# Compute

> A CLI tool that provisions fresh cloud GPUs for Python functions, streams output to your terminal, and bills a flat provider rate plus 7.5% platform fee per run.

Compute is a cloud GPU orchestration tool built by Theoric that lets developers point a CLI command at a Python function and have a fresh GPU machine provisioned, run, and terminated automatically. It streams stdout to the terminal in real time and produces a single itemized receipt per run covering provider usage and the platform fee. The service is currently in limited admission, with public self-service starting on RunPod Secure H100; MI300X access is noted as currently limited.

## What It Is

Compute sits in the cloud GPU access category, acting as a unified interface over multiple GPU providers. Instead of managing cloud accounts, instance types, or billing dashboards across providers, a developer installs the Compute CLI, adds prepaid credit, and runs a single command like `compute run simple_mlp.py::train --gpu H100`. Compute handles provisioning, execution, log streaming, and termination — then closes the run with a final receipt. The core value proposition is eliminating the operational overhead of one-off GPU jobs for workloads like fine-tuning, reinforcement learning, and batch inference.

## Provider Coverage

Compute routes jobs to GPU providers through a single interface. As of the current state of the product:

- **RunPod** — Secure H100 capacity, available now
- **Hot Aisle** — MI300X capacity, available now
- **AWS, GCP, Azure, Vast.ai** — listed as coming soon

A 70B parameter model is cited as fitting on a single 192 GB MI300X card, and the MI300X guide covers LoRA fine-tuning at that scale.

## Workflow and CLI Design

The end-to-end run flow is designed to be minimal:

1. Sign up and use an enabled sign-in method
2. Add at least $10 in prepaid credit via Stripe
3. Install the CLI with a single `curl | sh` command
4. Run a computation by passing a Python entry point and GPU type

During a run, Compute locks the provider rate at request time, provisions a fresh machine, streams stdout, captures the result artifact, terminates the machine, and emits a final receipt. A machine that never becomes ready costs $0 — no provider usage or fee is debited on a failed boot.

## Supported Workloads

The documentation organizes use cases into three guide categories:

- **SFT / Fine-tuning** — supervised fine-tuning of open models on task-specific examples
- **Reinforcement learning** — reward-signal training for cases where labeled examples are insufficient; supports detached runs with log following
- **Batch inference** — running a model across large input sets for evals, embeddings, synthetic data, or overnight labeling

## Billing Model

Compute uses a prepaid credit system with no subscriptions or usage tiers. Every run produces two line items: the provider's hourly list rate (metered by cumulative started minute, where 59 seconds rounds to one minute and 61 seconds rounds to two) and a flat 7.5% platform fee on provider usage. The homepage example shows a $1.00 total run broken down as $0.80 provider usage and $0.20 platform fee. Volume pricing is noted as available for significant recurring usage. New runs are blocked when the balance reaches $1 or less; active work stops at $0.50 or less.

## Current Status

Compute is live but in limited admission. The GitHub repository for the documentation (`theoriclabs/docs.compute.cx`) was created in August 2026 and is hosted via Mintlify. The product monorepo is separate and not public. The docs repository explicitly notes that unshipped surfaces — including PyPI packaging, persistent disks, and extra GPU SKUs — should not be documented until live.

## Features
- Provision fresh GPU machines per run
- Stream stdout to terminal in real time
- Automatic machine termination after run
- Single itemized receipt per run
- Prepaid credit billing via Stripe
- Flat 7.5% platform fee on provider usage
- Provider rate locked at request time
- Failed boot costs $0
- Support for H100 and MI300X GPUs
- Fine-tuning, RL, and batch inference guides
- Detached run mode with log following
- Balance controls to prevent overspend

## Integrations
RunPod, Hot Aisle, AWS (coming soon), GCP (coming soon), Azure (coming soon), Vast.ai (coming soon), Stripe, Mintlify

## Platforms
CLI, API, WEB

## Pricing
Paid

## Links
- Website: https://compute.cx
- Documentation: https://docs.compute.cx
- Repository: https://github.com/theoriclabs/docs.compute.cx
- EveryDev.ai: https://www.everydev.ai/tools/compute-cx