Build AI closer to home.

Before a data center gets a permit or a tax break, its owners show how much water, land and electricity it needs and who pays for power upgrades. Communities get limits they can enforce. Where it makes sense, run everyday AI on shared computers in buildings we already have.

A typical 250,000-square-foot data center is bigger than four football fields. Industry representatives told Virginia’s auditors it may have about 50 full-time workers, about half of them contractors. A small community cluster can create a different kind of opportunity. Local people can learn to build, run and maintain it, and train in machine learning, networking, computing, inference, AI agents and agentic coding. Each person trained can help train the next. The people who maintain the community system can go on to find good jobs in AI. We want to start with disadvantaged communities.

Cooling fans can make a steady low hum, and backup generators can add bursts of noise. Measure both where people live, during the day and at night. Ask residents how the noise affects their daily lives, sleep and well-being. Set limits people can enforce and give neighbors someone to call.

Don’t send a big request to a distant data center when a small model close to you can do the job. Run everyday work on the smallest model that does it well, on your own device or the local cluster. Save useful work so it isn’t done twice. Let idle machines sleep or power down, for example overnight. Send work to the cloud only when it is needed. This is a thesis we want to test.

Your private work stays on your device or a shared computer you trust. If something needs to go to an outside service, you are told first and you choose. You can see and delete what is saved about you, and private questions never go into shared answer libraries.

If an AI system goes rogue or behaves dangerously, people need a way to shut it down that does not depend on the AI itself. This is a hard, complex problem, and it may be beyond what a small community pilot can solve. We want to look into experiments on how to shut down or isolate a misbehaving AI, starting with our own cluster, and to share what we learn. Essential services should keep running safely under human control.

This is technically and computationally very complex.

Some knowledge should not be fed to big cloud AI. AI can supercharge information gathering, and that can turn extractive: taking a community’s knowledge without consent or benefit, which some call digital colonialism, or watching people who never agreed to it. Communities decide how their knowledge and data are used, what to pass on, what to keep private and who may use it. Ask permission first, keep the work on local computers, and let people withdraw access.

Who needs this protection

• Tribal nations
• Conservation and wildlife groups
• Elders and oral historians
• Traditional healers, farmers and seed keepers
• Small language communities
• Artists, writers and musicians
• Journalists
• Dissidents
• Unions and organizers
• Clinics, shelters and mutual-aid groups

Make local AI publicly accessible in community spaces, such as libraries, community centers and cooperatives. Offer a powerful standalone AI workstation or shared terminals connected to a local cluster, so people can use local AI without buying expensive equipment. The cluster can sit in the same building or nearby. The community decides who gets access, what to spend and who runs the service.

If fifty neighbors ask when the town picks up bulky waste, check the answer once and save it with the official source. Reuse it when the same public question comes up, and update it when the facts change. Personal questions are never saved to the shared library. Over time, checked answers could grow into a community archive that neighbors can search and share privately.

Run open-source software on computers in the community, so anyone can inspect, fix and adapt it. Prefer AI models you can download and run yourself, and say plainly how open each one really is: its code, its model weights, its training data and its license. Publish what you use, what it costs and how well it works, so another community can try it too.