European inference providers in 2026: the list

What counts as a European inference provider

An inference provider runs AI models on its GPUs and sells access to them, usually through an API billed per token. Calling one "European" only means something if three things hold, and they are worth checking separately because a provider can pass one and fail another:

• The company. Where it is incorporated, and who controls it. A US parent or listing brings US law along with it: the CLOUD Act lets US authorities require a company under US jurisdiction to disclose data in its possession, custody or control, wherever the servers are. Hosting location and legal exposure are two different questions.
• The place. Which data centres actually run your request, and whether the provider commits to them. Some publish one city; some route "dynamically" and say the region can change without notice.
• Your prompts. Whether inputs and outputs are kept, for how long, and whether they can be used for training.

The list below gives all three for each provider, quoted from its own pages, read on 2026-10-13. Prices are as each provider publishes them, per million tokens, input then output, before VAT unless stated.

Providers that run the models themselves

OVHcloud AI Endpoints (France)

OVHcloud says its AI Endpoints are "hosted and run in the European Union", with the infrastructure in Gravelines, and that it keeps "only the data required for billing purposes" (product page, docs). The API is OpenAI-compatible and priced in euros before VAT (prices); the catalogue covers Llama, Qwen, Mistral and gpt-oss, but no DeepSeek, GLM or Kimi model today. OVHcloud has a US entity, so the accurate phrase is outside the direct reach of the CLOUD Act.

Scaleway Generative APIs (France)

Scaleway, a subsidiary of the iliad group, says it hosts "all models in a secure datacenter located in France, Paris only" (product page). It applies zero data retention by default, does not train on your data, and may keep requests that cause errors for up to two weeks (data privacy). Its price list now includes DeepSeek V4.1 Flash and GLM-5.3 alongside Qwen, Llama and Mistral models, and new customers get one million free tokens (pricing). Like OVHcloud, Scaleway has a US entity.

Mistral AI (France)

Mistral sells its own models, plus a few third-party ones such as GLM-5.3. Where the request runs depends on the endpoint: on the default one, "Mistral does not commit to a specific inference location"; the EU endpoint runs in EU and EFTA data centres and is billed at 1.1 times the list price (regional inference). Its privacy policy keeps API inputs and outputs for 30 rolling days to monitor abuse unless zero retention is activated, and its terms let it train on data sent to preview models whatever your opt-out (§ 4.3). More in our Mistral API page.

IONOS AI Model Hub (Germany)

IONOS says that all inference is performed in its own data centres in Germany, that prompts and outputs are never logged or used to train any model, and that no third-party sub-processors are engaged (data handling). The service sits in the scope of the BSI C5 attestation, the API is OpenAI-compatible, and the offer is for businesses only, prices plus VAT (product page).

STACKIT AI Model Serving (Germany)

STACKIT is the cloud of the Schwarz group, owner of Lidl and Kaufland, and says it "is not subject to any non-European parent company" (facts). Its model serving neither stores queries nor trains on them, and bills three tiers — standard, plus and premium — rather than one price per model (product page). The shared catalogue lists Qwen, Llama 3.3 70B, gpt-oss and Gemma models (models).

T-Systems AI Foundation Services (Germany)

Deutsche Telekom's T-Systems serves open models such as GLM-5.2, GLM-5.3 Flash, Gemma 4 and Qwen from its T-Cloud in Germany, and says prompts and responses on the models it hosts are "never saved or used to train" (enterprise trust). The same API also forwards to closed models on Azure and Google Cloud, which it places in another category. Plans start at a €1,000 monthly minimum (plans): an enterprise offer, not a self-serve one.

Infomaniak AI Services (Switzerland)

Infomaniak runs its models in its own Swiss data centres and says queries are "neither recorded nor used to train models" (AI services). Switzerland is outside the EU but covered by an EU adequacy decision. The catalogue includes Kimi K2.6, Qwen3.5 397B, Mistral Small 4, Gemma 4 and the Swiss model Apertus; prices are in Swiss francs before VAT, with one million free credits (prices).

AKI.IO (Germany)

AKI.IO, in Berlin, offers "curated open-weight AI models hosted in Germany" with "no hyperscalers, no third-party routing", and says prompts and outputs are processed in volatile memory and not logged, stored or analysed (aki.io). Its GPUs are run by AIME GmbH in Berlin. Its sub-processor list also names Alibaba Cloud in Singapore; AKI.IO confirmed to us in writing on 2026-10-13 that this concerns marketplace models not launched yet (processors). It sells to businesses only, prepaid, and serves DeepSeek V4.1 Flash, GLM-5.3 and Gemma 4 among others.

Berget AI (Sweden)

Berget AI runs open models on its own hardware in Sweden, "under Swedish jurisdiction with zero data retention", and says it never trains on your data (berget.ai). It sells to businesses through monthly plans — there is no pure pay-as-you-go — starting at €25 of credit a month (pricing), and its live list includes Kimi K3 and GLM-5.3 Flash (models).

