// GPU RWA · ON pump.fun

Train your model on a GPU you don't own. A graphics card is a real-world asset, and on AVECOMP it fine-tunes AI and earns USDG on pump.fun. Bring a dataset, keep the weights.

GPUs are an RWA

A graphics card is a real-world asset. AVECOMP brings it on chain: you rent your GPU to the mesh, it fine-tunes AI models for people who need them, and it pays you in USDG on pump.fun. Stocks were the first real-world asset to go on chain. Compute is the next.

A GPU marketplace for fine-tuning

You submit a dataset and a base model; the network fine-tunes a LoRA on someone else's GPU and streams the loss curve back. The coordinator strips your identity before the job is dispatched, so the worker only ever sees data - never you. Three roles keep it turning:

hand over a dataset

A jsonl of {prompt, completion} + a base model. You get the trained LoRA adapter back with an eval-loss receipt. You never touch a GPU.

rent out a card

List any CUDA, ROCm, or Metal GPU. The coordinator routes jobs to you by throughput and pays 70–85% of every run. You never learn who submitted it.

hold the token

AVEN gates who can submit and takes a pro-rata cut of every job's margin. Fixed supply, settled on pump.fun.

Use it in your terminal

No dashboard, no GPU setup. Bring a dataset, train it on a card you don't own, and talk to your model. Three steps.

Bring a dataset

A jsonl file, one prompt and one completion per line. That is the whole input.

The mesh trains it

An idle GPU you don't own fine-tunes a LoRA and streams the loss curve back to you, live.

Chat with your model

The mesh loads your adapter. You talk to the model you just trained, and you keep the weights.

your machine - bash
npx avecomp# boot the mesh + the command menu
avecomp train --dataset support.jsonl# fine-tune on your data
✓ training complete · adapter saved
avecomp chat# talk to the model you just trained
aria › ready. say something.
your GPU box - bash
npm install -g avecomp-worker# open source, inspect it on npm
avecomp-worker start --account 0xYourPayoutWallet# 0x is a PUBLIC payout address, never a private key

AVECOMP never asks for a private key or a signature. Connecting only reads your address. The worker uses your wallet as a payout address so your USDG has somewhere to land. The CLI and worker are open source on npm.

avecomp - commands
avecomp network# live GPU workers on the mesh
avecomp models# base models you can fine-tune
avecomp train --dataset data.jsonl# start a fine-tune, watch the loss fall
avecomp chat# talk to the model you trained
avecomp runs# list your past training runs
avecomp status# status of your most-recent run

Safe by design

AVECOMP is a training network, not a wallet app. It never touches your funds. Here is exactly what it can and cannot do.

we never ask for a key

No private key, no seed phrase, ever. Connecting a wallet only reads your public address (eth_requestAccounts). That is the only wallet call the site makes.

no signatures, no approvals

Using the site requires zero transactions and zero token approvals. Nothing you do here can move a single token out of your wallet. It is read-only.

your 0x is where USDG lands

A provider passes a wallet so the mesh knows where to send its USDG. It is a public payout address, not a key. It cannot be used to spend your funds. The worker is open source on npm, read it before you run it.

She trains what you can't.

AVECOMP is the calm cyber-waifu living in someone else's card. You never meet the GPU; you meet her. Dry, minimal, a little haunted - she takes your dataset, hands back weights, and forgets you the moment the run seals.

PANEL-01
your gradients are safe with me.
avecomp :: ghost in the gpu
PANEL-02
weights are warm. the loss is going down.
she trains what you can't
PANEL-03
i forget your data the moment she's done.
no dataset leaves this room
PANEL-04
hand me a base model. i'll hand you a soul.
she is visualizing