In the field of Artificial Intelligence, one question remains unanswered: What are the environmental costs of training and inference of AI models? To address this question, IDED (formerly known as SDIA), in collaboration with the University of Stuttgart, developed a Kubernetes plugin that enables AI models to determine their own environmental impact. This project is supported as an AI flagship initiative by the Federal Ministry for the Environment and Consumer Protection (BMUV).

What is NADIKI?

NADIKI develops the capability for AI models to determine their own environmental costs within a Kubernetes cluster. This is particularly relevant during the inference phase, when the model is deployed in production environments and queried in real time—often across many data centers, in different regions, and on various hardware platforms.

NADIKI employs a three-tier approach: 1) mathematical estimation based on established formulas, 2) measurement via hardware and software-based APIs, and 3) direct linkage to the data center to query external power measurements, hardware information, and energy data.

The goal is a holistic view of the environmental impact—from used and reclaimed electricity, to server and network hardware employed, to water usage, cooling systems, and the data center building itself.

Transparency as a Basis for New Standards

Without data from an application's infrastructure, it is impossible for developers and operators to understand the environmental impact of their AI models. With NADIKI, it becomes possible—and straightforward—to measure, monitor, and report environmental costs: from energy to resource consumption to emissions.

Based on the metrics provided by NADIKI, AI applications can be optimized:

• Maximize server resources: Fully utilize existing infrastructure to avoid the construction of unnecessary new data centers or IT equipment.

• Understand actual resource consumption: Use signals from physical infrastructure to optimize or halt training when resource targets are not met—for example, only train when renewable energy is available.

• Make informed decisions about location and timing: Relocate training to another region, adjust speed, or operate AI applications at times of free capacity or favorable cooling conditions.

What is New About NADIKI?

For the first time, NADIKI provides environmental metrics for AI models and applications running on Kubernetes. In collaboration with the University of Stuttgart, NADIKI is developing data collection software for the data center that provides a unified API for Kubernetes: What is the current power mix? How much renewable energy is available? Are the diesel generators running? How much water is currently needed for cooling?

Within Kubernetes, NADIKI aggregates information from the data center, server infrastructure, and orchestration platform, enriches it, and delivers all relevant KPIs to the application and monitoring systems.

NADIKI uses a life cycle assessment (LCA) to measure actual resource consumption—including regional power generation, use of recycled IT hardware, and reuse of waste heat. It supports both virtual and physical IT infrastructure.

All software—for both Kubernetes and digital infrastructure—will be released as open source and can be utilized within existing IT infrastructure.

AI Flagship: Support by BMUV

The NADIKI funding was presented to the project team by Christian Kühn, Parliamentary State Secretary of the BMUV, and Corinna Enders, Managing Director of the Future-Environment-Society (ZUG), at an event for AI flagship projects from the Baden-Württemberg region.

The other funded AI flagship projects in the region: DESIRE4ELECTRONICS (AI for optimizing the recycling of small electronic devices), KiKKa (AI for climate-neutral sewage treatment plants), and RecycleBot (AI to increase the plastic recycling rate).