Compute funding for open source · Robinhood Chain

Your code, someone else's compute.

Every public GitHub repo can have one GPU pool on Robinhood Chain. Anyone funds it in USDG and holds the repo's share token. Only the verified maintainer can run compute — at a rate cap they set on-chain.

Pool asset
USDG on Robinhood Chain
Chain
Robinhood Chain · 4663
Shares
One share token per repo
Billing
Per second, capped on-chain
Illustrative preview · Compute offinterface example — not a live funding balance
ggml-org / llama.cpp
RTX 4090 · cap $2.50 / GPU-hour · goal 100 GPU hours
Goal progress: 55.2 of 100 GPU hours billed
Illustrative funded$950 USDG
Example balance$812 USDG
Illustrative GPU hours55.2
preview meter idle · open Pool workspace for live on-chain actions
Pools open—
USDG funded—
GPU hours—
The launchpad

A compute market for every repo, funded in rounds.

01Launch

A GitHub-verified maintainer opens a round: a USDG target, a deadline, and the repo's one share token.

02Contribute

Anyone contributes USDG. Everyone in the round gets the same price when it fills; a fresh pool starts at one share per USDG.

03Fund or refund

Hit the target and the money enters the GPU pool. Miss the deadline and contributors withdraw in full.

04Use it

Shares are access — each share buys daily requests to the repo's model — and a claim on unspent pool balance when compute is off and settled. When the pool runs low, open the next round.

How it works

Paste a repo. Let its users pay for the GPU.

01Fund

Paste any public GitHub repo and send USDG. The first deposit opens the pool and its share token.

02Verify

The repo admin signs in with GitHub. Once admin rights are confirmed, their Robinhood Chain wallet becomes maintainer.

03Run

The maintainer picks a GPU tier and the most they will pay per hour, then switches compute on.

04Bill

Compute is billed per second at or under the cap, with a receipt hash for every charge. No second is billed twice.

Compute goals

Support you can measure, in GPU hours.

Maintainers set the compute their repo needs. The community funds the target, progress stays public on-chain, and compute starts when the maintainer is ready.

10 GPU hours

About half a day of one GPU. Enough for release evals or keeping a demo alive through a launch.

100 GPU hours

Four days of one GPU around the clock. A small fine-tune, or a public endpoint through a busy week.

1,000 GPU hours

Six weeks of one GPU, or one week across six. The scale of a real training run.

The rules

Nobody bills more than the maintainer agreed to.

Rate capped on-chain

Every charge must satisfy amount × 3600 ≤ rate × seconds. Anything above the cap is rejected.

Every second billed once

Billing advances one clock. A second cannot be charged twice, and a future second cannot be charged at all.

The maintainer holds the switch

Only the GitHub-verified admin can switch compute on or off. The operator can stop a pool that runs dry.

Your share comes back

Burn shares while compute is off and settled, and your slice of the remaining pool returns to your wallet.

No rate changes mid-run

GPU tier and hourly rate can only change while compute is off, so every run keeps the terms it started with.

No key moves a balance

Money leaves a pool only two ways: a capped compute charge, or a holder burning their own shares.

The network

A permissionless GPU network for open source.

Idle GPUs become infrastructure for open-source code. Any developer can get compute, any GPU owner can provide it, and anyone can fund it.

Anyone can be a provider

Run the agent on an NVIDIA GPU, set an hourly rate, and post a USDG bond. No approval and no waitlist.

Matched automatically

Each job goes to the cheapest compatible GPU under the repo's cap, using the repo's repogpu.json.

Paid as it runs

Providers are paid for verified minutes, with heartbeat and GPU-check evidence behind every charge.

Reputation you can check

GPU checks passed, hours served, earnings, and strikes stay public for every provider.

Offline means rerouted

If a provider goes quiet, its lease ends, the gap is never billed, and the job moves to the next GPU.

One pool per repo

Every repo keeps one open USDG pool. Fund it any time and hold the same share token across rounds.

What pools holdUSDG
Share tokens1 per repo
Billing granularity1 second
Explore pools

Fund the repos you run on.

Paste any public GitHub repo. If it has no pool yet, creating it here reserves the repo in the RepoGPU registry; the first on-chain deposit opens its share pool.

—
Pools open
—
USDG funded
—
USDG spent
—
GPU hours billed

Open or find a pool

Checking contract configuration…

Loading pools…
Pool workspace

One repo. One GPU pool.

