GPU docs.

Rent real GPUs by the hour and pay from your Solana wallet. This page covers how renting works, what you get, how payment and refunds work, and the API for agents.

Overview

GPU sells hourly access to real machines. Most of the catalogue comes from RunPod Secure Cloud (data-center GPUs, from RTX cards to H100, H200 and B200). Independent providers can also list their own hardware.

Live stock. The catalogue shows every GPU with its current price. A machine marked Out of stock cannot be rented until capacity frees up.
Pay per hour in SOL or USDC, from Phantom. No account, no card: your wallet is your login.
Starts on payment. Cloud GPUs boot right after the payment is confirmed on Solana, with Jupyter and SSH access.
Stops by itself when your paid hours are over. Nothing keeps running or billing.

Rent a GPU

Open the app and sign in with Phantom (one signature, no transaction, no fee).
Pick a machine, the number of GPUs and the number of hours (1 to 144 for cloud GPUs). The total is shown before you pay.
Optional: paste your SSH public key (ssh-ed25519 …) to get terminal access.
Choose SOL or USDC, click Pay now and approve the transaction in Phantom.
Once the payment is confirmed (a few seconds), the machine starts. Its access details appear under Paid rentals. Jupyter is usually ready within 1 to 3 minutes.

Not sure what to pick? Describe your workload to the GPU Advisor on the home page.

Pricing

Prices are per GPU per hour, live from the catalogue. SOL prices follow the SOL/USD rate at the moment you pay. Network fees on Solana are a fraction of a cent.

GPU	Memory	Stock	USDC / h	SOL / h
Loading live prices…

Paying in USDC the first time to a new wallet may add about 0.002 SOL to create the receiving token account. This is a Solana network cost, paid once.

Payment

You pay with a single Solana transaction that you sign in Phantom. The server builds it, Phantom shows you exactly what leaves your wallet, and you approve it.

Before relaying, the server checks that the signed transaction pays exactly the agreed amount to the right wallet. Anything else is refused.
The rental only counts as paid once the transaction is confirmed on-chain. Every payment links to Solana Explorer.
For a cloud GPU, the capacity is checked before you pay, so you are not charged for a machine that cannot start.
Listings from independent providers send the payment to the provider's wallet; a 2.5 % platform fee is included in the price shown.

Access your machine

Jupyter

Click Open Jupyter on the rental and log in with the password shown next to it. You get JupyterLab with a terminal, Python and PyTorch with CUDA ready to use.

SSH

If you pasted an SSH public key when renting, click SSH on the rental to get the command, for example:

ssh root@203.0.113.10 -p 10341

The SSH port opens a minute or so after Jupyter. If it says "not open yet", try again shortly.

What is on the machine

Image: runpod/pytorch (Ubuntu 22.04, Python 3.11, PyTorch, CUDA 12.4).
40 GB system disk and a 20 GB volume mounted at /workspace.
You are root: install anything you need with pip or apt.

When time is up

The machine is stopped and deleted automatically at the end of the hours you paid for, with everything on its disks.

Save your work before the end. Download results from Jupyter, push to Git, or copy to your own storage. Data is not kept after a rental ends.

Need more time? Rent again before the end and move your files, or pick enough hours up front.

Problems & refunds

The machine could not start after you paid: the rental shows Could not start · refund due and you are refunded to the wallet you paid from.
Something is wrong while it runs (unreachable, broken GPU): click Report a problem on the rental. Each report is reviewed and settled, including a refund when the machine was at fault.

Payments go straight from your wallet and there is no on-chain escrow: refunds are sent back manually after review.

List your machines

Have GPUs or servers? In the app, open Provide compute, pick the hardware you run, set your price per GPU-hour in USDC and/or SOL, and publish. Renters pay your wallet directly.

When a rental is paid it appears under Rentals on your machines.
Give the renter access, then click Access handed over, start the clock. The paid hours count from that moment.
Pause a listing at any time; it disappears from the catalogue.

Only list machines you actually run. Renters can report problems, and repeated reports lead to removal.

GPU Advisor

On the home page, describe what you want to run ("Llama 70B in 4-bit for an agent", "fine-tune SDXL for 3 hours"). The Advisor sizes memory and compute, picks the cheapest machines in stock that fit, and estimates the total. Click Rent on a pick to go straight to checkout.

FAQ

Do I need an account?

No. Sign in with Phantom. Your rentals are tied to your wallet address.

Which wallets work?

Phantom on desktop and mobile. You need a little SOL for network fees, plus the SOL or USDC for the rental.

Can I stop early and get the rest back?

Not automatically: cloud GPUs are reserved for the hours you buy. Rent fewer hours if you are unsure, you can always rent again.

Can I rent several GPUs in one machine?

Yes when the GPU type offers it: the app shows the GPU counts available for each machine.

Is my data private?

The machine is yours alone for the rental and is wiped when it ends. Do not leave secrets on it after you are done.

API

Agents can use the platform programmatically. Base URL . Authenticate with Authorization: Bearer acp_live_… (create keys in Console → Settings). Errors are {"error":{"code","message"}}.

Live stock and prices

Start a workflow run

OpenAI-compatible inference

SDKs: acp.ts (zero dependencies) · acp.py (standard library). Full machine-readable spec: openapi.json. Webhooks are signed with ACP-Signature: t=…,v1=hex(hmac_sha256(secret, "t.body")).

Reference

Loading…