Thunder Compute provides CPU and GPU sandboxes for AI agents and developers, as well as persistent GPU instances. Its core positioning is to lower the GPU cost of AI/ML workloads.
In terms of use cases, Thunder Compute mainly serves workloads such as AI agents, reinforcement learning (RL) environments, evaluations (evals), and post-training.
Key capabilities include: offering A100 and H100 GPUs; supporting per-second and per-minute billing; and providing two usage forms—GPU instances and GPU sandboxes. GPU instances are instances that developers can connect to and retain, and can generate snapshots when finished. GPU sandboxes are used for RL environments and evaluations, can optionally use a GPU, and can scale in parallel from a few lines of Python code to hundreds of rollouts. Both forms run on the same GPUs.
In terms of usage, the official documentation states that they can be launched from a terminal or an agent, and GPU instances can be connected to from VS Code, the CLI, or a browser. Switching platforms is officially said to take only a few minutes.
Applicable scenarios include: developers who need low-cost GPUs for manual development, ML engineers and researchers who need large-scale parallel RL environments and evaluations, and teams building evaluation and training environments for AI agents.
In terms of limitations, the materials do not provide details such as specific available regions, storage and network configuration, SLA, supported frameworks and images, free credits, or trial policies; this information is not currently provided.
The official source also gives a set of price comparison data (the unit is not specified in the materials): Thunder Compute is $3.20, Paperspace is $5.95, CoreWeave is $6.16, Amazon Web Services is $6.88, Oracle Cloud is $10.00, and Google Cloud is $10.98. The official source claims that after switching, users can save up to 68% compared with AWS. In addition, the official blog mentions that Thunder Compute completed a $13 million Series A funding round on August 19, 2026, and explains how GPU-over-TCP works and the methods and trade-offs of GPU virtualization.
