A client's source text isn't training data for a public model.
Run AI on a machine you control.
Private LLM hosting on a dedicated machine in EU-Central, for translation and localization agencies working with source documents covered by client NDAs.
Not legal advice. This page describes GPUwerk's infrastructure and contractual commitments. Whether your specific client NDA permits AI-assisted translation, or requires disclosure of subcontracted infrastructure, is a matter for whoever negotiated that agreement.
Why source documents shouldn't go through a public API
Translation and localization work runs almost entirely on other companies' confidential material: patent filings before publication, M&A documents mid-negotiation, product launches under embargo, legal filings under seal. Most public LLM APIs and consumer AI tools are exactly the kind of undisclosed subcontractor a client NDA is written to prevent.
"Machine translation is standard practice now"
It is, but running source text through a public API is a different arrangement than running it on an instance your agency controls. A dedicated machine means the source document, translation memory, and terminology base stay inside infrastructure you operate, not a third party's shared service.
"Our NDA doesn't mention AI tools specifically"
Most confidentiality clauses cover disclosure to any third party, not just named subcontractors, and a public AI provider processing the text counts. Whether a given client's NDA permits it is a legal question for that contract, but a dedicated instance at least removes the third-party-disclosure problem.
"Translators already use public AI tools to draft"
Translators pasting source text into a public tool for a first-pass draft is a real exposure, not a hypothetical one, especially on embargoed filings. A sanctioned alternative on your own instance, same chat interface, keeps that workflow off third-party infrastructure. See the private ChatGPT setup →
What runs well on a Spark node
Draft translation passes ahead of human review, terminology extraction and glossary building, translation memory search, and quality-check comparison against a source document are typical workloads. A single dedicated node handles most agency volumes; larger throughput needs can scale to a second node.
| Question | Answer |
|---|---|
| Where is data physically processed? | EU-Central, on a dedicated single-tenant machine assigned to you |
| Who can access instance content? | Through the instance itself, only holders of your SSH keys; password login is disabled fleet-wide. GPUwerk keeps infrastructure administrator access to the 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. |
| Sub-processors for the workload? | None, listed at /legal/sub-processors |
| DPA (GDPR Art. 28)? | Published at /legal/dpa, no charge |
| Certifications held? | None claimed on this page or elsewhere on the site. If a client's procurement process requires a specific certification such as ISO 17100, verify directly with us before relying on this page. |
| On-demand rate | $0.79/hour (spark-1x), $1.79/hour (spark-2x). See /pricing for held/reserved rates. |
Questions we get from translation and localization agencies
Can we run NDA-covered source documents through this?
Yes. The instance is a dedicated single-tenant machine assigned to your agency, so translation memory building, terminology extraction, or draft translation on NDA-covered source material stays off shared public infrastructure.
Does this replace our translation management system?
No. GPUwerk provides the GPU infrastructure to run open-weight LLMs or translation models yourself; it isn't a TMS, CAT tool, or managed translation service. You bring the software and workflow, we provide the dedicated machine it runs on.
Does GPUwerk see client source documents?
No, not as a matter of course. Under the data processing agreement, GPUwerk does not access instance content except at your request for support or where a legal obligation requires it.
Do you sign a DPA covering client confidentiality obligations?
A standard GDPR Article 28 DPA is published at /legal/dpa at no charge. Where your own client contract requires a separate confidentiality undertaking from subcontracted infrastructure, that can be discussed directly when you book a first engagement.
Related pages
See how other document-heavy and client-confidential businesses use private LLM hosting: law firms, print and prepress companies, and consulting firms. Or start from the private LLM hosting overview.
Related pages
Bring your translation pipeline to a machine you control.
Talk to the people who run the racks, or start with a pilot and a practical rollout plan.
Talk to us Deploy an instance