Data centres consume 460-490 TWh in 2025, as much as France. The IEA projects 945 TWh by 2030. The 5 hyperscalers invest $660-690B capex in 2026. How to reconcile AI and low-carbon strategy?
Training a model like GPT-4 consumes ~2-15 GWh. But inference accounts for 80-90% of total global consumption, every prompt has a real energy cost.
Google search: ~0.3 Wh. ChatGPT request: ~10 Wh (33x more). DALL-E image: ~3 Wh. Edge AI (local NPU): ~0.001 Wh. Shifting to edge reduces impact by 10,000x.
Offload simple inferences to local NPU (smartphone, Copilot+ PC, IoT sensor). Consumption reduction: 100x to 10,000x per request vs cloud.
-99% per inferenceSLMs (1-8B) vs LLMs (70B+): 95% lower consumption for 80-90% of performance on targeted tasks. Phi-4 Mini, Llama 3.2 3B, Gemma 2 2B.
-95% energyAzure (100% renewable since 2025), Google Cloud (64% carbon-free energy), AWS (100% renewable 2025). Choose the greenest region for intensive workloads.
CO2 -80-100%Schedule intensive training and batches during off-peak hours (night, weekend) when energy is greener and cheaper. Savings: 20-40% on energy bill.
-20-40% billMolderez Consult SRL co-builds your 18-month AI roadmap.
AI strategy workshop