> Quelle: https://souverana.ch/en/insights/souveraene-ki-infrastruktur/
> Sprache: en

# Sovereign AI infrastructure: why the model is the easy part

The language model is the most replaceable part of your AI. Why embeddings and OCR decide your lock-in, and which Swiss hosters offer every component.

Guide · Published 14 Jul 2026 · Updated 24 Aug 2026 · Joel Barmettler

## What infrastructure actually means for AI

**When AI is meant to answer from your own documents, everything that sits in front of the language model decides the outcome. AI infrastructure includes the services that prepare your data: OCR for documents, transcription for audio, embeddings for search. You can swap the language model in an afternoon. Switch the embedding model instead, and you have to re-ingest every document.**

**In brief**

-   When comparing providers, check how much of the data preparation they take on. The more they take on, the more expensive your exit becomes.
-   Choose the embedding? model more carefully than the language model. Switching it means re-ingesting and re-indexing every document.
-   What sounds convenient binds the tightest: letting a provider ingest, embed and store your documents is the closest dependency in the whole setup.
-   Keep the models closest to your data, OCR? and embeddings, open. Only then can you actually switch provider later.

Public debate revolves almost entirely around language models. For a company, though, the language model is only the final step in a longer chain.

This is the third part of our series on sovereign AI and the first of three layer articles: it covers the bottom layer, where your data is prepared.

Series · Sovereign AI

1.  1[What sovereign AI actually means for Swiss companies](/en/insights/was-ist-souveraene-ki/)
2.  2[AI strategy for SMEs: the questions to answer first](/en/insights/ki-strategie-kmu/)
3.  3Sovereign AI infrastructure: why the model is the easy partYou are reading this part
4.  4[Sovereign AI platform: why it binds more than the model](/en/insights/souveraene-ki-plattform/)
5.  5[Sovereign AI integration: when AI starts to act](/en/insights/souveraene-ki-integration/)

## The model does not know your documents

A language model is trained on the public internet. It knows nothing about your company: it has never seen your contracts, your quotes or your meeting minutes either. Anyone using AI in operations wants exactly that, though: answers from their own documents.

Those documents exist as PDFs, as office files, as scanned contracts, as recordings of meetings. Before a language model can use them, they first have to be turned into searchable text. The rest of the infrastructure handles that: an OCR model makes documents and scans readable, transcription writes out audio, an embedding model makes the text searchable, a vector database stores it, and only at the end does the language model compose an answer from the matching passages.

This interplay is called RAG, Retrieval-Augmented Generation, and it runs in two phases. Once, you ingest your documents and store them in searchable form. After that, every question runs through the second phase, from the search through the matching passages to the evidenced answer.

A RAG pipeline as a graph. At the top, your knowledge is prepared once and stored in the vector database; at the bottom, every question runs through the search to an evidenced answer. The database connects both branches, and the embedding model is the same one in both. Four of the stations are AI models in their own right, not just the language model at the end.

Each of these stations is its own decision, with its own providers and its own consequences for sovereignty. Anyone who only looks at the language model overlooks most of it. So let us start with the question of how a company obtains AI in the first place.

## How companies obtain AI

How sovereign a solution becomes depends less on the model chosen than on how much work the provider takes off your hands. Whatever the provider does for you, it also keeps control over afterwards. Four paths, from the greatest convenience to the greatest control:

| Path | Provider | What you get | Data preparation |
| --- | --- | --- | --- |
| Suite with AI over your data | Microsoft, Google, Infomaniak | ready-made apps plus AI on top | the provider, in its own storage |
| AI platform with ingestion | OpenAI, Mistral | components plus managed storage | the provider, in its cloud |
| Cloud with model APIs | OVHcloud | models as a service, you build the pipeline | you |
| Open models, self-operated | Scaleway, PHOENIQS, on-premise | open models, full control | you |

The most convenient are the **suites**. Microsoft and Google do not dominate enterprise AI because of better models, but because your documents and mail already sit with them. [Copilot searches SharePoint and OneDrive↗](/en/insights/microsoft-copilot-datenschutz/), Gemini searches Google Drive; the AI reads the data you already keep there anyway. Infomaniak belongs in the same category, with Swiss infrastructure: Euria searches the files in your kDrive, hosted in Switzerland and running on open models. It is a hosting provider with built-in AI, not a company specialised in AI.

