🌏 PLANET GenAI Media Footprint Estimator

Methodology & assumptions

1. System boundary

• Scope: This tool estimates energy, carbon and water for the inference phase of GenAI models – the moment you generate text, images or video. It does not include model training, embodied hardware impacts, or user devices.
• Equations:
  IT energy (kWh) = (per-item Wh × count) / 1000
  Facility energy (kWh) = IT energy × PUE (1.3)
  CO₂e (kg) = Facility energy × grid intensity (0.4 kg CO₂e/kWh)
  Water (L) = IT energy × WUE (1.0 L/kWh) + Facility energy × grid water factor (2.5 L/kWh)

2. Per-asset energy assumptions (IT energy)

• Images (~2.9 Wh per image): based on measurements for Stable-Diffusion-class models (≈2.9 kWh for 1,000 images) on modern GPUs. Applied to NanoBanana / Imagen 3, MidJourney, DALL·E 3 (ChatGPT) and Runway image generation.
• Video (clip-level, not frame-by-frame): we assume that text-to-video models reuse computation across frames (temporal coherence, shared latent representations) and do not re-run a full image model for every frame. Using public benchmarks on H100-class hardware as a guide, we treat a short 720p clip as a multiple of a still image, not frames × image cost:
  • Runway / MidJourney video: ≈22 Wh per 5–6 s, 24 fps, 720p clip.
  • Google Veo: ≈56 Wh per 8 s clip (larger, higher-quality model).
  • Sora-style / very large models: ≈100 Wh per short clip (upper end of “central” range).
  Earlier versions of this estimator used a simple “frames × still-image” model, which produced much higher values; this clip-level approach is more realistic for modern video diffusion models.
• Text: calibrated to public estimates of around 4.3 g CO₂e per GPT-4-class request on a 0.4 kg CO₂e/kWh grid with PUE 1.3. That corresponds to:
  • Facility energy ≈ 0.0108 kWh per query.
  • IT energy ≈ 0.0083 kWh per query = 8.3 Wh.
  In the tool we use:
  • ChatGPT / GPT-4, Gemini, Claude: 8.3 Wh per query.
  • Llama / open models: 6.0 Wh per query (more efficient / smaller models).
  So 100 GPT-4-class queries give ≈0.43 kg CO₂e and ≈3.5 L water.

3. Why training is not included

• Training a frontier model can require tens of MWh of electricity, but that cost is amortised over billions of downstream inferences. The per-asset share is uncertain and speculative.
• To keep this tool usable at project level, we focus on the marginal footprint of generating each asset, which teams can actually influence through volume, media mix and asset reuse.
• Training impacts can be discussed separately (e.g. “this model required ≈X MWh to train”), but they are not allocated per asset in this estimator.

4. Impact of building this calculator

Developing and iterating this estimator involved a few hundred ChatGPT-style requests. Using the same text assumptions (~8.3 Wh IT, ~0.0108 kWh facility per query), a design process with 300 prompts would use roughly:

• Facility energy: ≈3.2 kWh
• CO₂e: ≈1.3 kg CO₂e
• Water: ≈10–12 L (combining data-centre cooling and grid water)

That is equivalent to only a few tens of kilometres of driving in a petrol car and a fraction of a standard shower. It is small compared to a GenAI-heavy campaign, but we note it here for transparency. This “tool-building” footprint is not added to project calculations.

5. Real-world analogies

• Driving: ≈0.192 kg CO₂e per km for an average petrol car.
• Showers: ≈80 L of water per standard 8-minute shower.
• Trees: ≈25 kg CO₂e sequestered per mature tree per year.
• Glasses of water: 250 ml per glass.

All numbers are indicative. The tool is designed for comparison, estimation and storytelling – not formal carbon accounting or supplier-specific reporting.