Community AI compute: a group buys the machine, not the hours

Community AI compute is a small group of people buying one AI machine together and sharing it, instead of each of them paying a large provider by the hour. Everyone puts in a share of the build cost, everyone holds a documented share of the box, and everyone gets time on it. After it is running, the only bills are power, connectivity, and whatever the group agrees to spend keeping it healthy.

It is deliberately unglamorous. Community-owned AI is not a financial product and nothing about it promises anyone money. It is a purchase, the same way four households buying one boat is a purchase. You share it because one person does not need a machine like this running all day, and one person struggles to justify the price alone.

The reason the question comes up now is arithmetic. A serious inference machine costs somewhere between $15,000 and $60,000 depending on the GPUs, and a person who wants a private, always-available, open-weight model rarely wants to write that cheque by themselves. Split four ways it becomes a large but ordinary purchase, and the machine is idle far less of the day.

Who community compute is actually for

It works when the members want roughly the same thing from a machine and do not all need it in the same second:

• People running open-weight models more or less continuously — agents, batch jobs, embeddings, overnight fine-tunes.
• People who cannot send their data to a third party at all: clinics, law practices, anyone working under a client-confidentiality clause.
• Small labs and research groups whose funding covers hardware once but not a cloud bill forever.
• People who want a model with nobody else's usage policy sitting on top of it.

It works badly for occasional, spiky use. If you touch a model twice a week for twenty minutes, renting is the correct answer, and we will tell you that before you spend anything. The self-hosted AI with friends guide walks through the duty-cycle question in detail, because it is the one that decides everything else.

You do not need to already know the other members

Most groups we form are people who did not know each other before the waitlist. That is not an edge case we tolerate — it is the service.

There are two ways in. Either you arrive with people already, three colleagues or a lab or a chat group that has been circling this for a year, or you arrive alone and we match you.

Matching is a paid service, 5% of project value, and it is more structured than people expect. It buys a waitlist sorted by tier, so somebody who wants a $12,000 machine is not put in a group with somebody who wants a $60,000 one; introductions; and a group agreement drafted before anyone spends a rupiah. The agreement is the part that matters. It records who holds which share of the hardware, how machine time is scheduled, what happens when a member wants out, and how decisions are taken — one member, one vote.

What the group actually buys

Five things, each priced on its own, each optional:

• Matching — 5% of project value, once, when the group closes.
• Parts from China — our cost plus 10%. We run the sourcing channel ourselves and you see the hardware invoices.
• Design — from $2,000. Hardware selection, power, cooling, network, the model-serving stack. The build plan belongs to the group.
• Assembly and commissioning — from $2,000. Built, burnt in, serving models, handed over working.
• Hosting in Bali — from $900 a month per machine, all-inclusive, plus $10 per member per month. Electricity, internet, monitoring and service.

A group can buy one of those or all five. Outside Bali we do the first four and the group hosts the machine wherever it likes — an office, a home rack, a local colocation cage. If you are deciding what hardware the group should even be arguing about, start with how to group buy a GPU server, which covers the tiers and what each one can actually serve.

Nobody collects the pot

This is the part of shared AI compute that people are right to be suspicious about, so here is exactly how it works.

Each member pays Heroic24 LLC directly, for their own share of each service — a deposit at order, the balance at delivery — as ordinary invoices for goods and services. There is no pooled fund. There is no escrow provider. There is no account anywhere with the group's money sitting in it, because that account never exists.

And we do not own the machine. Hardware title goes to the members on day one and stays with them; we hold no stake in it. We are paid for work we do — matching, parts, design, assembly, hosting — and not for a slice of anything the machine does afterwards. This is not an investment product, we never promise income, and the group agreement is written use-first for exactly that reason.

Start with the column where we lose

Renting wins more often than most write-ups about community-owned AI will admit. Four cases where it plainly does:

• Low duty cycle. A machine that idles twenty hours a day is a machine the group overpaid for. Rented compute bills only while it runs, and that is a real advantage, not a footnote.
• A moving target. If the group is still deciding which model family it needs, hardware bought this quarter locks in this quarter's answer for three years.
• Frontier-scale models. If what you actually want is the largest closed models, you cannot buy your way there. Nothing in this guide changes that.
• Nobody wants the admin. Someone in the group ends up being the person who notices a fan has died. If that person does not exist, the machine will teach you so.

Hosting moves that last problem rather than deleting it. Our Bali floor is $900 a month all-inclusive; a comparable single-GPU rented AI server from Hetzner runs around $965 a month. Those figures are close on purpose, and we are not going to pretend otherwise. The difference is not the monthly number — it is that after three years one group has a machine and the other has a folder of receipts. That is the honest case for owning, and it is the only case we make.

Run the numbers before you run a group chat

Every community AI compute project that went well started with a spreadsheet and not with enthusiasm. Two things to do in order.

First, price the machine and the month. Our cost calculator takes the tier, the group size and the hosting choice and gives you the build cost, each member's share, and the monthly figure per person. It is free and there is nothing to sign.

Second, price the workload rather than the hardware. If the group's real plan is a single large open-weight model served all day, the specific arithmetic in what it costs to split the cost of running DeepSeek is closer to your situation than any general guide, including this one.

The failure modes, plainly

Five things go wrong, and four of them are social:

1. One member goes quiet. Deposits are staged for this reason — parts are ordered against money already invoiced, never against a promise. But a member who stops answering during the design phase stalls everybody.
2. The schedule was never written down. Two members with overnight jobs and no agreed queue will find each other by the second week. Decide scheduling before the machine ships, not after.
3. Somebody wants out in month eight. Life changes. The exit and buyout terms belong in the group agreement on day one, when nobody is upset yet.
4. The machine was sized for the loudest member. The person with the strongest opinion about GPUs is not automatically the person with the heaviest workload. Size for measured use.
5. The room was wrong. Self-hosted groups underestimate power draw, heat and noise more than any other single thing. This is most of what Bali hosting is actually for.

None of these are reasons not to do it. They are the reasons the group agreement exists, and the reason matching is a service with a price on it rather than a chat thread that eventually goes quiet.