Cloud providers
Choose a GPU provider and storage, connect your accounts, and understand what training and serving require.
Train in your own RunPod, Verda, Lambda, AWS or Azure account. Keep your data in a private S3-compatible bucket or a RunPod network volume. Your cloud providers bill compute, storage and transfers separately from your Tensorant plan.
Choose your setup
| Provider | What it supports | Setup guide |
|---|---|---|
| RunPod | Training, managed evaluation, Playground deployments and published models, including Serverless | Connect RunPod |
| Verda | Training with a project client ID and secret, location and registered SSH key | Connect Verda |
| Lambda | Training with an API key, region, registered SSH key and compatible GPU image | Connect Lambda |
| AWS | EC2 training with a CloudShell setup command and optional custom network, image and recovery access | Connect AWS |
| Azure | VM training with a Cloud Shell setup command and optional custom network, image and recovery access | Connect Azure |
| S3-compatible storage | Data and adapter storage independent of the GPU provider, including AWS S3 and Cloudflare R2 | Connect storage |
RunPod is the simplest way to follow the complete quickstart. For organizations using their own connections, managed serving and the managed evaluation stages of experiments require both RunPod compute and a RunPod network volume. Custom S3 supports data preparation and training without a RunPod account.
Connect and verify
- An organization owner opens Settings → Connections and saves the provider credentials and storage connection. Members can check connections; viewers can read their status.
- Connect storage before creating a project. Custom S3 can be connected before compute; a RunPod volume needs a RunPod compute connection first.
- Prepare a dataset version and launch a short training trial with a price and time limit. Check that the adapter is verified and the provider confirms GPU release.
Saving credentials does not launch a GPU, reserve capacity or prove that an image can boot. Quota, GPU availability, outbound access and image requirements are covered in each provider guide. Creating a bucket or volume can create storage charges before your first run.
Before you change a connection
Once an organization has projects, its storage location is fixed in Connections. You can rotate keys for the same location; changing providers does not migrate data. Keep credentials valid until active work and recovery resources are released.
A failed run may retain billable storage, and Lambda retains a billable instance because it cannot suspend it. Follow your provider's recovery steps before discarding files. Shared RunPod storage provided by Tensorant can be used only by RunPod training GPUs.
See Connections for model endpoints, Hugging Face tokens, shared connections and credential management, and Training runs for launch checks, costs and recovery.
Need a hand? Visit troubleshooting