Lend the dead hours, borrow the whole garden.

This is the design we're building toward. Your gaming card sits idle most of the day. Lend those hours to the co-op, and when you need it, borrow the combined power of dozens of cards: enough to run models far too big for any one gaming PC.

1. Install the app and plant your GPU. It only runs while your machine is idle. The moment you sit down to play, the card goes back to you within a second or two.
2. Open models run across volunteer cards. Big models like Qwen and Llama are split into slices, with each card holding one slice. A smaller model fits whole on one card. Your unfinished answer is never lost: the coordinator holds the conversation, so if a card drops out, the work moves to another.
3. Your queries stay private. Prompts are never logged. Whoever runs a model's first slice could see the words, so private prompts go only to trusted or hardware-verified machines, and the app tells you which route you're on.
4. Verified work earns credits. Every hour your card does real work earns credits you can spend on AI. To keep it honest, the same job is quietly checked on a second machine, allowing for the tiny differences between graphics cards. If they don't match, that machine is benched, not banned: its owner sees the evidence and gets a chance to fix it. Broken is not the same as cheating: an unstable overclock or an overheating card could potentially cause wrong answers without the owner knowing, so a mismatch starts as a question, not an accusation.
5. Everyone gets a free base ration. Rig or no rig, every member gets a free daily ration whenever there's capacity. Contributors earn more on top and get served first. If free use ever has to be trimmed, the reason is shown to everyone, live.

When a model is split across cards: your words go into the first card, which turns them into the model's working math and passes that along. The other cards only see the math, not your words. It isn't encryption, and in principle parts of the text could be pieced back together, but no card in the middle sees your prompt plainly. The last card writes the reply, so it sees the answer.

Safeguards for those who lend

• Power limits keep your card running cooler and ease wear on your hardware.
• Sandboxed jobs can't touch your files or other devices on your home network.
• The co-op takes responsibility for what the network runs, not you.
• You can pause or pull your card at any time.

Only approved open models with a published fingerprint will run. Weights ship in a safe format, the worker runs in a sandbox, and updates are signed. Jobs are handed out at random, and content filtering happens at the coordinator, never on a volunteer's machine.

The captain's log

A live public dashboard will show capacity, usage, queue length and how the free tier is doing, plus the history of every throttle, outage, ban and policy change, along with the reason for it.

Root Compute is designed as a cooperative. The people who use it and power it will own it, and it can never be sold out from under them.

• One member, one vote. Votes aren't weighted by card size: a GTX 1070 counts the same as a rack.
• An asset lock in the charter. The co-op can't be sold for private gain or turned for-profit.
• Surplus goes back to members as credits, or as more co-op-owned GPUs.
• Public books. Every dollar is shown. A paid tier funds co-op-owned GPUs: the steady backbone and the trusted machines for private prompts.

Turnout sinks, a board settles in, and one day a "strategic offer" goes to a vote nobody is watching. That is how big mutual insurers like MetLife and Prudential were converted into for-profit companies around 2000: a one-time payday, and the member ownership never came back. So the protections will be written into the charter and bylaws from day one:

• Membership can't be bought, sold, transferred or stacked. One person, one vote. No one can buy up seats or pool votes by proxy. (The founders' temporary extra say, described below, is the one stated exception, and it ends.)
• Nothing to cash out. If the co-op is ever dissolved, what's left goes to another co-op or a nonprofit, never into members' pockets. With no payday to chase, there's far less reason to try.
• A supermajority of all members. Any sale, conversion or demutualization, or any change to these protections, needs it, counted against everyone who belongs, not just those who show up to vote. Apathy blocks a sell-off instead of enabling one.

Why have a legal structure at all?

So no volunteer or founder is personally on the hook, and so the promises above are written into a charter that binds the co-op rather than resting on goodwill.

If we ever fold, the garden stays

If Root Compute the organization ever folded, the open models and the software would still sit on everyone's machines. The aim is that the entity holds almost nothing the network needs to survive: you don't have to trust us, because nothing depends on us. That isn't fully true yet (the connection layer, the keys and the hubs still need work), and we'll say so as it changes. Whatever connects the members has to be swappable, and it can't rely on a single coordinator that one person controls.

Several keyholders, not one

No one person holds the keys. Changes take a majority of member-elected keyholders, and every change is public.

No surveillance work

License-plate cameras that record where people drive are the clearest example: the people in those records never agreed to be in them. Work that watches people who never agreed to it doesn't run here.

Said plainly While the co-op is being built, the founders will hold extra say. That will be written into the charter and public from day one. It hands over automatically to one member, one vote once the mesh reliably serves a big open model, cheat-checking and the captain's log are live, the co-op is legally formed and has held its first election, and the treasury pays its own way, or after five years, whichever comes first. The core principles are locked so that even the founders can't change them. The legal entity isn't formed yet, and a co-op lawyer will review it before launch.

The hard parts Named rather than hidden: latency between home GPUs, machines coming and going, the cost of checking work, privacy on hardware we don't control, legal liability, and making sure nobody can capture the governance. The plan answers each of these, and the code will be the track record.