Most of the conversation about AI is about the ceiling: the largest models, the newest capabilities, who gets there first. That race matters, and the labs running it are good at it. My concern is the other end. As these tools become the way people work, learn, and get things done, who guarantees that a capable version stays within reach of everyone, and not only whoever can afford to meter it?
Everyone deserves a dependable baseline of intelligence they actually own. Not the best model in the world. A good one that cannot be taken away.
The worry is simple. Capable AI is drifting toward a handful of providers who can raise the price, change the terms, or switch access off. That is a fragile place for something people are starting to depend on. A floor changes the picture. If a genuinely useful model runs on hardware you own, works when the connection does not, and answers to your rules rather than someone else's, then the ceiling can be a race and the floor stays solid underneath it.
What it could look like
This is a sketch, not a spec. The shape I am reasoning toward is local-first. A capable, open-weight model runs on hardware the user owns and keeps working offline, with no single point above it that can turn it off. That is the whole point of the floor: it holds even when the things above it move.
From there, nodes could federate by choice. Someone who wants more than one machine can offer should be able to join a pod of people they trust, and a capability router would spill heavier work outward, from your device to a trusted pod to a wider federation, but only within limits you set. Nothing leaves your control unless you say it can. The default is local; reaching out is a deliberate choice, never the price of entry.
What it would not have is just as defining. No token, no paid marketplace, no account you can be locked out of. The value is the floor being there, owned by the people standing on it, rather than a service someone rents to them.
Why say it now, before it is built
Because the ideas are worth reasoning about even at the sketch stage, and because getting them right early matters more than shipping something quickly. Putting the proposal in the open invites the objections and prior work that would make it better, or show where it is wrong. I would rather be corrected on a document than on a thousand deployed machines.