Daytona

NVIDIA H100, H200, RTX PRO 6000, RTX 5090 and RTX 4090, and the AMD Instinct MI355X, each available as a sandbox with one to eight GPUs. The NVIDIA B200 is on the rate card ahead of its launch. Every GPU is allocated exclusively to one sandbox at a time.

On-demand sandboxes keep their GPUs until you stop them. Preemptible sandboxes use spare capacity at a lower per-hour rate and can be reclaimed by Daytona when on-demand demand needs the GPUs, so they suit fault-tolerant work such as batch inference, evaluations and checkpointed training. Both are billed per second. Reserved capacity is available under enterprise agreements on request.

Per second, with no minimum commitment. The GPU rate covers the GPU only; vCPU, memory and disk are configured per sandbox and metered separately per hour. There is no subscription and no minimum spend, and new accounts start with free compute credit.

Daytona has GPU capacity in the United States, Canada, the United Kingdom, the European Union and Norway. GPU sandboxes are scheduled onto an aggregated shared region that spans that capacity, so a request takes the first available GPU of the type you ask for rather than targeting a single site. Dedicated and custom regions are available for organisations that need a fixed location.

From the dashboard, the REST API, the CLI, or the SDKs for Python, TypeScript, Ruby, Go and Java. You pick a Docker or OCI image, the number of GPUs and a GPU type, or an ordered list of acceptable types so the sandbox falls back to an alternative when the first choice is unavailable. A capacity endpoint reports live on-demand and preemptible availability per GPU type before you create one.

That is what the platform is built for. Each sandbox is isolated, starts from an image you control, and can be snapshotted and reused, so agent code execution, evaluations and model serving run without access to your own systems. Daytona is SOC 2 Type II audited, with the report available at trust.daytona.io.