The AI industry measures everything — except its own environmental impact. This presentation from EnviroInfo 2025 demonstrates a working end-to-end measurement system built through the NADIKI project: a proof of concept that traces every AI interaction back to real data center infrastructure, replacing industry estimates with actual measurements of energy use, water consumption, resource depletion, and CO₂ emissions per prompt.

The AI Measurement Paradox

The companies driving the AI revolution — OpenAI, Google, Anthropic, Microsoft — can optimize advertising to the microsecond and monitor server performance across global infrastructure. Yet when it comes to the environmental cost of a single AI prompt, they report estimates, model assumptions, and lab results from controlled conditions.

Existing tools reveal the depth of the problem:

• AI Energy Score — still guessing based on model parameters

• ML CO2 Impact Calculator — built on assumptions about hardware efficiency

• CodeCarbon — tracks the local machine, blind to the infrastructure behind the API

• Google's Gemini figure (0.24 Wh per prompt) — from controlled lab conditions, not billions of real production queries

• The MIT debate — whether ChatGPT uses 0.3 Wh or 3 Wh remains unresolved

Without real numbers, the rebound effect cannot be prevented: more efficient models simply lead to more consumption. The lack of environmental feedback removes user awareness and drives usage higher.

The NADIKI Solution: From Estimation to Measurement

The NADIKI project built a five-step architecture that connects each AI interaction to real infrastructure data — not model assumptions:

Step 1 — The Foundation: Direct connection to existing monitoring systems (Prometheus, Telegraf). No proprietary hardware required. Where data gaps exist, peer-reviewed assumptions from Green Coding Berlin and Boavizta fill in. The framework prioritizes real measurement wherever available.

Step 2 — Infrastructure Data Collection: Data centers report real-time energy consumption, water usage, and renewable energy share. Servers report resource utilization, power draw, and temperature. Kubernetes reports workload allocation. All stored in InfluxDB for precise temporal correlation.

Step 3 — The Registrar System: Data centers register with static information (location, energy mix, PUE, cooling systems). Servers register with hardware specifications (RAM, CPUs, GPUs, power consumption, efficiency). Both continuously send real-time performance data to the time-series database.

Step 4 — Real-time Impact Calculation: An API receives queries with start/end timestamps, identifies which servers ran the workload, and calculates the environmental impact for that specific timeframe. Transforming operational data into impact metrics.

Step 5 — The User Experience: An Open Web UI integration captures the timestamp when a message is sent and when the response is complete, then queries the system: "What was the environmental impact for this timeframe?" Every message shows real environmental impact — energy use, water consumption, resource depletion, operational and embedded CO₂ — at three levels: per message, per conversation, and cumulative for all activity.

The Breakthrough: End-to-End Accountability

For the first time, a specific AI interaction can be connected to its real infrastructure impact. Full traceability from query to data center. No more black boxes. The architecture uses standard, off-the-shelf components — proof that this level of measurement is technically feasible.

The system is a research prototype. The goal: move from prototype to adoption-ready system by end of 2025.

The Path Forward

The technology exists. The market lacks incentive — without regulation or economic pressure, AI companies have no reason to measure what they don't want to report. Regulation can create that incentive, just as fuel efficiency standards transformed the automotive industry.

The question is not whether AI's environmental impact can be measured in real time. We just showed it can. The question is whether we will.