I kind of love this hand-crank AI because it protests everything wrong with big tech
Just about every week another AI company announces a bigger model, a bigger data center, and a bigger electricity bill — while pocketing billions and quietly walking back its climate pledges. It’s pretty disheartening background noise. Then along comes CrankGPT, a tiny chatbot that refuses to answer your question unless you physically generate the electricity yourself by turning a hand crank. And honestly, I kind of love it.
What CrankGPT Actually Is
CrankGPT is a project from Squeez Labs — a voice assistant that runs entirely on a Raspberry Pi 5, powered by a 20W hand-crank generator. No Wi-Fi or cloud. No data center. Just your arm, a crank, and a local language model doing inference on-device.
The tech stack is impressively thoughtful for something this small. It runs llama.cpp with compact models like Liquid AI’s LFM2 350M and Gemma 3 1B, using Moonshine for speech recognition and Piper for text-to-speech. The whole thing boots in about 30 seconds from a cold start.
Squeez Labs built a custom capacitor board to smooth out the generator’s voltage spikes — because when the LLM and speech synthesis fire together, the crank literally gets harder to turn. You feel the inference load in your arm.
That last part is the whole point.
Why the Inconvenience Is the Feature
Squeez Labs isn’t pretending this is practical. They’re making a point: the cost of AI is real, you just don’t see it.
Right now it’s completely invisible. You type a prompt, get an answer, and somewhere a server farm draws more power. Global data center electricity consumption hit 460–490 TWh in 2025, up 17% year over year. AI infrastructure alone drove a 50% energy surge — growing at five times the rate of overall global electricity demand. The IEA projects that number will double by 2030.
Meanwhile, tech companies that once posted ambitious climate pledges have quietly started building gas-burning power plants to feed their GPU clusters. The five largest hyperscalers committed roughly $660–690 billion in capex in 2026, with about 75% tied directly to AI infrastructure. That money came from somewhere. Mostly from you, paying for subscriptions and API calls at margins that make the underlying energy cost look trivial.
CrankGPT makes the cost visible again. You want an answer? Then crank, baby.
The Broader Shift Toward Local AI
What Squeez Labs built as a provocation, a lot of people are building seriously. Enterprise AI inference has shifted dramatically, with 55% of enterprise AI inference now running on-premises or at the edge — up from just 12% in 2023. Privacy is the main driver. Running a model locally means your data never leaves your device, which matters for healthcare, legal, and anyone who’d prefer their most personal questions stay that way.
The hardware has gotten good enough to make this practical. Squeez Labs notes that smaller models on modest hardware can handle a surprising range of tasks — and that the industry’s reflex to reach for the biggest, most expensive model is often just laziness dressed up as quality.
The Uncomfortable Point
I’m not suggesting you swap ChatGPT for a hand crank—though that might be fun, for a while. But I think CrankGPT is onto something real. We’re so used to “free” AI that it’s easy to forget every prompt consumes actual resources somewhere — electricity, water, land, capital. Tech executives are gaining extraordinary profits while those costs get externalized onto the grid, the climate, and everyone who pays an electricity bill.
A project that makes you feel that cost in your bicep is a more honest interface than a clean white text box.
Squeez Labs isn’t selling CrankGPT — they open-sourced the whole build for under $300 in parts. That’s kind of the point too. The answer to resource-extractive AI isn’t more resource-extractive AI. Sometimes the right move is to make something smaller.
