Rent the cheapest GPU in stock from your terminal.

01Install and sign in

gpyou is on PyPI. Install it with uv or pipx on Python 3.12 or newer. You need a gpyou account to use it: sign up from the terminal with your email and a verification code, then add a passkey or an authenticator app if you like. The session lives in your OS keychain.

$ uv tool install gpyou   # or: pipx install gpyou
$ gpyou signup          # email code, then passkey, app or skip
$ gpyou credits add 50  # opens Stripe Checkout
$ gpyou status

Update with uv tool upgrade gpyou or pipx upgrade gpyou. Headless box? Add --no-browser to credits add and open the Checkout URL from the JSON on another device.

02Quickstart

gpyou gpus lists the cheapest in-stock offer of every GPU model with at least 24 GB, across all providers. Pick one and the CLI asks what to run on it, then waits with live state until it is running.

$ gpyou gpus
  GPU          VRAM    PROVIDER   WHERE   $/HR
› RTX 4090      24 GB  cheapest   ··      ····
  L40S          48 GB  cheapest   ··      ····
  A100 80GB     80 GB  cheapest   ··      ····
  H100 80GB     80 GB  cheapest   ··      ····
  H200         141 GB  cheapest   ··      ····

  What should run on it?  container · fine-tune · endpoint

03Providers

Three providers are live, all on demand. gpyou ranks offers by confirmed stock, then past launch success, then price. If a provider refuses before renting, it moves to the next offer. More providers are coming.

• Live container Runpod

  Secure Cloud. Data centres worldwide.

  --provider runpod
• Live container Vast.ai

  Verified hosts only. Widest range.

  --provider vast
• Live vm · EU Verda

  Finland and Iceland. Stays in the EU.

  --provider verda
• Coming soon More providers

  Each one passes the same lifecycle, isolation and billing tests before it gets traffic.

04Run a container

Quote first, then launch at the rate you saw. The quote and request ID go to stderr, one JSON operation goes to stdout. Retry with the same --quote-id and --request-id. Never mint a new request ID when the outcome is unclear.

$ gpyou offers ghcr.io/you/train:1 2 --gpu 'RTX 4090'
$ gpyou launch ghcr.io/you/train:1 2 --gpu 'RTX 4090' \
    --count 1 --max-duration 3600 --yes
$ gpyou logs OPERATION_ID
$ gpyou terminate OPERATION_ID

07Storage and registry

Buckets and images live in Cloudflare R2, EU jurisdiction, and survive GPU termination. Uploads resume with --upload-id. gpyou registry login installs a Docker credential helper for your private namespace. Transfers use short-lived, scoped credentials.

08For coding agents

Pass --json to any command for machine output. Prices in the CLI are USD; the API uses integer micro-USD. Without a terminal, interactive pickers return JSON instead of prompting.

$ gpyou --json instances get OPERATION_ID
{
  "id": "OPERATION_ID",
  "state": "running",
  "provider": "verda",
  "gpu": "RTX 4090",
  "max_hourly_micro_usd": 2000000
}

09Python SDK

The SDK reuses the CLI's keychain session. chat() and stream_chat() wait through cold starts and never retry once inference begins. The gateway speaks an OpenAI-style chat schema.

from gpyou.client import Client

client = Client()
endpoint = client.request("GET", "/v1/endpoints")[0]
result = client.chat(endpoint["id"], [
    {"role": "user", "content": "Explain gradient accumulation."},
])
print(result["choices"][0]["message"]["content"])

10CLI reference

Every command has --help with its arguments.

11Credit and billing