A peer-to-peer marketplace for on-device AI inference. Send a prompt over an encrypted P2P link, run it on a peer's hardware, and pay per token in real testnet USDT. No servers, no cloud, no cluster.
Two everyday devices, talking directly. One hosts a model, the other borrows it.
A laptop loads a model and opens a peer-to-peer endpoint over Holepunch, advertising a per-token price.
A device with no model at all sends a prompt straight to that peer. The answer streams back, token by token.
The work is metered by token and paid on-chain. The seller reads its own balance climb, straight off the chain.
Everything else can be mocked for a demo. These two cannot.
A buyer with no local model gets a real, streamed completion executed on the seller's hardware, routed through an encrypted peer link. It keeps working when the internet is cut.
Each request triggers a real on-chain USDT transfer on Sepolia. From a live run:
0x86ef9b16…eeb9f082 ↗A buyer dashboard and a seller dashboard, each wired to its own peer process.

