Switching provider
You can leave GPUwerk at any time and take your data with you, free of charge. The contract terms are in section 19 of our Terms of Service. This page describes how it works in practice.
How to switch
- Give a switching notice in the console under Your data, or with
POST /v1/account/switching. Say whether you are moving to another provider, to your own infrastructure, or want your data erased, and choose a transition period of up to 30 days. - During the transition period everything works as before. Copy your workspaces and download your account export.
- The retrieval period follows: at least 30 days in which your instances are kept, are not deleted or released for lack of credit, and carry no hold charge. A Partner-tier machine that is withdrawn or offline keeps your workspace until the period ends; one suspended or removed for tampering or a security problem may be erased at once. You cannot deploy new instances, but you can start existing ones to copy data off.
- When the retrieval period ends we delete every instance and workspace, close the account and confirm by email.
You can cancel the notice before the transition period ends. If you need longer, ask [email protected] to extend the transition period once.
If your account also holds a Partner account, a machine GPUwerk hosts or has not yet shipped to you, or a hardware order still being paid, closing it at the end of a switching would cut you off from them. For those accounts the console does not take the notice: write to [email protected] and we set the dates with you by hand, on the same terms.
Exportable data and formats
| Data | How to get it | Format |
|---|---|---|
Instance workspace (/workspace): your files, code, models, datasets, outputs | Copy over SSH, for example rsync -a or scp, from a running instance | Your own files, byte for byte; any format you used |
| Software you installed inside the container | Reinstall from your own setup scripts, or save the packages you need into /workspace first | Ubuntu 24.04 (arm64 on DGX Spark, x86_64 on RTX 5090) packages, Python environments, container images you built |
| Account, billing ledger, payments, instances and their event log, SSH public keys, hardware orders, Partner record, waitlist, contact and reservation requests, sanctions confirmations, switching requests, messages we sent you | GET /v1/account/export or the console | JSON (UTF-8). The file includes an exportable_data_and_assets list naming each category and its format |
| Invoices | Console, billing section, or GET /v1/invoices | PDF and UBL 2.1 XML |
| Instance configuration: machine type, image, model preset, ports, the country of the machine it runs on | Included in the JSON export | JSON |
The interfaces you use are standard: SSH (OpenSSH), the OpenAI-compatible HTTP API that our vLLM and llama.cpp presets serve, and the Linux filesystem. Nothing in your workspace is in a GPUwerk-specific format.
Known limits
- We keep no backup copy of workspaces. Data exists only on the instance's own machine, so copy it before the retrieval period ends.
- The operating system, the NVIDIA drivers and CUDA libraries, our images and staged model weights are third-party software under their own licences. They are not your exportable data. Your new provider or your own machine will supply its own copies.
- API key and token values cannot be exported, because we store only their hashes. Create new keys with your new provider.
- Moving a workload from DGX Spark (arm64, GB10) to x86_64 hardware may need rebuilt binaries and containers. That is a property of the hardware, not a restriction we impose.
No switching charges
We charge nothing for switching, exporting or transferring data, and nothing for bandwidth. Normal hourly rates apply while you run an instance, including while you use it to copy data off.
Where the infrastructure is
- GPU machines of GPUwerk's own tiers, including machines GPUwerk hosts for their owners: GPUwerk's datacenter in Hradec Králové, Czech Republic.
- Partner-tier machines ("DGX Spark 1× (Partner)"), used only if you choose that tier: at the partner's own premises, in the country the console shows for the machine. By default only machines in the European Economic Area are used; a machine outside it is used only if you choose "any location" (Terms, section 18).
- Control plane, console and account database: a Hetzner Cloud server in Helsinki, Finland.
GPUwerk is established in the Czech Republic and operates the service from the EU. We disclose data only on a binding order of a competent authority under EU or Member State law. If a third-country authority asks us for non-personal data in your account, we refuse unless an international agreement or the conditions of Article 32 of the Data Act are met, we give the minimum required, and we tell you before we disclose unless the request is for law-enforcement purposes and telling you would defeat it. Personal data is governed by Chapter V GDPR. On the Partner tier each instance's workspace is its own encrypted volume, whose key GPUwerk holds, not the partner (Privacy Policy, section 5a).