EVA and the Data-Center Question — Offline AI Without the Grid or Water Footprint
The Data-Center Question
Every cloud AI answer is manufactured inside a building. EVA's are not.
When you send a prompt to a cloud assistant, that answer is produced inside a data center. The building draws power from a regional grid, and most large facilities reject heat by evaporating water — water that does not come back.
📊 The Measured Baseline
Google is the only major provider to publish an audited per-prompt figure. In its May 2025 technical report, the median Gemini text prompt consumed:
- 0.24 watt-hours of electricity
- 0.26 millilitres of water — roughly five drops
- 0.03 grams of CO₂e
That figure is comprehensive: it includes idle standby capacity, host CPU and RAM, and cooling overhead — not just the AI accelerator.
One Prompt Is Trivial. The Aggregate Is Not.
The International Energy Agency puts global data centre electricity consumption at 415 terawatt-hours in 2024 — about 1.5% of the world's electricity — rising to roughly 945 TWh by 2030. That is more electricity than Japan uses today.
The IEA also warns that about 20% of planned data centre projects are at risk of delay because of grid-connection queues and equipment bottlenecks. The constraint is no longer chips. It is transformers, substations, land, and water rights.
EVA Runs on Hardware You Already Own and Already Power
| Per answer | EVA (offline) | Cloud AI |
|---|---|---|
| Data-center electricity drawn | None | Grid-supplied |
| Water consumed for cooling | 0 mL | ~0.26 mL per prompt |
| Load added to the grid interconnection queue | Zero | Cumulative |
| New facilities your usage argues for | None | Scales with demand |
- Zero water at the point of use. Your computer is air-cooled. It evaporates nothing. A session with EVA — a thousand prompts, ten thousand — consumes exactly 0 millilitres of water.
- Zero marginal data-center demand. EVA issues no query to any server. It contributes nothing to the load growth above, nothing to the interconnection queue, and nothing to the case for building the next facility, substation, or cooling tower.
- A fixed footprint, not a subscription to someone else's growth curve. You bought a drive once. There is no recurring server-side cost accruing behind your account for as long as you own it.
What EVA Changes
EVA does not make a data center more efficient. It removes your queries from one entirely. The electricity EVA uses is the electricity your computer was already drawing from your own outlet, on a machine that was already built and already connected to the grid.
EVA's language models were trained in data centers before they ever reached your drive. That training already happened. Running them locally adds nothing further to it — and nothing further to the grid, the build-out, or the water.
Sources
- Google Cloud — Measuring the environmental impact of AI inference (May 2025 analysis)
- International Energy Agency — Energy and AI, Executive Summary (2025)
- MIT Technology Review — Google releases data on AI prompt energy use
Figures verified September 6, 2026.