CrankGPT and the case for tiny, local AI

Forget the cloud. A new voice assistant called CrankGPT will only answer once a user has turned a handle for half a minute, and its makers argue that is the point. The device, built by computer scientist Katrin Tomanek and former Google ATAP technical project lead Alex Kauffmann through their company Squeez, is part provocation, part working prototype, and a deliberate answer to a question most AI companies prefer not to ask: what if an assistant did not need a datacenter at all?

CrankGPT looks like a small box with a crank on the side. Turning that crank charges a custom capacitor board that gives roughly 20 seconds of conversation before the system insists on more cranking. From the first turn of the handle to a working voice exchange, the documented cold start is 30 seconds. There is no web connection and no subscription tier, according to The Register's writeup of the project.

Squeez's pitch is not that crank power is the future of consumer AI. It is that small, private, specialized models running on cheap local hardware are. CrankGPT is the most theatrical version of that argument, a tongue-in-cheek demo designed to make the energy cost of an AI query something the user can feel in their forearm. The same hardware, Squeez argues, can host voice agents tuned for regional accents and speech impediments, or domain experts for gardening and auto repair that refuse questions outside their lane.

Kauffmann has framed the general-cloud alternative as "swatting a fly with a wrecking ball." The hyperscaler model, in this telling, pours water and electricity into answering questions a small, purpose-built model could handle in a living room. The hand crank puts that trade-off in physical form: every answer carries a visible mechanical cost.

The honest caveats matter. CrankGPT is a demonstration, not a product line. Squeez is not pitching spin-class AI stacks to engineering teams, and the 30-second startup and 20-second capacitor window are real user-experience costs, not bugs to be optimized away. The claim that small specialized models beat big general ones at most everyday tasks is Squeez's own argument, supported by their prototype rather than independent benchmarks. Treating it as proven at scale would overstate the case.

What the device does show, even as a novelty, is that the energy footprint of AI is a design choice, not a law of nature. The crank is the punchline. The engineering underneath it points to a different inference economy, one in which the cost of an answer is small, local, and visible to the person asking.