Dedicated instances
Enterprise
A machine of your own, under EU jurisdiction, with a written agreement. Dedicated instances are the first product of Sovereign AI Cloud.
What a dedicated instance is
A reserved GPU node
PlannedA 24 GB-class GPU node (RTX PRO 4000 class, 64 GB RAM) reserved for your organisation, expected to run the 9B and the 12B model resident at the same time. Quoted on 2026-10-12; provisioned after the infrastructure decision.
EU jurisdiction
PlannedThe seller is an Estonian company, the law is Estonian, the data centre is in the European Union, and no component of the request path is outside it.
Zero retention of content
BuiltThe gateway's audit line carries metadata only; your prompts and completions are never written anywhere. Built and in use on the inference gateway today.
Read the documentData Processing Agreement
BuiltArticle 28 GDPR, with Annex I (subject matter), Annex II (technical and organisational measures) and Annex III (sub-processors); a 48-hour breach notification so you can meet your own 72-hour duty. A complete draft is published for lawyer review.
Read the documentService Level Agreement
BuiltAvailability target, service credits, exclusions, maintenance windows and measurement through the status page. A complete draft is published; the numbers are fixed with the infrastructure decision.
Read the documentProcurement
PlannedInvoices with reverse-charge VAT for EU businesses, purchase orders, supplier onboarding forms and security questionnaires answered from the published Security and Trust page.
Security and trust, with status
BuiltThe technical and organisational measures are published with a status per measure — what is built, what is planned — so a questionnaire is answered by a page, not by a promise.
Read the documentConcurrency and throughput
PlannedToday's runtime is llama.cpp at one stream; a vLLM or SGLang runtime is quoted at about four times the throughput at concurrency 32 and will be measured before it is offered.
Bring your own open-weight model
PlannedAn open-weight model of your choice, resident on your instance, under its own licence — on request.
Status labels — MEASURED — exists and was measured on the stated date by the stated method. BUILT — exists and works, not yet measured under load. PLANNED — designed, not yet built. A line without a label is not a claim of this platform.
Cloud performance numbers are projected until a cloud node is measured; the measured numbers of the local machine are on the Benchmarks page.
How it goes
Write to us
The use case, the expected volume, the languages, the compliance requirements — to the commercial address on the Contact page.
Scoping and documents
A call to size the instance; the Data Processing Agreement and the Service Level Agreement for your review; the sub-processor list; your security questionnaire.
Order and provisioning
An order form with the fixed tariff; the node is provisioned, your models are loaded, your keys are created in your own organisation.
Acceptance
Your tests against your instance, your audit trail, the status page probing your endpoint. Then the first invoice.