The energy case for local AI, with our caveats

By our spec-sheet estimate, a 200-token answer on Base Router hardware takes about 0.03 Wh, against Google's 0.24 Wh for a median Gemini Apps text prompt. That gap is part of why the second belief on our about page (opens in a new tab) reads: "Running models locally answers the data-center problem." Both numbers carry caveats, and a Stanford preprint cuts against us on efficiency per watt.

How much electricity data centers use

Three sources set the scale.

In Berkeley Lab's high case, the US share nearly triples between 2023 and 2028 (4.4% to 12%).

Cheaper per answer, and many more answers

The second belief points to Jevons paradox. The cost per unit of intelligence keeps falling, but demand grows faster, so the total bill climbs anyway.

Google's own numbers show the pattern. In August 2025, Google reported that the energy for a median Gemini Apps text prompt had fallen 33x over the previous twelve months. Google's electricity use for 2025 still rose 37%. The two figures cover different spans and measure different things, so we read them as a pattern and leave cause and effect open.

The belief also says where big compute belongs: research, medicine and hard problems. Its example of a poor use is "another thank-you email." That kind of job fits a small model on your own machine, and our post What a million real questions say about small local models looks at which requests fall in that group.

Per answer: Google's figure and ours

Google measured a median Gemini Apps text prompt at 0.24 Wh, plus 0.03 grams of CO2e and 0.26 mL of water, about five drops (Google Cloud blog (opens in a new tab)). Google did the measuring, the figure has not been independently audited, and it covers median text prompts only. Images, video and long documents sit outside it.

Our figure is about 0.03 Wh for a 200-token answer on Base Router hardware. Here is the arithmetic:

1. NVIDIA's developer blog (opens in a new tab) lists about 43 tokens per second for Llama 3.2 3B on the Jetson Orin Nano Super.
2. We paired that speed with the board's 25 W power mode, the top of the 7 to 25 W range NVIDIA lists. NVIDIA's table gives the speed without naming a wattage, so the pairing is ours.
3. 200 tokens at 43 tokens per second takes about 4.7 seconds. At 25 W, that comes to about 116 joules, or roughly 0.03 Wh.

Our sustainability page (opens in a new tab) labels this spec-sheet arithmetic. It counts module power only, and the wall plug adds a little more.

The two rows describe different models doing different work, so treat the ratio between them as a rough guide at best. For more on the board, see Inside the Jetson Orin Nano Super, the computer behind Base Router.

One real reading from our fleet

We run open models on our own rack of Jetson Orin Nano boards, and our fleet page (opens in a new tab) shows their live power draw. At 21:40 UTC on October 14, 2026, the whole fleet was producing about 15 tokens per second at 26.3 W, or about 2,055 tokens per watt-hour. Four of its five boards were online.

That is one moment at light load. A couple of minutes earlier and a few minutes later (21:38 and 21:45 UTC), with no chats in flight, the same fleet drew 12.1 W and produced no tokens at all. Boards draw power while they wait, so tokens per watt-hour depends heavily on how busy they are.

For a sense of the gap, the spec-sheet arithmetic above implies about 6,200 tokens per watt-hour for one board at full power. Our fleet at light load reached about a third of that. Our post Two weeks of real traffic on our Jetson fleet covers a longer stretch.

The counterweight from Stanford

The first belief on our about page (opens in a new tab) leans on Stanford's Intelligence per Watt (opens in a new tab) preprint, which found that local models successfully answered 88.7% of a million real single-turn chat and reasoning queries. The same preprint has a finding that cuts the other way: local accelerators showed at least 1.4x lower intelligence per watt than cloud accelerators running identical models.

Model for model, then, the data-center chips in that study did more work per watt than the local chips it tested, such as Apple's M4 Max. It did not test Jetson boards, but we have no reason to expect them to do better. Our energy case depends on running a smaller model for everyday questions, on hardware that is already in the room. Our sustainability page states the claim this way: for everyday local tasks, the box "avoids the data-center overhead entirely."

What we can't claim yet

• Base Router's 7 to 25 W is a planned figure. Base Router is in development, and nothing we make can be ordered yet. Our Base Router preview has the rest of the plan.
• Base Drive is still coming soon. It runs on the laptop you already own, it is listed at $19.99, and our Base Drive post explains it.
• A box still costs energy to make and run. Our sustainability page says it directly: a desktop box draws power, and it had to be manufactured. The page's answer is that the box keeps working for years. Our computer buyback applies the same idea to machines people already have.
• Water and noise are open questions. The second belief calls water, power and noise real concerns, then adds: "That research is not fully established, and we say so."

The claim we stand behind

Our claim: everyday questions answered by a small model, on a laptop you own or a small box in the room, skip the data center. Our fleet page (opens in a new tab) keeps showing the live draw, so anyone can check the next reading against this one.