Compute

The simplest way for humans
and agents to use compute.

Across providers. Across SKUs. Across workloads.

Quote a GPU. Run a function. Get a receipt. CLI or agent.

Start a GPU runConnect your agent

Example

simple_mlp.py · H100example

$ compute run simple_mlp.py::train --gpu H100

  1. estimaterate locked · confirmation required
  2. provisionfresh H100-SXM requested
  3. ready80 GB HBM3 available
  4. stdoutstep 1/3 · loss 0.8421
  5. result{"status":"ok","artifact":"result.json"}
  6. terminatedmachine gone · receipt final
Provider usage$0.80
Platform fee$0.06
Total debit$0.86
Providers
RunpodHot AisleVast.aiAWSsoonGoogle Cloudsoon
SKUs
H100-SXM · 80 GBMI300X · 192 GBA100-80GBA100-40GBRTX 5090RTX 4090RTX 6000 AdaL4RTX 3090
Workloads
Fine-tune a 70B on your ticketsTrain a reward model from ranked answersScore 10,000 evals overnightEmbed a corpusDistill a teacher into an 8BGenerate preference pairsSweep LoRA ranks on one H100Label a dataset with a 70BTrain and keep the receiptLet an agent launch a quoted job
Providers

Three clouds live.

Grey is not live. All providers

  • RunpodH100-SXM
    Live
  • Hot AisleMI300X · 192 GB
    Live
  • Vast.aiSpot
    Live · spot
  • AWSTrainium
    Soon
  • Google CloudTPU
    Soon
Quickstart

First run.

  1. 01

    Sign up

  2. 02

    Buy credits

    $10 minimum.

  3. 03

    Install the CLI

    macOS or Linux.

    curl -fsSL https://compute.cx/install.sh | sh
  4. 04

    Run a function

    compute run simple_mlp.py::train --gpu H100
Guides

Guides.

SFT

Fine-tune a 70B on your data

What
Train an open model on your data.
When
When the base model misses your format or domain.
Run
compute run mi300x_lora_finetune.py::finetune --gpu MI300X
70B fits on one MI300X.
Guide
Preference

Train from a reward signal

What
Improve a model you can score.
When
Ranking, reward models, reasoning policies.
Run
compute run rl.py::train --gpu MI300X
Stay on --wait or read logs later.
Guide
Batch

Score a dataset overnight

What
Run a model over a file.
When
Evals, embeddings, labels, synthetic data.
Run
compute run batch_infer.py::generate --gpu MI300X
Billed while the machine exists.
Guide
Pricing

Pass-through. Plus a simple fee.

You pay the provider's locked rate. We add a simple fee. Failed boot is $0.

See pricing
Start

Start a run. Or connect an agent.

Start a GPU runConnect your agent