← AI Terminology
Energy Efficiency (Green AI)
Green AI is the movement to measure, minimise, and report the energy consumption and carbon footprint of AI training and inference — recognising that large-scale AI carries a significant and growing environmental cost.
Training a single large LLM can emit as much CO₂ as hundreds of transatlantic flights.
Training a single large LLM can emit as much CO₂ as hundreds of transatlantic flights.
Why It Matters in AI
AI compute demand is growing at ~4× per year, driven by larger models and more inference calls. Data centres for AI now consume a material fraction of national electricity grids. For companies, energy cost is a significant component of inference margins. For investors, energy efficiency is a competitive moat — models that deliver the same quality at lower compute cost win economically. For society, AI's carbon footprint is a policy and reputational risk.
Key Points
| Aspect | Description |
|---|---|
| Reporting | MLCommons MLPerf Energy track; model cards increasingly include energy use disclosure |
| Green hardware | NVIDIA H100 NVL, Google TPU v5e, AWS Trainium2 — designed for energy-efficient AI |
| Training carbon | GPT-3: ~552 tonnes CO₂eq; estimated GPT-4: >10,000 tonnes CO₂eq (based on reported compute) |
| Algorithmic gains | Quantization, distillation, MoE, speculative decoding — reduce FLOPS per inference |
| Efficiency metrics | FLOPS per token, tokens per watt, PUE (Power Usage Effectiveness) of data centre |
| Inference at scale | Serving millions of requests/day often exceeds training carbon over a model's lifetime |
Simple Analogy
Training a large AI model is like flying a jet across the Atlantic — efficient if amortised across millions of inference queries, but a significant one-time cost. Inference is the ongoing fuel burn: optimising model efficiency is like improving fuel economy rather than flying less.
Common Usage Examples
codecarbonlibrary:tracker = EmissionsTracker(); tracker.start(); ...; tracker.stop()— measures CO₂ during training- NVIDIA H100 NVL: 3.9 TFLOPS/W — twice the energy efficiency of previous generation
- Speculative decoding: 2–3× tokens/Joule improvement on LLM inference
- Google reports annual AI energy use in its sustainability reports; Microsoft's AI footprint is a growing line item
- Mixture of Experts: GPT-4 style MoE achieves dense-model quality while activating only a fraction of parameters per token
Summary
In short: Green AI is the recognition that intelligence has an energy cost — and that making AI more efficient is simultaneously a competitive, economic, and environmental imperative.