Skip to content
Get an API key

Idle GPUs in. Tokens out.

You call one OpenAI-style endpoint. Ballast sends the request to a healthy GPU in somebody’s house that already has the model loaded, and pays that person in dollars. If their machine drops, yours does not.

Get an API keyHost a GPU

Sample network: the points are not real hosts.

curl https://ballast.outerscope.xyz/v1/chat/completions \
  -H "Authorization: Bearer $BALLAST_API_KEY" \
  -d '{"model":"qwen3-14b","messages":[{"role":"user","content":"Hi"}]}'

The film, 45 s

  • Idle GPUs in. Tokens out.
  • If one fails, the next takes it.
  • Send a message.
  • Saved to recent calls.
  • Hosts earn per token served.

Real screens. The network, devices and payouts shown are sample data.

Sample network. These figures are not real traffic.

Three models. Prices you can read.

Pay by the million tokens. You are charged for what was delivered, and a call that fails on one host is not billed twice.

  • Llama 3.1 8B Instruct8 GB min
    $0.05 in / $0.10 out per M
    183
  • Qwen3 14B16 GB min
    $0.12 in / $0.30 out per M
    99
  • Qwen3 32B24 GB min
    $0.25 in / $0.55 out per M
    83

Sample data: host counts come from a sample network. Prices are the real price list.

Hosts online, last 24 h

24 h ago12 hNow

Sample data: hosts online on a sample network, one point an hour.

Rent it by the token or by the hour.

Start serverless with no commitment. Take a whole GPU when you want steady capacity.

Serverless, by the token

Pay by the million tokens. The router picks a healthy host for every call, you hold no capacity, and nothing is charged while you are idle.

  • Llama 3.1 8B Instruct$0.05 in / $0.10 out per M
  • Qwen3 14B$0.12 in / $0.30 out per M
  • Qwen3 32B$0.25 in / $0.55 out per M
Get an API key

Dedicated GPU, by the hour

Reserve one host GPU by the hour and run one catalog model behind your own endpoint. Billed per minute served, with a 10 minute minimum.

  • 8 to 15 GB$0.20 per GPU-hour
  • 16 to 23 GB$0.28 per GPU-hour
  • 24 to 31 GB, RTX 3090, 4090$0.35 per GPU-hour
  • 32 GB and up, RTX 5090$0.55 per GPU-hour
Rent a dedicated GPU

What a GPU keeps after power.

Pick your card, put in your wattage and your power price. Straight talk: serverless alone clears power by a few cents an hour. Renting the whole GPU by the hour is where it adds up.

Idle counts are from a sample network.

Left after power, per rented hourEstimate, not a promise

$0.20

After the 20% take and power.

Draw while rented450 W

Power price$0.17 per kWh

25% rented, 6 h a day$1.22
50% rented, 12 h a day$2.44
75% rented, 18 h a day$3.66

Rental rates are from the price list. Power is counted only for the hours the card is rented.

One request, start to finish.

  1. 01

    Your call comes in

    The same request an OpenAI SDK sends. You only change the base URL.

    POST /v1/chat/completions
model  qwen3-32b
in     3,771 tokens

  2. 02

    We pick a host

    Only healthy ones with the model already loaded. If one fails its check, the next takes it.

    ny-4090-27  score 0.91
model already loaded

  3. 03

    Tokens stream back

    Streaming works as you expect. You never see which machine answered.

    first token  667 ms
speed        38 tok/s
out          310 tokens

  4. 04

    The host gets paid

    Per token served, or per hour if you reserve their whole GPU.

    you pay   $0.001113
host gets $0.000891
estimate, not a promise

Plug in a key. Or plug in a GPU.

Renters: one OpenAI-style endpoint, serverless from $0.10 per million tokens out, or a whole GPU from $0.20 an hour.

Get an API key

Hosts: you keep 80% of what renters pay, and the page above shows what is left after power. An estimate, not a promise.

Host a GPU