Yes, a standard GDPR Article 28 data processing agreement is published at gpuwerk.com/legal/dpa at no charge. Our sub-processor list at gpuwerk.com/legal/sub-processors is short: there are none for instance workloads.

Private LLM hosting in EU-Central

Your models. Your data. One tenant: you.

Q: What happens to our data when we terminate?

Terminating an instance deletes its container and its /workspace volume from the node. Before a node is offered to another customer it is sanitised: every tenant container and volume is removed, and the node fails its readiness check and is not allocated while any remain. That is deletion at the filesystem level. We do not run self-encrypting drives, so we make no cryptographic-erase claim. GPUwerk keeps one recovery copy of a running or stopped workspace on its own server in the Czech Republic, refreshed about every six hours. It exists to recover from a failed drive and to restore a workspace stopped for lack of credit; it is not a backup service, and terminating deletes it. If your credit reaches zero or you hit your monthly budget, the instance is stopped, the machine is released, and the workspace is saved for 7 days before it is deleted. Low-balance mails warn you before that happens; auto top-up makes it less likely, but a payment can fail.

Private LLM hosting on whole dedicated machines in EU-Central. Your data is processed in the EU, never used for training, and no other customer ever touches your GPU.

Try a Spark yourself Get help setting it up

$0.79/hour, billed per minute · pay $10, get $20 in credit on your first top-up

At a glance.

1tenant

One tenant per machine

A whole dedicated machine, assigned to you. No shared GPUs.

Sub-processors for workloads

The sub-processor list is short: none for instance workloads.

Art.28

GDPR data processing agreement

Published at /legal/dpa, no charge.

The three fears

The three fears, answered.

Every company we talk to arrives with the same three worries. Here is exactly how each one is handled.

"Our data leaves the EU"

It doesn't. The hardware is physically in EU-Central and operated by PRINT IT! SE, a Societas Europaea registered in Prague. There is no US parent and no US region behind that machine. The few services GPUwerk uses around it, such as transactional email, are listed with their providers and countries in the sub-processor list. Data residency isn't a checkbox in a settings panel, it's where the machine bolts to the rack.

"The provider trains on our data"

GPUwerk runs infrastructure, not models: your prompts and documents are processed on your dedicated machine, by software you installed, reachable only through your SSH keys and your console's browser terminal. As on any hosted service, GPUwerk retains infrastructure administrator access to the underlying machine, but as a matter of policy set out in our data processing agreement, we do not access your workspace's content except at your request for support or where a legal obligation requires it, so there is nothing for us to train on. It's in the contract either way. The only training that happens is yours, on your own data.

"Employees paste secrets into ChatGPT"

Blocking pushes it to personal phones. The fix is a sanctioned alternative that's just as convenient: a ChatGPT-class assistant on your own instance, same chat UI, same API, every prompt staying on a dedicated machine assigned to you alone. See the private ChatGPT setup →

For your DPO

Written for your DPO to forward.

The compliance picture in one table, plain language, no marketing.

Question	Answer
Where is data physically processed?	EU-Central, on a dedicated single-tenant machine assigned to you
Who is the operator?	PRINT IT! SE, a Societas Europaea registered in Prague, Czech Republic (Company ID 069 60 090). No non-EU parent entity.
Who can access instance content?	Through the instance itself, only holders of your SSH keys, and you through the browser terminal in your GPUwerk console; password login is disabled fleet-wide. GPUwerk keeps infrastructure administrator access to the underlying machine, as on any hosted service, and under the DPA does not use it on your content except at your request for support or where a legal obligation requires it. A model image's HTTPS endpoint is reachable by anyone who has the URL until you put an authenticating proxy such as LiteLLM in front of it.
Does GPUwerk read prompt or model data?	No. GPUwerk hosts the hardware and runs billing and, under the DPA, does not access, read, copy, index or analyse workload content
Is any customer data used for training?	No, contractually, and since we do not access your content there is nothing to train on
Sub-processors for workloads?	None
DPA (GDPR Art. 28)?	Published at /legal/dpa, no charge
Data on termination?	Container and workspace volume deleted from the node, node sanitised before it is offered to anyone else. Filesystem deletion, not a cryptographic erase.

