Rent a whole GPU

Serverless is pay per token on whichever host is free. Dedicated is different: you reserve one GPU for yourself, by the hour, and run one catalog model on it. Requests to your endpoint go only to that GPU. They never queue behind other renters and they are not billed per token. You pay for the hour whatever you send.

GPU classes and prices

A model fits a class when its weights and a 4,096 token context fit in that much VRAM. The picker on Dedicated GPU only shows models that fit the class you chose.

Start a rental

1. You need at least one hour of credits. Start is disabled until you have them, with a link to Billing.
2. Open Dedicated GPU, choose a class and a model, and press Start rental. One hour is reserved from your credits.
3. Ballast finds an idle host in that class, tells it to load the model and checks the result with a spot check. This shows as Starting. If it takes more than 10 minutes the rental is canceled and the credits are released.
4. When it turns Active, the page shows your endpoint, the GPU class (never who owns it), the time so far and the cost so far.

Call your endpoint

The endpoint is https://ballast.outerscope.xyz/v1/dedicated/<rental_id>. It speaks the same chat API and takes your normal key. Use the model you rented.

curl https://ballast.outerscope.xyz/v1/dedicated/<rental_id>/chat/completions \
  -H "Authorization: Bearer $BALLAST_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3-32b", "messages": [{"role": "user", "content": "Hello"}]}'

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://ballast.outerscope.xyz/v1/dedicated/<rental_id>",
    api_key=os.environ["BALLAST_API_KEY"],
)
reply = client.chat.completions.create(
    model="qwen3-32b",
    messages=[{"role": "user", "content": "Hello"}],
)
print(reply.choices[0].message.content)

How you are billed

• One hour is reserved when you start, and the next hour is reserved at each hour boundary while the rental runs.
• You are charged per minute served, with a minimum of 10 minutes. Press Stop and the unused part of the reservation comes back.
• A minute only counts if the GPU sent a heartbeat that minute. Downtime is never billed.
• If the next hour cannot be reserved, the rental stops at the boundary and you get an email. A rental also stops after 24 hours. To keep going, start another.
• You can have 3 open rentals at a time.

If the host drops

If the GPU misses heartbeats for 60 seconds, the rental shows Interrupted and billing stops at the last healthy minute. Within 5 minutes Ballast tries to pin another idle GPU of the same class, load your model and carry on at the same endpoint. You will see "Moved to another host" and the gap marked as not billed. If no GPU turns up in 5 minutes, the rental ends and the unused reservation is released.

What the endpoint returns

The model is always one from the catalog. Dedicated rentals do not run your own weights, containers or shell. That is how hosts stay safe from renter code.