A seat holds for $1.88
Your email and a $1.88 deposit put your name on one of 1,500 seats when we accept you. The $1.88 comes off the price of your seat. If we cannot seat you, we refund it.
A private compute reserve
Stop renting tokens.
Three years of the best open-weights model on a private DGX node. No meter and no logs. You pay once.
What a seat buys.
No limits
No weekly caps and no five-hour windows. Start a run at 3am on a Sunday and nothing meters it. At peak hours a fair-usage stamina system paces the heaviest concurrent sessions so the node stays up.
Privacy
We never write prompts or completions to disk. They live in memory for as long as your request runs, and nowhere else. The node keeps token counts and timestamps for fair-use accounting, and nothing more.
Newest model by vote
When a stronger open-weights model ships, members vote and the node upgrades.
36 months guaranteed
Thirty-six months of node uptime. We add outages back to the end of the term for every member at once. If the machine cannot continue, we sell it early and pay every member an equal share of what it raises, under contract.
Drop-in tools
One OpenAI-compatible endpoint. Point Claude Code, Codex CLI, Cline, or Aider at it and keep working.
End-of-life sale, split 1,500 ways
You never own the machine or any part of it. You fund its working life, so the money it sells for comes back to you. After the 36 months it is sold and 100% of the sale price is divided equally among the members. We keep none of it.
How it works.
Fifteen hundred seats buy the machine
The campaign is all-or-nothing. At goal, the seats fund a new NVIDIA DGX B300 in a Tier-3 datacenter, run for members alone. Short of goal, no one pays a cent and every deposit comes back.
One base URL, one key
Your seat speaks the OpenAI API, so the tools you already run point straight at it. Cursor takes a custom base URL in its settings, and so does Continue. Open WebUI gives you a private chat front end over the same key. Anything built on the official openai SDK needs one line changed, which covers LangChain, LlamaIndex, CrewAI and most of what you have already written. n8n reads it as a normal OpenAI credential.
$ export OPENAI_BASE_URL=https://node.example.dev/v1$ export OPENAI_API_KEY=seat-xxxx$ aider --model openai/kimi-k3
36 months, no meter
The clock starts at general availability and counts 36 months of node uptime. We add outages back to the end of the term for every member at once, so the calendar runs until all 36 are delivered. Then we sell the machine, publish the price, and divide 100% of it equally among the members. If the node ever cannot continue, that sale happens early instead and the money is divided the same way.
The bar counts uptime, so every hour the node is down pushes M36 right for all 1,500 seats at once. If the node ever has to stop short of 36, the machine is sold early and every member is paid an equal share of what it raises.
If the register fills.
Fifteen hundred seats will not stretch. Leave your address, and if you miss this run the next node opens to this list first.
Fair questions.
Seven questions, answered
What exactly does $1.88 buy?
A numbered place in line for one of 1,500 seats, plus first access in a 48-hour window before the campaign opens to the public. We credit it against your seat if you take one, refund it whenever you ask, and refund everyone unprompted if the campaign never opens.
Where are the caps?
There is no billing meter and there are no weekly caps, around the clock, for 36 months of node uptime. During the busiest hours a pacing system slows the heaviest concurrent sessions so the node stays healthy. Nobody gets locked out, nights and weekends run wide open, and a public status page shows the live load.
What if the machine breaks?
The node is bought new from NVIDIA with a three-year support contract, insured, and backed by a ring-fenced hardware reserve. If a repair takes it down for long, service bridges to rented GPUs behind the same zero-retention gateway, and any downtime is added back to the end of the term for all 1,500 members at once.
What if the campaign doesn't reach its goal?
Then we charge no one. The campaign is all-or-nothing: seats fund only if we hit the goal, and every deposit comes back in full.
How do I know you keep no logs?
Zero retention is built into the architecture, and the full data-flow diagram is published before the campaign opens. The node stores token counts and timestamps for fair use, and we never write prompts or completions to disk. They do live in memory while a request runs, which holds for every provider there is, so the privacy notice names the two people who could reach that memory, the hardware key it takes to do it, and the session recording it leaves behind. No provider can honestly promise you more than that.
What happens to the machine at the end?
We sell it, by a process we publish before the campaign opens: a defined listing period, a written definition of the net price, and sale costs capped at 8% of the gross. The money goes straight to an attorney-held escrow account, never through ours, and pays out from there. Before we pay anyone we publish the listing, the bids, the final price, the deductions, the per-member figure, and the certificate showing the machine was wiped. Every dollar of the net, split 1,500 ways.
Why $5,300?
The machine costs about $500K, and three years of datacenter, bandwidth, staffing, and reserves cost the rest. The full spreadsheet, including the operator's margin, is published with the campaign. It works out to $147 a month, under the price of a capped $200-a-month plan. And at the end the machine is sold, with 100% of the sale divided among the 1,500. This project is community funded, so the residual value of the hardware goes back to the community, in full.