💧 Free · no sign-up · nothing leaves your browser
AI Water Cost Calculator
What did that prompt actually cost the planet? Pick a model, enter total tokens and requests, and read the money, litres of water, kilowatt-hours and CO₂ it takes to run.
Order-of-magnitude estimates built from published per-token energy use, datacenter water intensity and grid carbon intensity. Change any input to explore the trade-offs.
Estimate a workload
Readouts update as you type — no button required.
Your workload
Inference only — training, your device's power and network transmission to your screen are excluded. See the assumptions below.
How it works
Each model carries three coefficients: blended price per 1M tokens, energy per 1M tokens, and datacenter water intensity in litres per kWh.
Total tokens is what drives the numbers; the request count adds the fixed per-call overhead for network and orchestration.
Cost in USD, water in litres, energy in kWh and carbon in grams of CO₂e — plus a per-request and per-million-token breakdown.
Where the numbers come from
Four formulas, three shared constants. Change an input and every figure moves with it.
kWh × water intensity (L/kWh). Datacenters evaporate water for cooling and power plants consume it upstream; combined intensity runs about 3.5 L/kWh for hyperscale cloud and higher for older and self-hosted sites.
kWh × 390 g CO₂e/kWh, the approximate US grid average. Regions with cleaner grids will land lower; coal-heavy grids higher.
Energy per 1M tokens scales with active parameters and batching efficiency, from ~120 Wh for a small model to ~900 Wh for a dense 70B on rented GPUs.
Each call adds $0.00002 and 0.03 Wh regardless of size — negligible at scale, decisive when requests are short and numerous.
Published per-query figures differ by more than 10x because they assume different hardware, batching, cooling and grid mixes. Treat the output as an order of magnitude, not an invoice.
Model training, idle capacity, embodied hardware, and the electricity used by your own laptop, phone or network — all real, none of them per-prompt.
Model coefficients used
Blended prompt + completion pricing with datacenter energy and water intensities.
- Price
- $4.00 / 1M tok
- Energy
- 600 Wh / 1M tok
- Water intensity
- 3.5 L / kWh
Frontier multimodal; blended ~$4 per 1M tokens.
- Price
- $0.30 / 1M tok
- Energy
- 120 Wh / 1M tok
- Water intensity
- 3.5 L / kWh
Small model — roughly 5x cheaper and lighter per token.
- Price
- $1.00 / 1M tok
- Energy
- 200 Wh / 1M tok
- Water intensity
- 3.5 L / kWh
Legacy dense mid-size model.
- Price
- $6.00 / 1M tok
- Energy
- 700 Wh / 1M tok
- Water intensity
- 3.5 L / kWh
Frontier model, long-context friendly.
- Price
- $4.50 / 1M tok
- Energy
- 550 Wh / 1M tok
- Water intensity
- 3.3 L / kWh
Sparse mixture-of-experts frontier model.
- Price
- $0.60 / 1M tok
- Energy
- 900 Wh / 1M tok
- Water intensity
- 4.5 L / kWh
GPU-rental equivalent; dense 70B batches less efficiently.
- Price
- $2.00 / 1M tok
- Energy
- 500 Wh / 1M tok
- Water intensity
- 4 L / kWh
Blended stand-in when the model is unknown.
FAQ
How accurate is this?
It is an order-of-magnitude estimate, not a measurement. The physics (energy → water and carbon) is solid; the uncertainty sits in the per-model energy intensities and in where the datacenter draws its power and water.
Why does water show up at all?
Two places. On-site, datacenters evaporate water to reject heat. Off-site, the power plants feeding them use water for cooling too — often more than the datacenter itself.
Is a self-hosted model really cheaper?
Per token, usually yes — the row uses a typical GPU-rental equivalent rather than an API list price. But self-hosting is often less energy-efficient per token because your batch sizes are smaller.
Does 1M tokens mean 1M words?
No. One token is roughly 0.75 of an English word, so 1M tokens is about 750,000 words. Use the token counts from your API dashboard for real workloads.
Is anything sent to a server?
No. The coefficient table ships with the page and every calculation runs in your browser. Nothing you type is uploaded or logged.