Offline Base
Shop
The sustainability case

AI that doesn't need a power plant.

Every cloud query spins up hardware in a data center that runs hot, drinks water, and pulls from the grid around the clock. Offline Base answers from a box on your desk that sips power.

Watts, not megawatts

Base Router draws between 7 and 25 watts. That's the draw of a lightbulb, not a server hall.

No round trip

Your question doesn't travel hundreds of miles to a data center and back. It's answered in the room you're sitting in.

Hardware you keep

One box you use for years beats a constantly refreshed fleet of cloud hardware you never see and never own.

Watch the difference: a data center vs. your box
data center

one hall of thousands · running whether you ask or not

megawatts, around the clock

your box

7–25 W

the draw of a lightbulb, in the room you’re in

watts, only when you ask

The receipts

415 → ~945 TWh

Data-center electricity, 2024 to 2030 (projected)

The IEA counted about 415 terawatt-hours of data-center electricity use in 2024 — roughly 1.5% of the world's total — and projects it to more than double by 2030, with AI the most important driver. Base Router's answer to that curve draws 7 to 25 watts.

Source: IEA, Energy and AI

8.1 billion gallons

Google's water consumption in 2024, up 28% in one year

That's from Google's own environmental report, which also shows its data-center electricity use grew 27% the same year and has doubled over four years. A box on your desk doesn't need cooling water at all.

Source: RCR Wireless

0.24 Wh per prompt

Google's estimate for a median Gemini text prompt

Plus about five drops of cooling water. Straight talk: per prompt, that's small. It's also Google's own math, not independently verified, and covers only median text prompts — not images, video, or long documents.

Source: Google Cloud Blog

~43 tokens/sec at 25 W

NVIDIA's benchmark of the board inside Base Router

NVIDIA measured about 43 tokens per second on a 3B model with the module in its full 25-watt mode (module power — the wall plug adds a little more). By that spec-sheet arithmetic, not a lab measurement, a typical 200-token answer costs on the order of 0.03 watt-hours.

Source: NVIDIA Developer Blog

The greenest query is the one that never leaves the building.

No overclaiming

The honest part.

We won't pretend a desktop box is carbon-free. It draws power and it was manufactured. The honest claim is narrower and real: for everyday local tasks, it avoids the data-center overhead entirely, and it keeps working for years instead of being churned.