Thunder Compute

Thunder Compute provides cloud GPU instances and sandboxes for AI workloads. It addresses the cost and access constraints of running GPU-intensive work by offering persistent machines for hands-on development and isolated environments for reinforcement learning, evaluations, and agent tasks. The platform virtualizes GPUs so they can be scheduled independently of the physical hardware beneath them.

Instances are persistent machines available in RTX A6000, L40, A100 80GB, and H100 SXM configurations, connectable via SSH, the CLI, a VS Code extension, or a browser. Users can change GPU specs, take disk snapshots for later restoration, and launch pre-built templates such as Comfy UI, PyTorch, Ollama, and CUDA Kernels. An MCP server allows agents like Claude Code, Cursor, or Codex to create instances, run commands, and take snapshots. Sandboxes are isolated machines for agents and automated jobs, with or without a GPU, that scale up and down to zero and can fan out parallel rollouts from Python. A REST API is also available.

The service is aimed at developers, ML engineers, and teams running AI workloads. Instances are billed per minute while running, sandboxes per second, and snapshot storage per gigabyte per month. Reserved capacity is available for teams needing dedicated GPU clusters over a fixed term, with contract and volume pricing.