The next category is the **AI platforms**, whose actual product is the AI. OpenAI, and Mistral as the European answer, give you the components through an interface and offer to upload, ingest and make your documents searchable. Mistral covers the full width of this: its own assistant (Le Chat), a platform (AI Studio), OCR, transcription, embeddings and on-premise operation. A sovereign provider can accordingly keep up with the US corporations, from the assistant to the interface.

Next come the **cloud providers** such as OVHcloud, whose core business is infrastructure and who offer AI models as an additional service. Purely running open models goes furthest at Scaleway, at a Swiss provider such as PHOENIQS, or on your own hardware. The more self-contained the operation, the more you set up yourself, and the less can later be taken away from you.

## Where the lock-in actually sits

The language model is the most easily replaceable component. The prompt stays the same, and the same open model runs at a dozen providers.

The dependency forms with the models that process your data. For search and RAG, you convert your entire knowledge base into vectors with one specific embedding model. Documents and search queries must use the same model, or the vectors will not match. The embedding model therefore determines the structure of your entire data set. Switch it, and you have to **re-ingest and re-index every document**, because one model’s vectors are meaningless to another.

Highlighted: the embedding model. It appears twice, at indexing and at every question, and must be the same both times. A switch devalues the entire vector database.

This is exactly why OpenAI, Google and Mistral are so willing to offer to ingest and store your documents for you. At OpenAI, the resulting knowledge base then sits in their vector store, billed per gigabyte and day. At Google, the RAG engine handles ingestion, chunking and embedding. It is convenient, and this is exactly where the strongest dependency forms: the ingested, embedded and stored data set sits in one provider’s vector space and storage, and switching means rebuilding it in full.

The second demanding component is the preparation itself. Extracting clean, structured data from real documents, from Swiss PDFs, scans, tables and three national languages, is the actual engineering work. The pipeline built on top of it is tailored to one specific OCR service’s output format and is just as hard to replace afterwards.

Highlighted: OCR and transcription at the start of the chain. Your output format shapes everything that follows, and locks you into one provider.

At this layer, sovereignty therefore mainly means keeping the models closest to your data, embeddings and OCR, open and able to be self-operated. The language model can be replaced at any time. Switching the embedding model, by contrast, forces you to rebuild your entire indexed data set.

## Choosing the model

That leaves the question most people ask first: the one about the model. It matters, but it is the most easily corrected decision of all. Two reference points are enough to start. For performance, the independent Intelligence Index from Artificial Analysis ranks models by measurable capability, across open and closed alike. For cost, what we worked through in our article on [LLM cost](/en/insights/llm-kosten/) applies: the cheapest model on the price list is rarely the cheapest one in operation.

What matters for sovereignty is only that the model has open weights?. Then you can run it at several hosters, compare their prices and switch whenever you want.

## Choosing the hoster

Anyone who wants to become sovereign chooses a hoster for open models instead of a closed platform. The comparison should go beyond the language model and ask who offers all the models closest to your data on European or Swiss soil. If a hoster only runs language models and you have to fall back on a US API for embeddings or OCR, sovereignty is lost again at exactly the point that matters most.

Key figure

six model classes

A single Swiss provider, PHOENIQS/kvant, today operates language models, OCR, embeddings, a reranker, transcription and vision, all as open models on Swiss servers (among others Apertus 70B, MinerU, BGE-M3, Whisper).

Source: PHOENIQS Model-as-a-Service, catalogue as of July 2026

Four model types sit inside the pipeline (in red): OCR and transcription, embedding, reranking, language model. A sovereign hoster has to offer all four, not just the language model at the end.

European providers cover this offering to differing degrees. OVHcloud lists over forty open models complete with transcription, embeddings and image models. Scaleway operates language models, vision and embeddings on sovereign infrastructure in Paris. Infomaniak offers the same models through an OpenAI-compatible API from Switzerland. And PHOENIQS shows that a Swiss provider can run the complete set closest to the data openly, OCR and embeddings included.

The greatest control comes from self-operating open models on your own or rented GPUs. That is a topic of its own, [which we cover in a dedicated article↗](/en/insights/llm-selbst-hosten/): the setup is demanding, but ongoing operation afterwards is surprisingly modest.