Regolo.ai (Italy)

Regolo.ai is the inference service of Seeweb, an Italian cloud company, with data centres in Lazio and Lombardy (infrastructure). It states that it does not retain, log or reuse any input or output (zero data retention), and offers pay-as-you-go prices or subscriptions from €39 a month, with a 30-day trial without a card (pricing).

TensorX (Ireland)

TensorX, formerly Tensorix, runs its GPUs in Dublin and Helsinki and says prompts and completions are held in volatile memory only, never stored, logged or used for training (trust). It has one of the widest open-weight catalogues in Europe — DeepSeek V4 Flash, V4.1 Flash and V4 Pro, GLM-5.3, Kimi K3 — priced in US dollars before VAT (pricing). Its sub-processors include AWS in Ireland and Cloudflare.

Nebius Token Factory (Netherlands, with caveats)

Nebius B.V. is a Dutch company, but its parent, Nebius Group N.V., is listed on Nasdaq (legal quick guide). Its public endpoints carry "No Region Commitment": the processing location "is decided dynamically" and "can change at any time without prior notice", and model pages showed US regions for several DeepSeek and GLM models on 2026-10-13. Its sub-processors include US entities, and prompts may be stored for speculative decoding unless you turn zero retention on. It belongs on a European list, but not as a default for data that must stay in Europe.

Routers: one API over several providers

A router does not run GPUs. It sends each request to one of several providers and bills you once — which is what Frontière AI does too. What separates them is who picks the provider, and whether the route can leave Europe.

• cortecs (Cortecs GmbH, Vienna) routes to about sixteen providers — European ones such as AKI.IO, Berget, IONOS, OVHcloud, Scaleway and TensorX, plus the EU regions of Amazon, Google and Microsoft — at the upstream rate plus a 5 % fee on top-ups. Setting eu_native: true keeps a project on EU-native providers only (sovereign cloud).
• Melious (Melious AI GmbH, Saarbrücken) routes across twelve EU inference providers listed on its sub-processor page, says every request runs inside the EU with no US or APAC fallback, and sells to businesses only (sub-processors). Compared in Frontière AI vs Melious.
• LowRouter (Carbonifer SAS, Nantes) passes upstream prices through with a fee on top-ups, across EU and US providers; you keep a request in Europe by pinning the region. Compared in Frontière AI vs LowRouter.
• Requesty is a UK company whose EU endpoint runs on AWS in Frankfurt; its privacy policy says that where inference happens "follows the model you select" (privacy).
• Frontière AI (FLEECE AI SASU, France) — see where we fit, below.

OrcaRouter, often listed next to these, is a Singapore company whose servers run in Singapore, the United States and Japan: not a European provider (comparison).

Close, but not pay-per-token

Three European names come up in these searches without selling a shared per-token API. Gcore (Luxembourg) answers its own FAQ question on pay-per-token hosted models with dedicated deployments billed by GPU time (Everywhere Inference). Koyeb (France), which announced it is joining Mistral AI, sells serverless GPU containers you deploy yourself (pricing). Verda, formerly DataCrunch (Finland), rents GPUs and serverless containers; a hosted per-token API appears in its changelog but not on its public pricing. All three suit teams that want to run a model themselves on European hardware.

Who serves DeepSeek, GLM and Kimi in Europe

The most searched open-weight families come from Chinese labs, and their own APIs run outside Europe — DeepSeek's in China, Z.ai's in Singapore. Because the weights are published, European providers can serve them instead. What the price pages listed on 2026-10-13:

Our own pages go further for each family — which models we serve, who runs them, and what the vendor's API does with your data: DeepSeek API hosted in Europe, GLM API in Europe and Qwen API in Europe.

How to choose

1. Read the company, not the region. A European region inside a company under US control answers where, not who. Check the legal notice, the parent company, and the sub-processor list.
2. Check that the region is committed. "Hosted in Europe" on a product page and "no region commitment" in the legal guide can describe the same service.
3. Read the retention sentence word for word. Zero retention by default, retention with an exception (errors, abuse), or 30 days unless you ask: three different answers.
4. Check who you can buy as. IONOS, AKI.IO, Berget and Melious sell to businesses only; T-Systems starts at €1,000 a month.
5. Keep your code portable. Nearly every provider here is OpenAI-compatible, so switching is a base URL, a key and a model ID.

Where Frontière AI fits

Frontière AI is a router, not a GPU operator. Its 17 sovereign models are served by Scaleway (France), Mistral AI (France), Infomaniak (Switzerland) and AKI.IO (Germany), each model with one named provider shown before you call it and in every response. What it adds: one OpenAI-compatible key across these providers, prices in euros with VAT included, prepaid from €10, access for individuals as well as companies, and no prompt or completion written to our database. What it costs: buying direct is cheaper. OVHcloud lists Llama 3.3 70B at €0.67 per million input tokens before VAT; through Frontière AI, the same model from the same provider costs €1.51 VAT included, about €1.26 before VAT. If you only need one provider, sign up with it directly. If you want several of them behind one key, with the jurisdiction named on every call, that is what we are for. See the catalogue and how we label sovereignty.

FAQ