to AI data centres in the coming years.
In addition to this energy consumption, there is also a great thirst for water that arises from the cooling of data centres.
As AI applications place heavy demands on data centre hardware, the need for graphics cards and server infrastructure is also increasing. This is also reflected in more electronic waste in the long term.
The currently established language models also focus on a few languages such as English or Mandarin. Minority languages are thus marginalised, which significantly worsens participation in AI models in the global South. Yet it is precisely people in poorer countries who are suffering the consequences of the AI boom.
The EcoLogits Calculator is a free web app available on HuggingFace, making it accessible to users without coding knowledge. It allows users to compare different language models and select from predefined use cases, such as generating a 170-token email or creating a 15,000-token app code. As we previously noted, resource consumption increases with the number of output tokens requested. For users who know the desired text length, the calculator offers a “Tokens estimator” to approximate token consumption and a “professional mode” for a more precise calculation.
After a short calculation time, the AI resource calculator spits out information about the energy consumption, the CO2 equivalents released and the amount of precious metals and minerals required. In the same step, GenAI compares these rather abstract values with the range of an electric car or other parameters.
Life cycle analyses for language models
Caroline Jean-Pierre also refers to other tools for calculating the emissions of language models. “But we don’t just want to focus on energy and the CO2 emissions released. Instead, we want to get as close as possible to a life cycle analysis of language models.” However, GenAI Impact only focuses on the inference – in other words, the ongoing operation – of the respective language model in order to present more information about the use of the models. Resource-intensive processes such as the training of language models, which takes place for months in huge data centres, are therefore left out of the methodology used.
The French AI company Mistral presented such an analysis in July 2025. In a unique analysis that has been cross-checked by independent bodies, Mistral shows how resource-intensive the entire life cycle of a language model is. This also allows us to see how large the share of inference is in the total emissions of an AI model.
How does GenAI Impact determine this data?
If we look around the EcoLogits Calculator, we see repeated references to the inaccuracy of the results. Shortly after our conversation with Caroline Jean-Pierre in July, GenAI Impact even added deviations to the web tool that represent the minimum and maximum estimated values. So why can’t we calculate energy consumption accurately?
“There are open models in which we can measure resource consumption. And they show a linear dependency between parameters and energy consumption per token. Together with leaks and estimates of the development costs of certain language models, we are trying to transfer these results to large language models.”
How can we ensure a green digital future?
Growing e-waste, carbon emissions from AI, data centre water usage—is rampant digitalisation compatible with a healthy planet? Our latest project explores how digital tools and services can be developed with sustainability in mind.
GenAI Impact presents these estimates according to open data principles. Other companies or researchers can check and correct them accordingly. “Everything is open and transparent – and we also provide error probabilities to make the accuracy of the results more transparent.”
When asked about possible reactions from big tech companies to her estimates, Caroline Jean-Pierre also replied: “As soon as the big tech companies disclose the real data, we are happy to change our estimates.” With Mistral, at least one established AI manufacturer has now disclosed measured values. A comparison shows that the values given are within the estimated range of the EcoLogits Calculator.
For now, however, the GenAI Impact team is looking for volunteer programmers to be able to offer EcoLogits in the Java programming language in the future. GenAI Impact is also currently still in need of support in the form of funding. More information can be found on the company’s homepage.
We would like to thank Caroline Jean-Pierre for the interview!