## Do not decide in isolation

One more point on order: you do not choose these models one at a time, because the platform your team works with requires a particular interface and often a matching embedding model. It therefore dictates part of the infrastructure for you. That is why we stay with the model and the data-adjacent services here, and take up the platform in the next article of this series.

## What this means for your company

Sovereign infrastructure delivers AI that is just as good. You assemble the individual components yourself instead of renting a finished, closed solution. The model is the easy part of it. The care belongs with the models that process your data, and with the question of whether you can get them back if it comes to that.

Want model, hoster and the data-adjacent models chosen independently for your case?

[Architecture & Solution Selection](/en/leistungen/ki-architektur/)

Above infrastructure sits the layer your team works with every day. That is what the next part of the series covers, before the last one turns to the integration layer.

The author

![Portrait of Joel Barmettler](/_astro/joel-barmettler.CGKHGWrV_sJ0IG.webp)

Joel Barmettler

AI Architect · Souverana, Zurich

Joel Barmettler guides Swiss companies from AI strategy to integration: sovereign, confidential and production-ready. He built the Swiss AI Hub as its architect and today co-owns its architecture; he personally leads every Souverana mandate. Mandates from one-person firms to Fortune 500 corporations.

[Book an intro call](https://meet.brevo.com/joel-barmettler/30-minute-meeting) [More about Souverana](/en/) [LinkedIn](https://www.linkedin.com/in/joel-barmettler-b9ab361b7)

## Frequently asked questions

What belongs to AI infrastructure?

More than a language model. AI infrastructure also includes the models and services that prepare your data: OCR for documents, transcription for audio, embeddings for search, and a layer for cost control and access. The language model is the most easily replaceable component of the set.

Which AI model should a company choose?

The model that solves the task at the best price, not the one with the biggest headline. The independent Intelligence Index from Artificial Analysis ranks models by capability; which one is cheap in operation depends on token consumption, which we work through in our article on LLM cost. An open model can also be run at several hosters and switched at any time.

Where can I host LLMs in Switzerland or Europe?

As an API at European providers such as OVHcloud, Scaleway or Infomaniak, or on a Swiss model-as-a-service platform such as PHOENIQS/kvant, which runs open models complete with OCR and embeddings. Anyone who wants maximum control operates the models themselves; that is a topic of its own.

What is the difference between an AI platform and an integration?

An integration is AI built into a product you already use (Microsoft Copilot, Google Gemini, Infomaniak Euria). You do not choose the infrastructure. A platform gives you the components through an interface, so you can build your own solution and switch provider.

Why are embeddings a lock-in risk?

Because your entire knowledge base is stored as vectors in the space of one specific embedding model. Documents and search queries must use the same model. Switching models means re-ingesting and re-indexing every document. An open, self-operated embedding model keeps that option open.

LinkedIn

## Share this article

Ready-formatted graphics and a suggested post for your LinkedIn feed: download, copy, post.

![Matrix of European and American AI providers by model type: language model, OCR, transcription, embeddings and managed ingestion. PHOENIQS/kvant and Mistral cover the whole set, Infomaniak and Scaleway have gaps.](/media/souveraene-ki-infrastruktur-en/infografik.png)

[Download infographic (PNG)](/media/souveraene-ki-infrastruktur-en/infografik.png)

Suggested post

Most companies compare the language model when buying AI. Do you know which model reads your documents and translates them into vectors?

That is exactly where the dependency forms, and exactly where the European offering is thin: anyone who needs embeddings or OCR quickly falls back on a US API and loses sovereignty at the point where their own data is actually processed.

A look at the catalogues (as of July 2026) shows the pattern: every provider has a language model. Beyond that, coverage gets patchy. Infomaniak offers no OCR, Scaleway no transcription, OVHcloud only partial OCR. A complete set across six model classes is delivered in Switzerland today only by PHOENIQS/kvant, with Apertus 70B, MinerU, BGE-M3 and Whisper.

Our conclusion: the care belongs with the embedding model, not the language model. You can swap the language model in an afternoon. Switching the embedding model means re-ingesting and re-indexing every document, because one model's vectors are meaningless to another.

The article has the seven-row provider matrix, the RAG chain step by step, and seven questions for your AI hoster. Link in the comments.

#AI #DataSovereignty #RAG #Switzerland