You’ve no doubt heard that websites, software and AI are consuming ever-increasing amounts of electricity, water and resources. According to estimates, our information technology will account for roughly half of global energy consumption over the coming decades. Experts suggest that data centres will be responsible for the lion’s share of digitalisation’s carbon emissions in the future. This is partly because we need to use fossil fuels to fulfil their ever-growing energy requirements.
Companies that produce their own digital products or operate entirely digitally are particularly well-placed to save resources by using green alternatives. We’ve already reported on green hosting, green coding and other measures companies can take. In this guide, we’ll highlight ways companies and freelancers can track their progress towards a green transformation of their digital tools.
Below, we present tools and online services that allow you to track the energy consumption of software, your website or your AI usage (generative or traditional). To this end, we spoke to Arne Tarara from Green Coding Solutions, amongst other experts. Arne pointed out a number of tools that serve as CO2 trackers for digital services. You’ll find these alternatives at the end of this article.
What exactly is being measured or estimated here?
Before we begin, we need to consider our objective when measuring or estimating.
Track or estimate?
It’s fairly straightforward to measure the power consumption of your own hardware—but things get complicated when it comes to a data centre.
For this reason, in some cases it’s advisable to carry out measurements, whilst in others it makes more sense to extrapolate using known data and apply heuristics.
To make this as transparent as possible, we have added the labels ‘measurements’ and ‘estimates’ to the respective solutions.
What our smartphones, laptops, data centres and the like all have in common is that they are powered by electricity. So, electricity is important to track. There are also solutions that measure the power consumption of hardware and can even compare this with the energy mix where you live. This gives you a particularly accurate picture of an application’s carbon footprint.
In addition to energy requirements, the demand for hardware and water is particularly relevant in data centres. For instance, servers in AI data centres are subjected to such heavy use that they need to be replaced after just a few years and end up as electronic waste in landfills. Meanwhile, their cooling systems require pure fresh water. Because of this, some tools also estimate water consumption.
Tracker for generative AI
Let’s start with generative AI. This refers to all AI technologies used to create text content, images and videos, as well as to generate code. The range of services based in some way on large language models has exploded since OpenAI released ChatGPT in 2022. Many companies feel compelled to integrate generative AI into products, work processes and much more. This is a consequence of the AI boom and the narrative that companies without AI models are no longer competitive.
However, a study by MIT suggests that, in the long run, investments in AI do not pay off for 95 percent of the companies it surveyed. Furthermore, terms such as ‘AI fatigue’ or ‘AI brain fry’ indicate that employee satisfaction can drop significantly if an AI strategy is not implemented effectively.
Generative AI also poses major challenges for global IT infrastructure. The technologies are so computationally intensive that GAFAM (Google, Apple, Facebook / Meta, Amazon, Microsoft) are building new data centres all over the world. This is mostly at the expense of locals, who suffer from noise pollution, water shortages and high energy consumption.
For businesses and freelancers, it therefore makes sense to use generative AI with a degree of caution—or to opt out of an AI strategy altogether. That is why we are starting with tracking solutions that specialise in generative AI.
EcoLogits Calculator estimates the inferences of various language models (Estimates)
The US-based non-profit organisation CodeCarbon offers the EcoLogits Calculator, a comprehensive tool for assessing the environmental impact of text-based GenAI activities. Users can configure exact token lengths (i.e. the length of the generated responses) and specific language models. By considering the location of the data centre an application uses, EcoLogits generates particularly realistic calculations. These include the CO2 emissions, as well as electricity, water and hardware consumption.
The EcoLogits Calculator is available as a free web app via Hugging Face. Alternatively, developers can integrate the GenAI calculator directly into other programmes via an API. We discussed exactly how this works with Caroline Jean-Pierre from Gen AI Impact.
Green Algorithms Calculator (Measurements & Estimates)
The “Green Algorithms Calculator for High Performance Computing” can also be installed. The tool is designed for servers that use SLURM for workload management. However, according to the documentation, the Green Algorithms Calculator can also be adapted for other solutions.
The tool can also be used to monitor large language models running locally. The Green Algorithms Calculator calculates the energy required and the carbon footprint based on CO2 equivalents. You can view the exact results in a web version without installing the software. As this cannot access the hardware’s analysis data directly, the web version estimates the metrics using various parameters.
