AI, The Better Way
We have been told there is only one way to build artificial intelligence: bigger. Bigger models, bigger buildings, bigger power plants, bigger checks from a handful of companies. If you want the benefits of AI, the story goes, you rent them from the people who own the warehouse.
That story is a choice, not a law of physics. The same chips that fill a hyperscale campus can sit in a closet, a basement or a rooftop enclosure, powered by the sun and wind, owned by the people down the hall. This post is about that choice: what is wrong with the way AI is being built today, and what a better way looks like in practice.
What’s wrong with the way AI is being built
It is an energy land-grab. Data centres used about 415 terawatt-hours of electricity in 2024, roughly 1.5% of all the electricity on Earth. The International Energy Agency expects that to more than double, to about 945 TWh, by 2030. In the United States, data centres account for nearly half of all growth in electricity demand between now and 2030.
It piles the burden on a few places. Nearly half of US data centre capacity sits in five regional clusters, and half of the capacity under development is going into clusters that already exist. Those communities absorb the transmission lines, the grid upgrades, the noise and the rate pressure. The profits go elsewhere.
It drinks from the local tap. US data centres consumed an estimated 17 billion gallons of water for cooling in 2023, and about 211 billion more through the power plants that run them. A single Google campus in Council Bluffs, Iowa used about a billion gallons in 2024.
It wastes what it cannot sell. Every watt a server uses becomes heat. Giant facilities spend still more energy and water to dump that heat into the air, while schools, pools and apartment buildings nearby pay to heat themselves.
It moves your data, and your power, away from you. When intelligence lives in a distant cloud, your records, your prompts and your bargaining power live there too. AI is not a disembodied mind but an extractive supply chain of minerals, megawatts, water and underpaid human labor.
The better way: community mini data centres on sun and wind
Picture AI that lives in the building instead of the cloud. A small, shared AI appliance sits in a utility room. Solar panels on the roof, and a small wind turbine where the site allows, feed a battery. The battery feeds the compute. The heat warms the hot-water tank. The building’s own broadband connects every apartment to it, and a link to the wider internet connects it to its neighbours.
This is no longer science fiction. Desktop-class AI appliances such as the ASUS Ascent GX10, built on NVIDIA’s GB10 chip with 128 GB of memory, ship with a 240-watt power adapter. Reviewers measured it at about 42 to 47 watts idle and under 200 watts under load. Running flat out all day, that is under 5 kWh a day. In a sunny climate, a rooftop array of roughly 1.5 to 2 kW plus a home-sized battery can carry it (an approximate sizing, worth checking site by site).
The principles are simple:
- Generate where you compute. Pair every node with its own solar, wind and storage, so new AI demand adds clean supply instead of straining the grid. Electricity that does not travel far also avoids much of the roughly 5% the US grid loses in transmission and distribution.
- Run work when the sun shines. Training jobs, indexing and batch tasks wait for surplus power; everyday questions run anytime from the battery.
- Put the heat to work. In Exmouth, England, a washing-machine-sized data centre supplies about 62% of a public pool’s heat, saving about $24,000 and 25.8 tonnes of CO2 a year. A building node can do the same for showers and radiators.
- Keep data at home. Records, prompts and fine-tuned models stay inside the building’s boundary, under rules the residents set.
- Connect as a mesh. Many small nodes share capacity over community-owned fiber and wireless links, so the network grows one building at a time and no single failure takes it down.
Not every job belongs in a basement. Training the largest frontier models still needs big clusters. But most of what people actually use AI for (answering questions, summarizing documents, translating, tutoring, searching a group’s own records) runs well on open-weight models that fit on one appliance.
Who owns it matters as much as how it runs
A solar-powered server owned by a monopoly is still a monopoly. The better way is as much about governance as hardware.
- Cooperative ownership. Residents, tenants’ associations, credit unions, libraries or worker co-ops own the nodes together, one member, one vote.
- Community broadband as the backbone. Building-owned wholesale networks, like the multi-unit model the Broadband Institute Foundation pioneered at the 200-unit Veridian at County Farm in Ann Arbor, already connect every unit. The same wiring can carry local AI.
- Mutual aid as the economy. Members contribute time, skills and spare compute and earn the same in return, an hour given for an hour earned, rather than paying rent to a platform.
- Public partners. Municipal utilities, community choice energy programs and cities can host nodes, buy services, and treat local compute like any other public infrastructure.
This is what the solidarity economy has always done with electricity, water, credit and housing: take essential infrastructure out of the hands of distant owners and put it under democratic control.
Honest tradeoffs
A better way still has to be built, and it helps to name the hard parts.
| Objection | Honest answer |
|---|---|
| Small nodes can’t run the biggest models. | True. Frontier training stays in large clusters for now; most daily use runs on open-weight models that fit one appliance, and nodes can pool capacity. |
| Solar is intermittent. | Batteries cover nights; flexible jobs shift to sunny hours; the grid stays a backup, not the main course. |
| Many small sites are harder to maintain. | Shared tooling, remote monitoring and trained local stewards make it manageable, and create local jobs. |
| Big data centres are more efficient per server. | Efficiency inside the fence ignores the water, the transmission losses and the wasted heat outside it. |
| Hardware still comes from extractive supply chains. | Yes. Buying less, keeping it longer, sharing it widely and repairing it locally shrinks that footprint. |
| Who pays up front? | Co-op equity, community solar financing, grants and public procurement, the same tools that built rural electric co-ops. |
The choice in front of us
A century ago, big utilities decided rural America wasn’t worth wiring. Farmers formed electric cooperatives and wired it themselves. Today, many of those co-ops still deliver power to their members.
AI is at the same fork. We can let a few companies build ever-larger warehouses that drain local grids and aquifers, or we can build intelligence the way communities built power, water and broadband: small, local, renewable and owned by the people it serves.
What you can do now:
- Ask your building, co-op or city what it would take to host a shared AI node.
- Map the roofs, broadband and waste-heat users in your neighbourhood.
- Join a cooperative AI or community broadband effort, or start one.
- Ask every AI vendor where its electricity and water come from, and who owns your data.
AI, the better way, is not bigger. It is closer, cleaner and ours.
