Table of Contents
AI is changing the world. It is also consuming a significant and growing portion of it. The energy demands of training large AI models, running inference at scale, and powering the data centres behind it all represent a material and increasingly scrutinised environmental cost.
For technology leaders, sustainability is no longer just a values conversation. It is a procurement question, a regulatory consideration, an ESG reporting obligation, and increasingly, a competitive differentiator. Companies that ignore the environmental footprint of their AI stack are building a liability. Companies that address it seriously are building an advantage.
| 🌱 By the numbers: Training a single large language model can emit as much CO₂ as five cars over their entire lifetime. Running that model at scale over the years multiplies the impact many times over. |
Where the Environmental Cost Actually Lives
Model training
Training a large foundational model from scratch is among the most computationally intensive tasks in modern computing. The energy required is enormous, and unless the training runs are powered by renewable energy, the carbon impact is significant. Most organisations will never train a foundational model, but they may fine-tune one, which carries its own footprint.
Inference at scale
Every query you send to an AI model, every search, every recommendation, every document summarised, requires computation. At the scale of millions of queries per day, inference becomes a major ongoing energy cost. Unlike training, which happens once, inference never stops.
Data storage and management
The datasets that power AI are often vast. Storing, replicating, and retrieving them continuously requires energy. Cold storage helps, but data that is actively used in training pipelines or feature engineering creates a persistent load.
Data centre infrastructure
Cooling is the hidden cost of computing. Data centres use significant water and energy to keep hardware at operating temperatures. The location of your cloud infrastructure matters: a data centre powered primarily by coal has a very different footprint than one powered by hydroelectric or wind energy.
Practical Steps to Reduce Your AI Environmental Footprint
1. Choose the right model size for the task
This is the single most impactful decision you can make. A 7-billion-parameter model may perform nearly as well as a 70-billion-parameter model for your specific use case, at one-tenth the compute cost. Model selection should be driven by fit-for-purpose reasoning, not prestige.
| 💡 Tip: Benchmark smaller, open-source models against your specific tasks before defaulting to the largest commercial option available. |
2. Prefer fine-tuning and retrieval-augmented generation over training from scratch
Fine-tuning an existing foundation model on your domain-specific data is dramatically more efficient than training from scratch. Retrieval-augmented generation (RAG), which allows a model to pull from a curated knowledge base at inference time, can reduce the need for large, generalised models while improving accuracy and reducing hallucination.
3. Optimise inference
Techniques like model quantisation (reducing precision from 32-bit to 8-bit or 4-bit), pruning, and distillation can reduce inference compute requirements by 50–80% with minimal accuracy loss for most applications. These are not experimental techniques; they are production-ready and should be part of every AI deployment review.
4. Select cloud regions powered by renewable energy
Major cloud providers, AWS, GCP, and Azure, all publish carbon intensity data by region. Routing workloads to regions with high renewable energy penetration can significantly reduce scope 2 emissions. This is a configuration decision that requires almost no additional cost.
5. Implement compute scheduling
Batch processing and scheduled inference runs can shift compute demand to off-peak hours, when grid energy is often greener and cheaper. Not every AI workload needs to run in real time. Identifying which ones do and which ones do not is a useful exercise with both cost and sustainability benefits.
6. Measure and report
You cannot manage what you do not measure. Tools like CodeCarbon (open source), cloud provider sustainability dashboards, and emissions calculators from providers like Hugging Face enable you to track the actual carbon footprint of your AI workloads. Building this into your ESG reporting signals maturity to investors, clients, and regulators.
What to Ask Your Technology Partner
- Do you measure and report the carbon footprint of your development and deployment infrastructure?
- What model size recommendations do you make, and are sustainability considerations part of that analysis?
- Do you deploy into renewable-energy-powered cloud regions by default?
- Can you provide guidance on model efficiency techniques such as quantisation and distillation?
Conclusion
Green AI is not a niche concern. As ESG disclosure requirements tighten globally, as energy costs rise, and as enterprise clients increasingly require sustainability commitments from vendors, the environmental footprint of your technology stack will matter more, not less.
Sustainability is a consideration in every AI architecture decision we make. We help clients select the most efficient model architectures, deploy into greener infrastructure, and measure the footprint of what they build. If you want to talk about how to make your AI programme both powerful and responsible, we would be glad to hear from you.