The GreenPrompt Whitepaper

A primer on the energy economics of large language models.

1. Why prompt efficiency matters

Training a single large language model can emit hundreds of tonnes of CO₂ — but inference, repeated billions of times daily, now dominates lifetime emissions. Every redundant token in a prompt directly translates to GPU cycles, energy, and carbon.

2. Energy coefficients

GreenPrompt uses these research-aligned coefficients (Wh per 1k tokens):

• GPT-4o: 0.005 Wh
• GPT-4o-mini: 0.0004 Wh
• Gemini 1.5 Pro: 0.004 Wh
• Gemini 1.5 Flash: 0.0003 Wh

3. Carbon intensity

Global average grid intensity: 0.475 g CO₂e per Wh (IEA, 2023). Real values vary by region and time of day.

4. Real-world equivalents

• Smartphone charge ≈ 12 Wh
• LED bulb hour ≈ 9 Wh
• Laptop minute ≈ 0.83 Wh (50W avg)
• Google search ≈ 0.3 Wh
• EV km ≈ 180 Wh (Tesla Model 3)
• Tree absorption ≈ 21 kg CO₂/year

5. References

1. Patterson, D. et al. (2021). Carbon Emissions and Large Neural Network Training. arXiv:2104.10350.
2. Luccioni, A. S., Viguier, S., Ligozat, A. L. (2023). Estimating the Carbon Footprint of BLOOM. JMLR.
3. Strubell, E., Ganesh, A., McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP. ACL.
4. IEA (2023). Electricity Grid Emissions Factors.
5. Google (2024). Environmental Report.

6. Limitations

Coefficients are public-research estimates — true per-token energy depends on hardware, batching, KV-cache reuse, and datacenter PUE. Treat numbers as comparative, not absolute.