Environmental Accounting
This page estimates the energy, water, and carbon footprint associated with producing the poster, writing the site, and conducting the dialogue that produced them. These values are estimates, not direct measurements from a data center. They are presented to make infrastructure visible, not to manufacture certainty.
Transparency note: Actual impacts depend on hardware, batching, model configuration, data center cooling design, local weather, and electricity mix. This page reports ranges and clearly states assumptions.
Inputs used for this estimate
- Text prompts: 86 (human messages)
- Generated images: ~14 (poster iterations + 1 QR code)
- Poster resolution: 1024×1536 pixels
Summary
Energy
Text + image generation (range-based).
Water
Cooling sensitivity using 1–9 L/kWh (see method).
CO₂ (sensitivity)
Using U.S. grid-average emissions factor.
| Quantity | Estimated range | Notes |
|---|---|---|
| Energy (total) | 0.021–0.087 kWh | Text prompts + image generation. Image generation dominates the uncertainty. |
| Water (cooling sensitivity) | 0.02–0.78 L | Computed by multiplying total energy by a 1–9 L/kWh cooling-intensity band. |
| Carbon (grid-average sensitivity) | 8–32 g CO₂ | Computed using U.S. grid-average emissions (0.81 lb CO₂/kWh). |
| Text prompts | 86 | Human messages in this project. |
| Images generated | ~14 | Poster iterations + QR code (counted from project artifacts available here). |
Method
Text prompts
For text, we use published production estimates of energy, carbon, and water per prompt. Google reports a median Apps prompt at roughly 0.24 Wh with 0.03 gCO₂e and 0.26 mL water per prompt in their measurement framework: Measuring the environmental impact of AI inference .
We also present a wider energy band using 0.43 Wh per short query from a benchmarking study (for sensitivity): arXiv:2505.09598.
Images
For image generation, we use a published diffusion-model energy range of approximately 0.051–3.58 Wh per image across model/configuration choices: arXiv:2511.17031. Because the exact generator configuration used here is not directly measurable from this page, image energy is treated as the dominant uncertainty band.
Water and carbon conversion
Water is reported as a cooling sensitivity using a band of 1–9 liters per kWh described for data center evaporative cooling under varying conditions: University of Illinois: AI’s Challenging Waters.
Carbon is reported as a grid-average sensitivity using the U.S. electricity emissions factor: 0.81 lb CO₂ per kWh (2023): U.S. EIA FAQ.
Interpretation
- The total energy estimate (0.021–0.087 kWh) is roughly comparable to running a 100W bulb for about 13–52 minutes.
- The water range (0.02–0.78 L) is less than a typical 1-liter bottle, but it can vary widely by cooling design and climate.
- The carbon range (8–32 g CO₂) is small in absolute terms, but the point of this page is visibility and traceability, not magnitude minimization.
What this does not include
- Visitor traffic: page views and downloads after publication (these can be estimated separately if you track analytics or server logs).
- Device energy: the electricity used by readers’ computers/phones.
- Embodied hardware impact: manufacturing impacts of servers/GPUs and end-user devices (important, but not attributable to a single short project run).