I kept researching the same AI providers. So I built a tool.

I regularly spend time researching AI tools. Every time I build something new for a client, I first have to pick the right tools to give them the best possible setup. The market moves fast, new tools keep appearing, and so a good chunk of my time goes into the ongoing hunt for new AI tools.

Somewhere along the way I noticed I kept researching the same tools over and over, and that certain questions never really got answered. The big one: GDPR-compliant AI. Everyone knows and uses the major providers: Microsoft, AWS, Google Cloud, and of course OpenAI and Anthropic. But the lesser-known ones? What's actually out there, and which providers can you trust with customer data?

So I went looking, sent my research agents out, and collected a mountain of material. While sorting through it, it occurred to me that this could be more than a pile of notes. I built a tool, and it's now live at ki-markt.jonaskrauss.de. 126 providers, compared, categorized, searchable, filterable by hosting region, group, and features. Every provider has a profile card with strengths, weaknesses, and quirks, and all product pages are linked so you can click straight through.

Here's how I structured it, and which providers stood out.

The three levels of data sovereignty

Data sovereignty is the filter I care about most, and for it I adapted a three-layer model.

Layer 1 means: your data ends up in the EU. Whatever you send gets stored and processed in European datacenters. That's the baseline, and many providers offer it. One feature deserves a separate mention: no data retention, meaning the provider doesn't store prompts or responses at all. Any layer can have it, or lack it.

Layer 2 goes further: the control plane lives in the EU too, along with the data. The services and software that actually deliver the offering, including identity and access management. What can still be foreign at this level is the owner. And most companies in this space are American. As good as they are, and some are genuinely excellent, US law has a catch: a US company can be compelled to hand over requested data to the US government, and the data owners don't get told. That applies even when the data never physically leaves Frankfurt.

Layer 3 is the top level: the parent company itself sits in the EU, and the entire offering falls under EU law only. Who needs that? Public administration, healthcare, regulated industries, and anyone who wants to sleep well.

Many providers fall into none of these levels. The tool shows that too.

The categories

"Provider" is a stretchy word. It hides completely different products, which is why the tool sorts them into groups.

The models themselves. The large language models everyone talks about, from the US flagships to the Chinese providers, which are often a lot cheaper. Many Chinese providers come with a practical catch though: no contract template for a data processing agreement (DPA). If I let a provider process my data, I have to sign one. Good providers offer a standard contract you can click through in minutes. With others you have to email a sales team and negotiate, which turns a legal formality into a project.

Harnesses and agents. A model alone is just a model. You can use it for programming or as a general agent, say through a desktop tool that calls a cloud model in the background. Or you run the agent on your own VPS and only keep a local interface. The software around the model, the harness, has become almost as important as the model itself. And if you want to go further: with OpenHands you can even host your own team of agents with automations, fully open source.

Vector embeddings and RAG. This category is really about search. Anything that exists as text, documents in a database, files on a disk, maybe emails, becomes usable for an AI through vector embeddings. Retrieval augmented generation then gives the model fast, precise access to a knowledge corpus on every query. If it's a few documents, maybe up to a book, you can simply hand over the whole thing each time. Beyond that, you hand over only what the current query needs, both for quality and for token cost. Vector databases are what figure that out. They're interesting beyond AI too, by the way: they make large collections semantically searchable, so a slightly different spelling still finds what you meant.

OCR and document intelligence. Paper still runs a lot of offices. Scanned, it arrives as PDFs: incoming invoices, quotes, requests of every kind, letters. And in many, many offices, a human reads what's in those PDFs and types it into a system. That's changing right now. OCR works really well these days, and document intelligence systems pull structured data out of messy scans, and in the other direction help generate documents like quotes and invoices.

Speech. Text-to-speech and speech-to-text let you attach a natural voice interface to any of the above: a person speaks, it becomes text, the text flows into the system. And the answer comes back as natural-sounding speech instead of dry text. That changes what it feels like to work with these systems.

Training and hosting. Once you run serious AI workloads, training a specialized model often beats calling a giant one: better results, and often cheaper at the same time. This category is about who sells you the compute for that.

AI gateways. One API key, hundreds of models behind it, plus routing and failover. OpenRouter is the best known, but the tool tracks over a dozen, and some are built specifically for the European market.

Neoclouds and self-hosting. Neoclouds are a young generation of cloud companies built around serverless GPU inference: you don't rent servers, you call an API. And self-hosting, from Ollama on your own server to a dedicated GPU box, remains the most GDPR-friendly option that exists, because nobody else touches the data. It takes operational effort, so it's a trade, not a default.

What caught my attention

In no particular order, the finds from this round of research:

Perplexity has a lovely API. Sonar does web search and returns answers with citations and links included. If your application needs current facts with sources, that replaces an entire search pipeline.

Cohere is regarded as the multilingual embedding champion. Not the cheapest, but whenever retrieval has to work across German and English documents, this is the name that keeps coming up.

Scaleway stands out among the European full-service providers. They offer the complete cloud portfolio the big US hyperscalers have, plus generative APIs with an OpenAI-compatible endpoint on top. Of the European providers I looked at, they're the furthest along.

Noxtua is the specialist for the legal niche. A sovereign legal AI, built with major German law firms, on European infrastructure, aimed at law firms and legal departments of larger companies. Pricing is only on request, which also tells you something about the target audience.

Rossum for documents and invoices. Document intelligence and automated invoice processing for the enterprise segment. The starter plan officially starts at $18,000 per year. That one number is the whole positioning.

Berget.ai is the one I'll be trying myself for sure. A strong Swedish offering for developers: an OpenAI-compatible endpoint, a good selection of models, plus databases/storage, compute and connectivity, and container management. Fully European and GDPR compliant.

Exoscale sells compute including GPUs. Pay-as-you-go, operated from Switzerland, no procurement process. Interesting for anyone who needs a GPU for a few hours rather than permanently.

Hetzner does have GPUs after all. I always thought they didn't. Wrong. They do, though without per-minute billing: dedicated servers only. The entry model currently costs 214 euros per month, plus a one-time 209 euro setup fee. Not as flexible as a neocloud, but predictable, German, and cheap for sustained workloads.

Requesty is a British company with a serious European answer. Their EU gateway runs in Frankfurt, supports over 200 models, and adds PII detection and scrubbing on the way through: personal data gets masked before a model ever sees it. In my model, that's Layer 2, from a provider outside the EU.

Eden AI offers the most comprehensive package on the whole list. A French aggregator with 500 models: frontier models, open models, OCR, speech, embeddings, all behind one API. The platform sits in France with an EU endpoint and zero retention, so the platform level is Layer 3, even though US models obviously route through the catalog as well. That breadth under full EU jurisdiction was unknown to me before this research.

LiteLLM is an open-source LLM router. Free, self-hostable, speaks to well over a hundred providers. For teams that can run a small service, it's the cleanest answer to the gateway question, because there's no third party left to trust.

Try it

The tool is free, no signup: ki-markt.jonaskrauss.de. Searching it is wonderfully simple. If you have any requirement around the topics I described, you'll find the candidates in a few clicks, with product pages, strengths, and weaknesses. I expect I'll be updating it regularly...