Tracker for “traditional” AI
Although the distinction is made less and less frequently, alongside generative AI, there are technologies that utilise machine learning on a much smaller scale. The German start-up Dryad, for example, uses artificial intelligence to detect forest fires. It keeps its proprietary models small enough to run on low-cost, local hardware.
Generative and traditional AI—why does the difference matter?
In their book *AI Snake Oil*, authors Arvind Narayanan and Sayash Kapoor pose an interesting question right at the start:
“Imagine an alternative universe in which we humans have no separate words for modes of transport—only the collective term ‘vehicle’.
We use this word for cars, buses, bicycles, spaceships and all other means of getting from A to B. Conversations in this world would be confusing. There are heated debates about whether vehicles are environmentally friendly or not, with nobody realising that some people are talking about bicycles and others about lorries.
There might be a breakthrough in rocket technology, but the media focuses on how fast vehicles have become—so people ring their car dealer to ask when faster models will be available.
Meanwhile, fraudsters have exploited the fact that consumers don’t know what to believe when it comes to vehicle technology, so fraud is rife in the automotive industry.
Now replace the word ‘vehicle’ with ‘artificial intelligence’ and we have a pretty good description of the world we live in.”
This thought experiment illustrates why we should be particularly precise when talking about ‘AI’. Tech blogger Ketan Joshi explains what can happen when we deliberately conflate these terms.
There are cloud-based solutions available for training such models, which provide researchers and businesses with sufficient computing power. Alternatively, these small models can also be trained locally on your own hardware. CO2-Tracker is available for both scenarios:
MLCO2 (Estimates)
The MLCO2 tool is a simple web app where you can set four parameters. The type of hardware used, its runtime, the cloud environment provider, and the region in which the training is processed. MLCO2 then calculates the independent carbon footprint of your ML training, as well as the footprint generated at the respective location. In doing so, the tool takes into account the energy mix of the respective cloud provider.
The developers Victor Schmidt, Sasha Luccioni, Alexandre Lacoste and Thomas Dandres point out that these calculations do not take into account the Power Usage Effectiveness (PUE) of the respective data centres. Users will need to research these figures separately and factor them into their calculations. But bear in mind: according to Jens Gröger, caution is advised when considering the PUE of data centres.
Code Carbon (Measurements)
As a further recommendation, MLCO2 suggests the Code Carbon tool for measuring local AI training. The tool can be installed as a Python package and calculates the power consumption of the CPU, GPU and installed RAM. Code Carbon then correlates these values with the electricity mix of the respective location to obtain more accurate forecasts.
Code Carbon is generally recommended as a tracking solution for local computing operations. However, we are including it in our recommended tools for tracking AI application training, as it includes GPU usage in the calculated carbon footprint. As graphics cards are particularly well-suited to ML training, Code Carbon is also highly recommended in this context.
Tracker for green coding and software
Green coding refers to programming practices that have a positive impact on a software’s energy consumption. As Max Westing from the ‘Green Coding’ project explains, programmers can integrate energy-efficient libraries into their programmes or design software. This instructs their software to perform computationally intensive tasks when there is more green energy in the electricity mix.
There are numerous CO2 trackers for software available online that take different approaches or specialise in various programming languages.
Carbonara estimates CO₂ emissions right from the coding stage (Estimates)
The Carbonara tool, which we have already covered in a separate article, takes an interesting approach. The CO2 tracker is designed to provide a forecast of the software’s future emissions whilst the code is still being written. According to the developers, this is intended to make optimisations more effective.
As Daniela Grau from Carbonara explained to us, code is usually only optimised for energy efficiency at the end of a project. However, as a great deal of money has already been invested by that point, such optimisations are not particularly efficient. This is precisely what the team aims to change with Carbonara, though the tool is only compatible with VSCode-based editors.
Claude Carbon demonstrates what CO2 tracking might look like for “vibe coding” (Estimates)
Using a small tool, we would like to highlight the intersection of two fields of application that are becoming increasingly popular: in so-called ‘vibe coding’, generative AI is used to generate executable code. Depending on how developers use these tools, code generators such as Claude Code can certainly produce efficient program code. However, just like ChatGPT or Google Gemini, they require a great deal of computing power to do so.
