Because water is the hidden cost of intelligence.
Live app: lesterhau.github.io/tokenwater
No download. No account. Open in any browser and you're done.
You use AI every day. You have no idea how much water that costs.
Microsoft used 6.4 billion liters in 2022 — up 34% year-over-year. Google: +20% the same period. The servers running Claude, ChatGPT, Gemini, and Grok consume enormous amounts of water for cooling and electricity generation. That number grows every time you send a message. No one tells you. There's no meter running. There's no bill.
TokenWater is an estimator, not a meter. It makes an invisible infrastructure cost visible while keeping the uncertainty visible too.
TokenWater is a setup tool, not a dashboard. You open it once, set your weekly water quota, pick your AI platforms, and copy a memory snippet for each one. Paste the snippet into your AI's memory settings. Close the tab. Done.
From that point on, your AI can maintain a directional usage estimate from conversation context. Exact token counts are only available when the platform exposes usage telemetry; otherwise the estimate must be labeled as approximate.
You do not need to keep this page open after setup.
Every morning, your first conversation shows you:
- Yesterday's water usage (in liters and a physical comparison — "about one toilet flush")
- Your weekly pace and projection
- How you compare with the personal budget you chose
- Alerts at 50%, 75%, 90%, 95%, and 100% of your weekly quota
- 13 platforms supported across persistent memory and session-only tiers
- Water comparisons in plain English (water bottles, showers, toilet flushes)
- Anonymous community impact counter — see how many people are tracking
- Monthly refresh reminder baked into every snippet
- Discourages unnecessary AI follow-up questions to reduce token waste
- Praise system for good usage — irreverent, never sycophantic
- Shareable weekly stat with #TokenWater hashtag auto-generated
- Over-quota alerts at 100% and 110%
| Platform | Memory | Notes |
|---|---|---|
| Claude | ✅ Persistent | Settings → Capabilities → Memory → Manage edits |
| ChatGPT | ✅ Persistent | Settings → Personalization → Memory → Add memory (app or browser) |
| Gemini | ✅ Persistent | Settings → Extensions & Personalization → About you |
| Microsoft Copilot | ✅ Persistent | Settings → Account → Privacy → Personalization and memory |
| Claude Code | ✅ Persistent | Paste into CLAUDE.md in your project root |
| OpenClaw | ✅ Persistent | Add to persona/memory context during onboarding |
| Grok | ⚡ Session only | Paste at start of each conversation |
| Perplexity | ⚡ Session only | Paste at start of each conversation |
| Perplexity Computer | ⚡ Session only | High token burn — consider higher quota |
| DeepSeek | ⚡ Session only | Paste at start of each conversation |
| Claude Cowork | ⚡ Session only | Paste at start of each session |
| Ernie Bot (Baidu) | ⚡ Session only | Paste at start of each conversation |
| HyperCLOVA X (Naver) | ⚡ Session only | Paste at start of each conversation |
There is no universal water-per-token constant.
Modern measurements show large variation by model, hardware, serving configuration, query type, cooling system, climate, and electricity source. TokenWater is moving to a versioned low/mid/high scenario methodology rather than presenting one fixed coefficient as measured reality. See METHODOLOGY.md.
Current sources:
- Chung, J.-W. et al. (2025). "The ML.ENERGY Benchmark." NeurIPS Datasets & Benchmarks. https://ml.energy/leaderboard/
- Li, P. et al. (2025). "Making AI Less 'Thirsty'." Communications of the ACM. https://doi.org/10.1145/3724499
- Google (2025). "Measuring the environmental impact of AI inference." https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
- Oviedo, F. et al. (2026). "Energy use of AI inference, efficiency pathways, and test-time scaling." Joule. https://doi.org/10.1016/j.joule.2026.102430
- Data-center workload water-use review (2025). Resources, Conservation & Recycling. https://doi.org/10.1016/j.resconrec.2025.108310
If the methodology is wrong, open an issue and cite your source.
This is not an app about guilt. Guilt doesn't change behavior.
The Opower experiment sent households a simple letter: here's your energy use, here's your neighbor's. No moral lecture. Just a comparison. It drove a 2% national reduction in household energy consumption.
TokenWater runs the same mechanic. "You used 34% fewer tokens than the average user this week" is more effective than any environmental messaging. The comparison is built in.
Single HTML file. Zero dependencies. MIT licensed. You do not need to be a developer.
Highest priority:
- Phase 2 platform connectors — If you use Mistral, Cohere, Qwen, or Groq, open an issue with the API response format and the token field names. That's the whole spec.
- Better water math — Per-provider PUE and water intensity data. Open a PR against METHODOLOGY.md with a source.
- Browser extension — Auto-captures token counts without manual entry. Highest-leverage expansion.
- Translations — Japanese, Spanish, Arabic, Hindi are highest priority.
- Cloudflare Worker — Community impact counter backend. See CONTRIBUTING.md.
- No frameworks. Vanilla HTML/CSS/JS.
- No server. Everything runs client-side.
- localStorage only. API keys never leave your browser.
- Single file. Fork it, host it on GitHub Pages in 5 minutes.
- MIT licensed. Take it, improve it, redistribute it.
Built by Ryan Lester — Navy veteran, economics student at the University of Houston, foresight researcher. TokenWater started as a thought experiment about invisible infrastructure costs and became a tool because nothing like it existed.
Essays on systems, power, and the future: RLPerspectives.com
MIT licensed. Take it. Make it better. The water is still being used whether you track it or not.
v1.0 — MIT License — rlperspectives.com