Solarpunk AI: what AI looks like in a solarpunk world

Solarpunk pictures a future where the technology works and the planet is still alive. Solar panels on every roof, trains instead of traffic, food growing on balconies, tools that get mended instead of replaced. The movement shows up in novels, album covers, video games, and actual city plans, and one idea runs through all of it, which is that technology should serve life rather than eat it.

AI has spent the last few years building in the other direction. New data centers, new power plants to feed them, water pulled from dry places for cooling, and a business model that charges rent on thinking. A solar-powered village with a repair cafe on the corner has no room for any of that.

There is a second version, and it already runs on the phone in your pocket. It points at a different future from the one being built in the desert.

What is solarpunk?

Solarpunk began as an aesthetic and grew into something closer to a design brief. The imagery is easy to spot, with plants climbing through architecture, wind turbines on the horizon, bright glass, cargo bikes, and people who look like they spend their days outdoors. Underneath the pictures sits a consistent set of preferences.

Energy comes from the sun and the wind, and it is generated near the people using it. Objects are built to be opened, understood, and fixed. Communities make decisions about the systems they depend on. Growth stops being the point, and durability takes its place. Nature is not a backdrop, it is part of the infrastructure.

Cyberpunk asked what happens when technology wins and people lose. Solarpunk asks what technology would look like if it were designed by people who planned to stay.

Why AI looks like the villain right now

Cloud AI runs on the largest computers ever built. Every question you ask travels to a warehouse full of accelerators, gets answered there, and travels back. Multiply that by billions of prompts a day and you get the industry we have now, one that measures progress in gigawatts and signs deals for nuclear plants.

The physical footprint is the part that clashes hardest with solarpunk. A single accelerator can pull several hundred watts while it works, and they run in racks, in halls, in buildings that need constant cooling. The energy is centralized, the hardware is unrepairable by anyone outside the company that owns it, and the whole arrangement is invisible to the person typing.

Then there is the ownership question. Cloud AI is rented. The model can change under you, get retired, get more expensive, or get restricted, and you have no say in any of it. A movement built on local control and long-lived tools has an obvious problem with a mind you have to lease by the month.

The version of AI that fits

Open models changed the shape of this. A language model is a file. Once it has been trained, running it is a much smaller job than making it, and small models have gotten good enough that a phone can handle real work. You download the file once, and every question after that is answered by hardware you already own.

That single change fixes most of the conflict. The energy comes from whatever charged your battery, which in a solarpunk world means a roof or a community array. The compute is distributed by definition, because it lives in devices people already carry. Nothing is rented. Nothing phones home. The model keeps working when the network does not, which matters more in a world of local grids and patchy coverage than it does in a city with fiber.

A phone battery holds roughly 15 to 20 watt hours in total, and it runs your whole day. A data center accelerator draws several hundred watts on its own, continuously, and it never works alone. Answering an everyday question on the device you are holding costs a rounding error of the energy that the same question costs when it travels to a warehouse and back.

Six jobs for AI in a solarpunk town

Abstract arguments about energy only go so far, so here is the concrete version. These are the jobs a small offline model is already good at, and every one of them gets more useful as the network gets less reliable.

Repair. This is the most solarpunk use of AI there is. Point a model at a broken bike derailleur, a seized water pump, or a solar charge controller that stopped charging, and let it walk you through diagnosis with no signal and no shop. Repair knowledge is exactly the kind of thing that used to live in a neighbor’s head, and a model on a phone puts it back within reach. Every fix keeps a working object out of the waste stream.

Growing food. Planting calendars, companion planting, crop rotation, what to do about the aphids on the beans. A gardener needs answers in the garden, which is usually the place with the worst reception on the property.

Sharing energy. Community solar and battery setups need scheduling, and scheduling is a reasoning problem. When to run the washing machine, when to charge the cargo bike, when to draw down the battery and when to hold it. A local model can reason about a local grid without exporting a full picture of the neighborhood’s habits to anyone.

Language. Small models translate well enough for daily life, and they do it without sending a private conversation to a company. That matters for neighbors who do not share a language, for arrivals in a new place, and for smaller languages that big services support badly or not at all.

Learning. A patient tutor that works in a classroom with no bandwidth is worth more than a brilliant one that needs a data plan. A school can hold a model on a shelf of secondhand tablets and use it every day, offline, for years.

Building. Local materials, load calculations, permit paperwork, water catchment sizing. The paperwork side of building anything is a wall that stops a lot of good projects, and a model that reads and drafts is a useful crowbar.

What has to be true for this to work

Four things, and all four are technical choices rather than wishes.

Models have to stay open. Weights you can download, keep, and run are the difference between a tool and a subscription. Formats like GGUF matter for the same reason standard screw threads matter.

Models have to stay small enough for hardware people already own. A model that only runs on this year’s flagship rebuilds the exclusion that local AI was supposed to remove. A model that runs on a five year old phone extends that phone’s life, which is worth more environmentally than most of the efficiency gains being marketed.

Offline has to be the default rather than a fallback. Software that quietly needs a check-in is not offline software, and it will fail on exactly the day the grid or the tower goes down.

Devices have to last. Local AI on disposable hardware is a wash. The gain shows up when a model keeps an old device useful for years longer than it would have been otherwise.

Where to start

You can run this today. The pieces are ordinary, and the whole setup takes about ten minutes on a phone you already have. If you want the background first, what a local LLM is covers how a model runs on your own device, and how to run AI offline on iPhone walks through the actual setup.

Offeline does this part. You download an open model once, and after that your phone answers every prompt on its own, in airplane mode, on a plane, in a field, or during an outage, with no account and nothing sent anywhere.

Solarpunk is a bet on the small, local, repairable, community-owned version of technology, and on the giant centralized version being a phase rather than a destination. AI is about to be a big part of how people think and work. Whether it looks like a data center in a drought or a file on a solar-charged phone is being decided now.