Greener by Design
How to Measure the Carbon Footprint of AI Projects
The Hidden Emissions of Intelligence
Artificial intelligence may be digital, but its footprint is very real.
Every data centre, GPU, and training cycle consumes energy, and the global expansion of AI infrastructure is now a measurable contributor to carbon emissions. For small and medium-sized enterprises (SMEs) and cultural organisations, this raises an important question: how do you know if your AI system is sustainable?
The good news is that measuring AI’s environmental impact is becoming more practical, even for smaller teams. By understanding where emissions originate and how to quantify them, organisations can make informed design choices that balance performance, cost, and responsibility.
Where AI’s Carbon Comes From
AI’s carbon footprint arises from three main sources:
Training: The initial creation of models, especially large-scale ones, is highly energy intensive. Training an LLM typically consumes more energy than a Jumbo Jet circumnavigating the earth, twice.
Inference: Running predictions or generating outputs over time also consumes significant power, particularly at scale.
Data Lifecycle: Storing, transferring, and cleaning datasets all contribute to energy use.
While many discussions focus on training emissions, inference (the day-to-day running of models) often represents a larger cumulative cost for SMEs. A modest image-classification model, for example, may run thousands of times a day, generating a steady stream of emissions that are rarely accounted for.
How to Start Measuring
You don’t need enterprise-level tools to begin tracking impact. A few practical methods can help:
Energy-based estimation: Monitor electricity use (in kWh) during training or operation, then convert to CO₂ using local energy mix data. Tools like CodeCarbon and Green Algorithms provide straightforward calculators.
Compute-based estimation: Track GPU/CPU usage time and memory requirements. Comparing model configurations can reveal efficiency gains before deployment.
Cloud provider metrics: Major platforms (AWS, Azure, Google Cloud) now include sustainability dashboards showing emissions by service.
Lifecycle accounting: Consider emissions across data capture, preprocessing, and model updates, not just compute time.
Even approximate tracking helps identify “hot spots” where optimisation will yield the greatest environmental return.
Reducing Impact Without Reducing Performance
Once you’ve measured, the next step is mitigation. Several proven strategies can cut your AI’s carbon footprint without undermining its effectiveness:
Choose smaller or fine-tuned models instead of training from scratch.
Use hybrid or Physics-Informed Neural Networks (PINNs) to reduce unnecessary computation.
Optimise batch sizes and learning rates to shorten training time.
Run workloads during off-peak energy hours or in regions with cleaner grids.
Cache and reuse intermediate results where possible, especially in creative workflows like rendering or reconstruction.
Each of these actions contributes to a broader goal: making efficiency a core part of design, not an afterthought.
Reporting and Responsibility
In the heritage and creative sectors, demonstrating environmental responsibility is increasingly a funding requirement. Innovate UK, Horizon Europe, and other frameworks now request sustainability statements for R&D projects.
By integrating carbon measurement into AI workflows, SMEs can strengthen bids, improve transparency, and align with institutional values.
Simple reporting templates can help:
Scope 1 & 2 emissions for direct and purchased energy.
Scope 3 equivalents for outsourced compute or data storage.
Per-model carbon intensity metrics (e.g., kg CO₂ per training cycle).
These data points not only support sustainability reporting, they also tell a story about innovation with integrity.
Final Thought
AI’s success shouldn’t come at the planet’s expense. Measuring and reducing emissions isn’t about perfection, it’s about awareness, accountability, and incremental progress.
For SMEs and cultural organisations, sustainable AI design offers more than environmental benefits: it builds trust, reduces cost, and demonstrates leadership in responsible innovation.
Green AI isn’t a niche movement anymore; it’s becoming a defining mark of quality.
