Compute

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.