Claude Carbon is designed to visualise the energy consumption of the “Claude Code” tool. It was developed by Gaëtan Wittebolle, a French programmer who wanted to visualise his own energy consumption. Based on a study of the energy consumption of Amazon’s server infrastructure, the tool estimates the carbon footprint based on executed sessions.
Gaëtan demonstrates exactly how this works in a blog post: after four months, Claude Carbon showed 215.5 kg of CO2 equivalents. Extrapolated over a year, this would amount to 10 percent of the individual CO2 budget of an average person in France. However, by adjusting his use of Claude Code, Gaëtan was able to reduce the high energy consumption.
CO2 tracker for websites
Every website you visit on the internet is hosted on a server. As soon as you access it, the server transmits data to your computer or smartphone, which requires energy. Truly energy-efficient websites can be so small that they can be powered by a solar cell on a mini-PC, as demonstrated by LowTechMagazine, which even goes offline sometimes when there isn’t enough sunlight.
Designing websites to be truly efficient and carbon-neutral depends on many factors: what energy mix powers the server, how large the uploaded images are, and which cookies, trackers and other files are loaded per session. If you want to optimise all these aspects, the following trackers can help you:
webNRG by green-coding.io (Measurements)
The online tool webNRG offers three key benefits: it measures the energy required to render the content. It measures the traffic generated and provides a forecast of how many CO2 equivalents the transmitted data would generate in the network infrastructure over the course of a year, based on 10,000 monthly visits. And, if desired, it can send regular updates on page performance via email.
Green-Coding.io uses the Running Average Power Limit (RAPL for short), which the manufacturer Intel supports in its server architectures, for measurement. To view the website in isolation from the browser, which may further delay loading times, webNRG isolates the loaded webpage. You can read exactly how this works in the webNRG methodology.
If you do not wish to retrieve this information via an external service such as webNRG, Green-Coding.io also offers the Green Metrics Tool. This service is also free of charge and works via Docker, amongst other methods. If you use Docker containers on your server, you can install the tool with just a few clicks. Alternatively, installation guides for Linux, Windows and macOS are available at Green-Coding.io.
CO2.js as a JavaScript library (Measurements)
The Green Web Foundation offers another interesting alternative in the form of CO2.js, which is based on JavaScript. As the tool can, in principle, measure the emissions of all data transfers, you can use it not only for websites but also for software and apps.
What’s interesting about CO2.js is that the developers have linked their tool directly to the Green Web API. Through this, the Green Web Foundation provides information on the current electricity mix, based on the server’s location. Fershad Irani explained exactly how this works in the context of grid-aware websites.
This allows CO2.js to incorporate location data as a variable in its calculations. Fershad Irani explains the exact benefits of this using the example of the Italian ‘Sitigreen website carbon calculator’ in the linked video.
Ecograder (Measurements & Estimates)
Ecograder is a particularly practical and easy-to-use tool that’s available online. For every webpage scanned, Ecograder assigns a score on a scale of 0 to 100. The higher the score, the more energy-efficient your website is. For each page view, Ecograder also calculates the amount of CO2 emissions released using the Sustainable Web Design Model, which was developed by the Green Web Foundation and Wholegrain Digital, among others.
In addition to these two variables, Ecograder assigns a score for UX design, green hosting and the size of your website. For these three areas, the website then suggests ways to optimise your site, such as reducing the size of images or the number of requests to the server. For some of these suggestions, you don’t even need any technical knowledge. You can, for example, reduce the size of your website’s images using the open-source tool GIMP to improve your Ecograder score.
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.
If you’d like to explore the topic in more depth, Ecograder is also available in a Pro version. Here, you can create an account and then save multiple measurements to compare them over a longer period.
Further tracking solutions
Our list of resources and CO2 trackers is intended as an introduction to the topic. If you’d never heard of this topic ten minutes ago, you may now already know what to look for to suit your own needs. With this knowledge, we’d now like to direct you to comprehensive lists from the Green Web Foundation and the German Green Software Association:
