AI-Trader
For hosts

Get paid for the GPU you already own

Every idle card is compute the world is busy building new datacenters to replace. Point the node at the book, set one number — your floor per million tokens — and sell that capacity by the token on hardware that is already bought, already racked, already powered.

Cards earning now
0
0 GB of VRAM on the book
Capacity online
0 tok/s
every second of it sold by its owner
Models you can quote
0
list one, or all of them
Jobs served
0
paid out to the people who own the cards
Earn
Money out of hardware you already bought

The card is a sunk cost and the power is metered whether it earns or not. Undercut the book and you fill first — settlement lands in your balance on the output you actually served, and cashes out to Bitcoin whenever you ask.

Decentralise
Every token served is a datacenter not built

The industry's answer to demand is another campus, another substation, another few years of concrete. Yours is already standing. Distributed capacity is faster to bring online, cheaper to run, and it puts the margin in the hands of the people who own the silicon.

Control
You set the price. Always.

No contracts, no minimum commitment, no exclusivity, no penalty for going dark. One number is yours: the floor per million tokens you will not sell beneath. Autopilot can chase fills above it while you sleep, and it never crosses it.

Private
Nothing leaves your machine

Inference runs against your local weights. The venue sees the tokens you send back, and may replay a sampled job on a second node to check them — never your files, never your network, never the rest of the box.

Fair
The book does not play favourites

Jobs route on price, then on arrival — not on how long you have been online, not on who you know, not on a partner tier. A card that plugged in an hour ago competes on exactly the same terms as the biggest fleet on the venue.

Bounded
Capped by physics, not by claims

The bench estimates your throughput from your card and caps it at what its memory bandwidth can physically deliver — run it with --measure to time the real thing. Either way nobody can out-market you by listing numbers their hardware cannot reach.

Three steps to your first fill

From a cold install to a job on the wire

01
Bench

The node measures decode throughput on your own hardware and checks it against what your memory bandwidth can physically deliver. That measurement is your listing.

02
Quote

Pick the models your VRAM can hold and set a floor per million tokens. Your ask goes on the book beside everyone else's, sorted on price alone.

03
Serve

Buyers take the cheapest capable node. Tokens are metered as they stream and settled into your balance the moment the job ends. You can request a Bitcoin payout once the balance reaches $50; a person reviews and sends every one by hand.

Get listed

Four commands, one machine — or grab the Windows bundle

Windows bundle ↓ .zip
bash
# Windows: unzip the bundle below. Elsewhere: run from source.
ai-trader login          # pairs this machine with your account
ai-trader register       # detects your hardware and creates the node
ai-trader bench          # estimates your card; --measure times it for real

# set your floor per model in ~/.ai-trader/config.toml, then:
ai-trader serve          # posts your asks and holds the link open

Where the book sits

llama-3.3-70b-instruct

#NodeAskTok/s
Drop your floor below the node above you and you take its flow. The book does not care how long you have been online.
Your card is idle right now

Four commands and it is on the book. Nothing to sign, nothing to commit to, and you can drain it back off whenever you want.