Load a registered repository, inspect its live contract state, then use your wallet for funding, rounds, redemption, and maintainer controls.

Load pool

Verify GitHub admin

Fund or redeem

Funding mints the repo share. Redemption burns shares and is available only while compute is off and the last run is settled.

Round actions

No pool loaded.

Maintainer controls

Only the wallet bound through GitHub admin verification can use these controls.

Open a round

The target is also the cap. A missed deadline can be refunded in full.

Request compute

Launch a raise

Raise for the compute your repo needs.

Rounds have one target and one deadline. Hit the target and every cent enters the GPU pool. Miss it and contributors are refunded in full. Only a verified maintainer can launch.

Round terms

Allocation is paid only if the round funds, and only from that round — earlier holders are never diluted by it.

What contributors get

At a $2,500 target, contributors claim the repo's shares at the pool value when the round funds. A fresh pool starts at 1 share per 1 USDG.

  1. Open a round.
  2. Target hit: funds enter the pool; shares are claimed at one price.
  3. Deadline missed: USDG returns in full.
  4. When the pool runs low, open the next round.
Providers

Rent out your GPU. Get paid in USDG.

Run the RepoGPU agent on the machine that owns the NVIDIA GPU, set your hourly price, and keep the agent online. When compute is requested, the operator can assign the cheapest compatible provider. The current release records verified work off-chain and settles from the pool only when its USDG balance is sufficient; a provider bond contract is not deployed yet.

—
Providers online
—
USDG charged
Local
Execution mode
$250
Bond target (roadmap)

1. Run the agent

Needs Python 3.9+, an NVIDIA driver, and Docker with the NVIDIA container toolkit.

cd provider
npm install
REPOGPU_API_URL=https://repogpu.xyz/api node agent.mjs

Run this on the machine that owns the NVIDIA GPU — for example your laptop. It signs a heartbeat every 15 seconds. On this server there is no NVIDIA GPU, so the check correctly reports nvidia-smi unavailable. Repository job execution stays disabled until a reviewed local runner is configured.

How you are paid

Price. You set USDG per GPU-hour. A pool leases you only at or under its cap.

Billing. For verified seconds only, after compute stops, at or under the maintainer rate cap and never above the available pool balance. A completed job with an empty pool stays blocked instead of pretending it was paid.

Bond. The original market design uses a provider bond and slashing for failed GPU checks. That bond contract is not part of the current pool deployment, so the web does not claim bonded USDG yet.

Docs

Everything RepoGPU does, and how to check it.

Overview

RepoGPU gives every public GitHub repository a GPU pool on Robinhood Chain, held in USDG. Anyone can fund a pool and receive the repo's share token. The verified maintainer decides when compute runs and the most it may cost per GPU-hour. Charges are billed per second, never above that rate, and every charge carries a receipt hash.

Money leaves a pool in exactly two ways: a capped compute charge, or a holder burning shares for their slice.

Maintainer verification

Only a GitHub admin can become maintainer — write access is not enough. The admin signs in with GitHub, RepoGPU checks permissions, and the chosen EVM wallet is bound as maintainer. If a different admin verifies later, they replace the old maintainer and any running meter is stopped first.

Rounds

One round at a time. Target is also the cap. Deadline missed means a full refund. Creator allocation is 0–10% of that round's shares, only if it funds, and never dilutes earlier holders.

Jobs from your repo

A repo can describe its GPU job in repogpu.json at its root. RepoGPU schedules the cheapest compatible provider under the cap; if none fits, a fallback GPU is rented so the pool never waits.

Trust model

A contract cannot see a GPU. That is why the rate cap bounds every charge, billing moves one clock forward, and every charge names a receipt hash that can be matched to a job log. Never place private secrets on a community GPU.

Robinhood Chain status

This web app reads Robinhood Chain (4663) live. Factory 0xd73a60b98640877a99b2bc2bc7e551fe5c213250 is deployed, and the ggml-org/llama.cpp pool currently holds 0 USDG, so real charge settlement is blocked until a user funds that pool. The operator is running and will reject charges above the pool balance or rate cap. GitHub maintainer verification still requires the actual repository admin to sign in; a test binding by the verifier wallet is not GitHub-admin proof.

This repo is not on the trail.

The page does not exist. Pools and funds live on-chain, not on this page.

Back home