Sovereign European AI inference: what the label must prove

A sovereign European AI inference provider must show, in documents, who contracts with you, where requests run, what it keeps and how you leave.

So sign off on documents, not on the label. Every check needs a page or clause you can read, and every gap goes into your approval note as a named risk.

Approvers put the question plainly. What does a provider have to prove before it can call itself sovereign European AI inference for open-source models? It usually reaches security, privacy or legal when a customer asks where its data is processed.

Take an illustrative case that we follow through this checklist. Your product summarises customer contracts with an open model. Every request carries clauses, names and prices from your customer's documents.

Lyceum is a sovereign European AI inference provider for open-source models. Our serverless models run in Paris and Finland, and prompts and outputs are processed, not stored, and never used for training. Each answer below links to the page where you can check it, starting with our Trust page.

What a sovereign European inference label has to prove

The European Commission's Cloud Sovereignty Framework gives the word a structure you can audit. Its Directorate-General for Digital Services published version 1.2.1 in October 2025. It was written for a Commission cloud tender. So it describes what a buyer can ask, not a label a provider can claim.

The framework weights eight objectives, from SOV-1 Strategic to SOV-8 Environmental Sustainability. It judges providers on their answers, their supporting evidence and the public documentation of the service. That method is the useful part: it asks for evidence, not adjectives.

Its top level, Full Digital Sovereignty, means complete EU control, subject only to EU law, with no critical non-EU dependencies. That is a demanding bar. We use the framework for its questions only and claim no score or level against it.

Five of its objectives turn into checks an inference buyer can verify:

The strategic, supply chain and environmental objectives go beyond a data review, so we leave them out here.

Check who contracts with you and which law applies

The contract check asks which company signs and which national law governs it. The framework lists exactly this under legal and jurisdictional sovereignty. It also names exposure to non-EU laws with cross-border reach, such as the US CLOUD Act.

For the contract-summary feature, this decides which law you would rely on if something goes wrong. Foreign-law exposure is a question for your counsel, so record it in your note.

Our answer is in our terms of service. Your contract is with Lyceum Technology Germany GmbH, and the terms are governed by the law of the Federal Republic of Germany. The data processing agreement (DPA) is Annex 1 of those terms, so it applies when you accept them. Under it, you are the controller and we are the processor.

Check where requests run and which subprocessors touch them

The request-path check asks where processing happens and who else takes part. The framework looks for storage and processing confined to European jurisdictions, with no fallback to third countries. A region name answers only part of that.

Where the GPUs sit. Our homepage names Paris and Finland as the GPU locations of our serverless models. Dedicated and VM locations are agreed per contract.

Who else touches the request. Our subprocessors are mainly in the EU and the EEA. Anything outside is covered by the GDPR's safeguards, such as standard contractual clauses. The GDPR lists standard data protection clauses adopted by the Commission as one such safeguard. You can ask us for the subprocessor list by email. Which subprocessors apply depends on the service you choose, so your choice can rule one out.

What happens when that list changes. We tell you before we add or replace a subprocessor, and you can object within 10 business days. The GDPR requires a processor to give the controller that chance to object.

Staff access is the fourth piece. The people who run our platform work for Lyceum Technology Germany GmbH in Berlin and Lyceum Switzerland GmbH in Zurich. They can access account details, usage records and system logs, not your prompts or responses. If your policy covers staff locations, record both entities in your note.

For the contract feature, the subprocessor list shows whether a third country appears in the chain. Our guide to vetting EU inference providers covers the follow-up questions.

Check what is stored and what trains models

The retention check asks what a request leaves behind. The framework asks for visibility into when, where and by whom data is accessed, and for verifiable removal. For an inference API, the sharper question is narrower: which copies of the prompt exist after the answer?

Our Trust page walks one request through each stop. During a request, your prompt sits in memory on our gateway and a GPU while the model works. Our gateway keeps a usage record with your organisation and user, the model, token counts and the time. It never logs the prompt text or writes it to a database.

One thing does stay for a while. Recent prompt prefixes remain in GPU memory until newer requests replace them, which is what makes cached input cheaper. Our terms put the same rule in contract form for the inference product. Inputs and outputs are not stored beyond what is technically necessary, nor used for training.

For the contract feature, that means no clause text sits in a log you would later have to delete.

Check that your models and code can leave

The exit check asks whether you can move without rebuilding. The framework looks at migration without vendor lock-in and at integration through non-proprietary APIs. It also values software under open licences with rights to audit and modify.

Open models cover the licence half. You can inspect the weights and run them elsewhere, subject to each model's own licence. Our press page states the aim: open models you can inspect and move, on GPUs in European data centres.

The API covers the code half. Our llms.txt puts it in one line: swap one endpoint to move in, keep your code to move out. Your client talks to an OpenAI-compatible API, so the exit is a base URL and a model ID.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.lyceum.technology/openai/v1",
    api_key=os.environ["API_KEY"],
)
reply = client.chat.completions.create(
    model="moonshotai/kimi-k3",
    messages=[{"role": "user", "content": "Summarise clause 4 of this contract."}],
)
print(reply.choices[0].message.content)

To leave, point `base_url` at another OpenAI-compatible endpoint and pick its ID for the same open model. Your prompts and tests stay as they are.

Data follows the same logic. When you leave, you email us whether you want your data deleted or returned. Prompts and responses were never stored, so there is nothing of theirs to return. If you need rights under the EU Data Act, our terms let you ask for an annex in your contract. Our exit planning guide covers the test you should run before signing.

Ask for the evidence before you sign

The evidence check asks which documents back the answers above. The framework looks at recognised certifications such as ISO, and at transparent, timely reporting of breaches.

Here is where we stand. Our ISO 27001:2022 audit is in progress, and the engagement letter is open to read. Our DPA obliges us to tell you immediately about any breach involving your personal data. Service incidents appear on our public status page. Most review documents are open to read in Vanta, our compliance platform, and the technical and organisational measures (TOMs) can be requested there.

Use this as the skeleton of your approval note. For the contract feature, every row needs a document or a named gap.

Approve when every row holds a document or a gap you accept in writing. A provider that cannot produce the document for a row has given you its answer for that row.

Send your security questionnaire or DPA questions to our compliance team via our Trust page.