Architecture

The architecture is the argument.

Few arrows, each one documented: the shorter the DPO review.

your office / VPN            your apps
      │                          │
      │  SSH (keys only,         │  HTTPS model endpoint
      │  dedicated IP)           │  (add an auth proxy)
      ▼                          ▼
┌────────────────────────────────────────────┐
│  your dedicated DGX Spark · EU-Central     │
│  model · vector store · chat UI · API      │
│  128 GB unified memory · 1 TB NVMe         │
└────────────────────────────────────────────┘
      │
      │  /workspace recovery copy, ~6-hourly
      ▼
GPUwerk backup server · Czech Republic

      three arrows, all listed in the DPA.

Weighing this against a per-token API instead? Read private LLM hosting vs Azure OpenAI or private LLM hosting vs AWS Bedrock for data location, who can see your prompts, and measured throughput.

Guides

Guides by provider, industry and country.

Every comparison, industry and country guide, in one place. Open the comparison index.

How it compares

A dedicated EU Spark against the per-token APIs buyers cross-shop it with. Data location, who can see your prompts, and where each one wins.Browse 62 guides

Private LLM hosting by industry

The same dedicated-machine setup, argued for the constraints each field actually has.Browse 76 guides

Private LLM hosting by country

The same dedicated-machine setup, argued for the regulatory and language context of each market.Browse 50 guides

Compliance FAQ.

Is it GDPR compliant?

We provide the evidence you need: hardware and operations in EU-Central, an Article 28 data processing agreement, no sub-processors for workloads, and dedicated single-tenant machines whose content GPUwerk does not access, read, copy, index or analyse. Infrastructure administrator access to the underlying machine exists, as on any hosted service, and under the DPA it is not used on your content except at your request for support or where a legal obligation requires it. You remain the controller, so compliance of the processing itself is yours to assert.

Do you sign a DPA?

Yes, a standard GDPR Article 28 DPA is published at /legal/dpa at no charge. Our sub-processor list for workloads is short: there are none.

Can our DPO audit the setup?

Yes. We provide documentation of the physical location, access controls, and wipe procedures, and answer DPO questionnaires directly. Send your DPO's questions.

What happens to our data when we terminate?

Terminating deletes your container and its /workspace volume from the node. Before that node is offered to another customer it is sanitised: every tenant container and volume is removed, and the node fails its readiness check and is not allocated while any of them remain. This is deletion at the filesystem level. We do not run self-encrypting drives or full-disk encryption on the nodes, so we make no cryptographic-erase claim. GPUwerk keeps one recovery copy of a running or stopped workspace on its own server in the Czech Republic, refreshed about every six hours. It exists to recover from a failed drive and to restore a workspace stopped for lack of credit; it is not a backup service, and terminating deletes it. If your credit reaches zero or you hit your monthly budget, the instance is stopped, the machine is released, and the workspace is saved for 7 days before it is deleted. Low-balance mails warn you before that happens; auto top-up makes it less likely, but a payment can fail.

Does this also cover the EU AI Act, not just GDPR?

Dedicated hardware doesn't change your EU AI Act risk tier or provider/deployer status, those depend on what your system does, not on where it runs. What it does give you is a clean answer to who can access your data while the model runs, which is useful evidence for the deployer-side obligations most self-hosting companies actually face. See GDPR and the EU AI Act on dedicated hardware for the combined picture in one page.

Beyond infrastructure

Practical rollout package.

If you need more than infrastructure, we can add a scoped engagement with clear deliverables for team onboarding and integration.

Project intake
Define use case boundaries, latency goals, and who owns production approval.
Team training
Practical sessions on prompt quality, governance patterns, and common operations.
Integration roadmap
Plan the API rollout, including what your tools should do when the instance is stopped or busy.

See how the engagement works

Bring your AI project to life.

Deploy a Spark and test it yourself, or ask the person who runs the racks. First top-up: pay $10, get $20 in credit.

Deploy now Ask a question