Our Impact
of cooling water a single AI query is estimated to draw in a data center — water your reply doesn’t spend when the model runs on your own device.
Estimated, not measured. Eco keeps no per-query telemetry. See the methodology below.
Water
Every query to a cloud AI drinks a sip of water.
Large AI data centers rely on evaporative cooling towers that consume enormous volumes of fresh water. Microsoft's 2023 sustainability report revealed a 34% year-over-year increase in water consumption [2], driven largely by AI workloads. Google's data centers consumed 5.6 billion gallons of water that same year [3].
Researchers at UC Riverside found that a single conversation with GPT-4 uses roughly 500 mL of cooling water — about a full water bottle — which works out to an estimated 250 mL per query [1].
When the model runs on your own device, that query never reaches a data center, so it doesn't draw on a cooling tower at all. Your laptop or phone still uses electricity and may warm up, but it carries no evaporative-cooling footprint to spend.
per query — data-center API call
data-center cooling — on your device
Energy
The device in your hand was already on.
Data centers currently consume 1–1.5% of global electricity [4]. Goldman Sachs projects that AI workloads alone will drive a 160% increase in data center power demand by 2030 [5].
A small model running on the laptop or phone you already own draws a tiny fraction of what a dedicated data-center GPU rack pulls to serve the same query — and it skips the data center's surrounding overhead entirely: the cooling, the networking, the idle capacity kept warm for the next request.
We don't claim a precise per-query energy number — consumer hardware varies far too much for an honest figure. What we can say is that running AI on a device that's already powered on adds no new power-hungry infrastructure to the grid.
of global electricity — data centers
new infrastructure — on your device
Privacy
The greenest query is the one that never travels.
Centralized AI forces a false trade-off: use a service that collects your data, or don't use AI at all. Eco answers a third way — the model runs on your own device, so the conversation has no reason to leave it.
That privacy is also why the impact adds up. A query that never reaches a data center has no remote GPU to power and no cooling tower to feed. The two stories are the same story: keeping the work close to you keeps it light on the planet.
To be precise about what “on your device” means — your conversation can still be saved locally in your browser's storage so you can return to it. It just isn't shipped to us, or to anyone else, to generate a reply.
Runs in your browser
The model is downloaded once and runs on your own device
Stays on your device
Your prompts and replies aren’t sent to a server to answer them
No data-center round-trip
No remote GPU to power, cool, or trust with your words
How we calculate impact
- Water savings per query
- Each AI query to a traditional data center is estimated to use about 250 mL of cooling water — the midpoint of the 200–300 mL range identified by researchers at the University of California, Riverside for GPT-4 class models. When the model runs on your own device, that query never reaches a data center and so avoids this evaporative-cooling footprint. Your device still consumes electricity and may warm up under ordinary hardware cooling.
- What this figure is
- A published research estimate, not a measurement of your usage. Eco keeps no per-query telemetry, so the number above describes the data-center footprint a typical cloud query would carry — not a count of queries you’ve run.
- What we don’t count
- We don’t claim carbon offsets. We don’t publish a precise per-query energy saving — there are too many variables in consumer hardware configurations to make an honest figure. We report only the water estimate, clearly labeled as an estimate.
est_water_per_query ≈ 0.25 L of data-center cooling avoidedOur methodology is deliberately conservative. We'd rather understate our impact than overclaim it. All source code — including the calculation you see on this page — is open source under AGPL-3.0.