Unmasking AI’s Footprint - The Real Cost of Large Language Models
The Hidden Environmental Cost of AI
TL;DR ⏱️
- Mistral AI published the first full life-cycle analysis (LCA) of an LLM
- Training has huge up-front CO₂, water, and resource costs
- Smart choices (renewables, smaller models, optimisation) reduce the footprint
Background
🌍 Mistral AI just released the first full life-cycle analysis (LCA) of an LLM (Mistral Large 2).
The numbers are eye-opening:
- 20.4 kt CO₂e, 281 000 m³ water & 660 kg Sb-eq to train over 18 months
- 1.14 g CO₂e, 45 mL water & 0.16 mg Sb-eq for a single 400-token response
- Footprint scales roughly linearly with model size (10× parameters ≈ 10× impact)
This raises questions about sustainability, usage patterns, and model selection.
What have I done:
I reviewed the Mistral post and analysed where the environmental impact of LLMs comes from:
- Training once, paying for years – the up-front carbon loan dominates unless utilisation is high.
- Hardware & supply chain – producing GPUs and cooling systems can rival the energy cost.
- Inference at scale – small grams per query add up massively with millions of requests.
- Utilisation ratio – measuring "total inference ÷ total lifecycle" helps evaluate if training is “earned back.”
IMHO:
🔬 Key takeaways for sustainable AI:
- Cool-climate, renewable-powered datacentres cut both carbon and water draw.
- Always choose the smallest model that solves the task – big isn’t always better.
- Use distillation, quantisation, batching, and caching to reduce impact.
- Skip LLMs altogether if simpler methods (regex, heuristics, traditional ML) do the job.
AI has the potential to optimise logistics, drug discovery, and code generation — but its own footprint is real and must be addressed.
At Comma Soft AG, we always validate the most reasonable approach to solve a problem and optimise resource spend, especially as we host LLMs ourselves within Alan.de.
❤️ Feel free to reach out and like if you want to see more of such content.
#artificalintelligence #llm #oneearth