How Much Water Does It Take to Generate One Video?
About 2 milliliters. That is a teaspoon of water for two videos, or one 500 mL bottle for about 240 videos. Data-center cooling alone is closer to 0.2 mL. Most of the rest is water used to generate the electricity.
This is an estimate for one typical Merra AI studio video, not a meter on the servers. We mapped the real pipeline, timed the fast steps from wall clock, and converted energy with published water-use factors.
What “generate one video” actually does
Merra AI does not invent pixels the way Sora, Runway, or Luma do. A studio generate run:
- Writes one voiceover script with Gemini 2.5 Flash Lite (~3 seconds)
- Speaks it with ElevenLabs Flash (~3 seconds)
- Edits your existing 1080×1920 clips on AWS Lambda with FFmpeg (~45 seconds)
That last step is the only one that actually runs for tens of seconds. The model calls are short.
Assumptions for this estimate: one script version, 45-second output, Lambda billed for 45 seconds, script and TTS each about 3 seconds at ~100 W.
The number
| Water | |
|---|---|
| Best estimate (cooling + electricity) | ~2 mL |
| Data-center cooling only | ~0.2 mL |
| Plausible range | 1–4 mL |
A few drops more than Google’s published cooling figure for a median Gemini text prompt (0.26 mL). A toilet flush is about 2,900 of these videos.
Where the water goes
Full footprint for one 45-second video, including electricity-generation water (US average 3.1 L/kWh):
| Step | Energy | Cooling | Full water |
|---|---|---|---|
| 1× Gemini Flash Lite (~3s) | 0.10 Wh | 0.11 mL | 0.4 mL |
| ElevenLabs Flash TTS (~3s) | 0.10 Wh | 0.03 mL | 0.3 mL |
| Lambda FFmpeg (45s, 3 GB) | 0.38 Wh | 0.05 mL | 1.2 mL |
| API, S3, Mongo | 0.05 Wh | 0.01 mL | 0.2 mL |
| Total | 0.63 Wh | 0.2 mL | 2.1 mL |
Encode is more than half. Script and voiceover are the same order of magnitude as each other, and both are small next to 45 seconds of CPU video work.
Fast models are not 0.24 Wh calls
Google’s 0.24 Wh figure is a median Gemini Apps prompt across mixed models, with idle machines amortized in. Flash Lite returning in a few seconds is a small TPU slice for ~3 seconds: about 0.10 Wh.
Same story for TTS. ElevenLabs Flash returns a 45-second voiceover in about 3 seconds (~15× realtime). At ~100 W for GPU, host, and PUE, that is also ~0.10 Wh, not several GPU-seconds of work.
If a step finishes in a few seconds, it cannot dominate a job that encodes video for 45 seconds.
Duration
Script time stays around 3 seconds (the prompt is already long; the spoken output is short). TTS and encode scale with the clip.
| Output length | Cooling | Full water |
|---|---|---|
| 15s | 0.1 mL | 1.1 mL |
| 30s | 0.2 mL | 1.6 mL |
| 45s (typical) | 0.2 mL | 2.1 mL |
| 60s | 0.2 mL | 2.6 mL |
Everyday comparisons
| Activity | Water | Equivalent Merra AI videos |
|---|---|---|
| One 45s Merra AI studio video | 2.1 mL | 1 |
| Median Gemini text prompt (Google, cooling) | 0.26 mL | — |
| Average ChatGPT query (OpenAI, cooling) | 0.32 mL | — |
| Teaspoon | 5 mL | ~2 |
| 500 mL bottle | 500 mL | ~240 |
| Toilet flush | 6 L | ~2,900 |
| One-minute shower | ~8 L | ~3,800 |
| Open text-to-video clip (WAN 2.1 14B, ~415 Wh) | ~1.4 L | ~670 |
Scale
- 8 videos / month: ~17 mL
- 3 videos / day: ~6 mL/day, ~2 L/year
- 1,000 videos: ~2.1 L, about a third of a toilet flush
This is not generative AI video
Pixel-generating models (Runway, Luma, Sora) synthesize frames on GPUs. A short diffusion clip can be 50–400 Wh — 0.2–1.4 L of water including electricity, 100–670× a Merra AI montage.
Studio generate writes a script, speaks it, and edits cuts you already filmed. That is why the water number stays in milliliters.
How we built the numbers
Energy. Script and TTS: ~3 s wall clock × ~100 W (accelerator + host + PUE) → ~0.10 Wh each. Lambda: 3008 MB ≈ 1.7 vCPU, x86 FFmpeg for 45 s at ~25 W package share plus PUE 1.15 → ~0.38 Wh. S3/API is a small remainder (~0.05 Wh).
Water. Cooling: Google implied ~1.08 L/kWh from 0.26 mL / 0.24 Wh; AWS global WUE 0.12 L/kWh (2025); ElevenLabs unknown, Microsoft-like 0.27–0.30 L/kWh. Electricity: US average consumption 3.1 L/kWh (Li, Yang & Ren, Making AI Less Thirsty). That electricity term is most of the 2 mL. On-site cooling is ~0.2 mL.
Not included. First-time clip ingest and auto-cut (a separate Lambda job, often heavier than generate). Extra script variants, voice clone, translation, retries. Model training amortized across all customers. The viewer’s phone. Regenerating script, voice, or video multiplies the figure.
This is a transparent estimate from the pipeline we run, not a certified lifecycle assessment. If a step is fast, we counted it as fast.
Sources: Google, Measuring the environmental impact of delivering AI at Google Scale (2025); OpenAI / Altman query figures (2025); AWS WUE 0.12 L/kWh (2025).
Go to Merra AI · Enable JavaScript for the full interactive Merra AI app, or continue reading this page content above.