PLANET GenAI Media Footprint Estimator

Methodology & assumptions

This estimator is designed for quick, order-of-magnitude comparisons across media types and engines. It uses a combination of published measurements and simple scaling rules.

System boundary

• IT energy per item (Wh): GPU/CPU/RAM energy for one image, video, or text response.
• Facility energy (kWh): IT energy × PUE.
• CO₂e (kg): Facility energy × grid emission factor (kg CO₂e/kWh).
• Water (L): on-site water = IT energy × WUE (L/kWh IT), upstream water = facility energy × grid water intensity.

Per-item energy baselines

• Images: ~2.9 Wh per image (IT only) based on measurements of Stable Diffusion-style models (≈2.9 kWh per 1,000 images).
• Videos: open text-to-video models ≈90 Wh for a 5.4 s, 720p, 15 FPS clip on an H100 GPU; Veo/Runway/MidJourney video profiles scale from this by duration and FPS.
• Text requests: LLM prompts roughly in the 0.03–0.3 Wh range per query; this tool uses 0.05–0.07 Wh as a conservative planning assumption depending on engine.

Shared factors

• PUE = 1.3 (midpoint between hyperscale averages ≈1.1–1.2 and global data-centre averages ≈1.47).
• Grid emission factor = 0.40 kg CO₂e/kWh (global-average style placeholder for electricity).
• WUE = 1.0 L/kWh IT (within reported ranges 0.36–1.8 L/kWh; reflects moderately water-cooled, AI-heavy workloads).
• Grid water = 2.5 L/kWh facility as a generic estimate for upstream electricity water consumption.

Analogies

• Driving: ≈0.2 kg CO₂e per km for an average petrol car.
• Showers: ≈80 L per standard shower (≈8 minutes).
• Trees: ≈25 kg CO₂e/year sequestered by a single mature tree.