Much has been written about the environmental impact of artificial intelligence (AI) models. A standardized assessment model and mandatory reporting, however, are still missing. As part of our NADIKI research project, we aim to establish a standardized measurement infrastructure built on Kubernetes. And with the following overview, we want to give a conceptual idea of how and where we see the added value of measurement across the lifecycle of an AI model.

As with many software systems, this can be seen as a manufacturing process for a machine part that is then integrated into a final product. It begins with creating a design for the machine part, considering the materials to be used and the manufacturing process (e.g., dataset, approach to training). Then the machine part is produced using machinery (training infrastructure) and is then subject to quality control (testing). This creates the machine part (the model), which can then be implemented into our own or third-party products (model API).

Let's look at the different phases, their environmental impact (there is a definition at the end of the document), and output.

Lifecycle of an Artificial Intelligence Model

Design

During the design phase of a new model, developers make two key decisions that can result in more or less environmental impact during training and use.

1. The dataset that is going to be used for training the model. If it is synthetic, how much environmental impact was created by the computation necessary for generating it? If it is organic or scraped content, the same question applies. Collecting data uses compute and storage, which in turn creates environmental impact. Considering a dataset that has a limited footprint can reduce the impact of the resulting model.

2. The approach to training and algorithms is logically the largest multiplier in the design phase on both the computational demand of the training itself and the environmental impact created during training. Putting efficiency at the center of this decision (e.g., how we can train this model using the least amount of computing) will have a positive effect on the total footprint of the model.

Development

The impact in this phase is straightforward – the majority of it stems from the training infrastructure that is being used to build the model. Here the amount of infrastructure matters but also the type of infrastructure, for example:

• Is the infrastructure shared or exclusive? Meaning, can it be used by others (and thus increase the value of the materials in use) or is it tied to a single provider?

• Is the infrastructure running in a region with a high penetration of physical renewable energy? For example, is there hydro or wind power physically close to the data center site used for the training?

• Does the infrastructure utilize refurbished or existing equipment or is it all newly built?

• Does the infrastructure recover the energy used for training (e.g., by ejecting it as heat to a nearby city or other industry)?

Another aspect, which was recently demonstrated by DeepSeek, is how well the infrastructure is orchestrated to maximize the computation that can be harvested for training (i.e., how well does the training utilize the resources that are present)?

Lastly, model development also involves a fair amount of testing and validation. These activities also require infrastructure which should be taken into consideration when calculating the total impact of the model.

The output of this phase is the model itself. Up to this point, the model has generated environmental impact through design choices (dataset, algorithms), infrastructure choices, and quality of orchestration, which can be calculated into a footprint for the model and published alongside it.

Use

Now for the use phase, the model is being deployed again to infrastructure (for inference) so that it can answer questions (queries), usually via an API (technical interface to pose questions to the models and process the results). For the infrastructure part, the same questions apply as for the development phase.

Now, aside from the footprint the model has from its production, it creates environmental impact for each query it answers, depending on the number of output tokens it generates (akin to how long the answer to the question is). For the user of the model, understanding the environmental impact per token is relevant to determine the total footprint of a query/response.

In the use phase, there is an additional environmental impact occurring – which stems from the system running the model waiting for a query. During that time, the model is loaded into memory, blocking important system resources. This can be improved through shared infrastructure orchestration. However, if the model is idling but still reserving significant resources, this also needs to be reported as non-usage based environmental impact from model operations.

Integration

In the integration phase, the model's API is used in a digital product or service to provide some kind of value-added functionality (e.g., to answer questions by the user in a chat-like application). For this integration, the model API needs to deliver the impact per token together with each reply it sends, so that it can be displayed to the user. This is important as it helps the user gauge the impact of its action and enables choice (i.e., is this worth it?).

An example from Notion, a popular knowledge management platform, shows a chat-like command box for issuing AI queries. In the response, it neither shows how many tokens have been used nor how much environmental impact has been caused.

Summary of Environmental Impact Measurement Points

1. Design Phase

2. Development Phase

3. Use Phase

4. Integration Phase

Environmental Impact: What Do We Want to Measure?

When we use the term environmental impact, we refer to a set of environmental and resource costs that stem from the infrastructure, its inputs, the equipment used for training and inference, and the manufacturing of the equipment.

It is important to capture both the static environmental costs – e.g., from the manufacturing of equipment or construction of the data center – and the dynamic costs, e.g., from energy use and associated emissions.

• Energy Use (kWh)

• Materials Use (ADP, Abiabatic Depletion Potential)

• Water Use (m³)

• GHG Emissions (CO₂-eq)

• Pollution (NOₓ and others)

What’s next?

In the next article we are looking into how to set up an development process for an AI model training how explore how the environmental impact can be measured.