Sovereignty & compliance

European inference providers: who actually runs AI models in Europe

A European inference provider is a company incorporated in Europe that runs AI models on its own GPUs in European data centres, sells access per token, and says in writing what happens to your prompts. In October 2026 that describes OVHcloud and Scaleway in France, Mistral AI on its EU endpoint, IONOS, STACKIT, T-Systems and AKI.IO in Germany, Infomaniak in Switzerland, Berget AI in Sweden, Regolo in Italy and TensorX in Ireland; Nebius, in the Netherlands, comes with caveats. Routers such as cortecs and Melious sit in front of several of them. Below: who each one is, where it runs models, what it keeps, and which of them serve DeepSeek, GLM and Kimi.

Published 13 Oct 2026

What counts as a European inference provider

“Inference” is the work of running a trained model on your prompt. An inference provider does it for you, on its GPUs, behind an API. Many companies sell that; far fewer are European in a way that holds up when a data protection officer reads the contract. This list applies three tests, each checked on the provider's own pages on 13 October 2026:

The company. The entity you contract with is incorporated in the EU, the EEA or Switzerland — read in the legal notice or the terms, not on the home page. Who owns it is noted where it is published.
The place. The provider names where inference runs: a country, a city or a region it commits to. “Hosted in Europe” with no commitment for the endpoint you call does not count.
The prompts. A written statement on what is kept, for how long, and whether it is used for training.

The first test is the one most often skipped. The EU region of a US cloud passes the second test, not the first: the US CLOUD Act lets US authorities require a provider under US jurisdiction to disclose data in its “possession, custody, or control”, wherever it is stored source. Our article on the GDPR and the CLOUD Act explains why EU hosting alone does not settle it.

Two other kinds of company appear further down, because they are often counted as inference providers: routers, which forward your request to someone else's GPUs, and GPU clouds, which rent you the hardware but sell no per-token API.

The providers that run the models themselves

Each of these runs open-weight models on infrastructure it operates or rents in Europe, and sells them through an OpenAI-compatible API. Each cell links to the page it comes from.

EUROPEAN INFERENCE PROVIDERS · READ ON 13 OCTOBER 2026
Provider	Company	Where models run	What it says about prompts
OVHcloud AI Endpoints	OVH SAS, Roubaix, France, a subsidiary of OVH Groupe SA source	Gravelines, France source	“Your data is never used to train our models, and we only keep what is strictly necessary for billing” source
Scaleway Generative APIs	Scaleway SAS, Paris, France source, part of the iliad group source	Paris only source	Zero data retention by default, no training; requests that cause errors or look abusive may be kept and read for up to 14 days source
Mistral AI	Mistral AI, Paris, France source	EU endpoint: data centres in EU and EFTA countries; the default global endpoint commits to no location source	API inputs and outputs kept 30 rolling days to monitor abuse, unless zero data retention is on source; training is subject to your opt-out, except on Labs and Preview models, where opt-outs do not apply source
IONOS AI Model Hub	IONOS Cloud GmbH, Montabaur, Germany source	“All inference is performed in IONOS CLOUD data centers in Germany” source	Prompts and outputs never logged or stored, not used for training; no third-party sub-processors source
STACKIT AI Model Serving	Schwarz Digits Cloud GmbH & Co. KG, Germany source, in the Schwarz Group with Lidl and Kaufland source	Region “Germany South” source	“Your data and queries are neither stored nor used to train models” source
T-Systems AI Foundation Services	T-Systems International GmbH, Germany source	T-Cloud in Germany for open models source	For the models it hosts, prompts and responses “are never saved or used to train or fine-tune any model” source
Infomaniak AI Services	Infomaniak Network SA, Geneva, Switzerland source	Its data centres in Switzerland source	“Your queries are neither recorded nor used to train models or improve our services” source
AKI.IO	AKI.IO GmbH, Berlin, Germany source; its shareholders are not published	Germany, on GPUs of AIME GmbH, Berlin source: “No hyperscalers, no third-party routing” source	Prompts and outputs “are not stored, logged, or used for training” source
Berget AI	Berget AI AB, Stockholm, Sweden source	Its own hardware in Sweden source	“Never stores the content of your prompts or generated output, and never trains on your data” source
Regolo	Seeweb S.r.l., Frosinone, Italy source, in the DHH group source	Data centres in Lazio and Lombardy, Italy source	“We do not retain, log, or reuse any input or output data” source
TensorX (formerly Tensorix)	TensorX Ltd, Dublin, Ireland source	GPUs in Dublin and Helsinki source	Held in memory for the request only; “We do not log, retain or train on your inference data” source
Nebius Token Factory	Nebius B.V., Schiphol, Netherlands source; parent Nebius Group N.V., listed on Nasdaq source	Public endpoints: “No Region Commitment”; a fixed region needs a dedicated endpoint source	Content not used for training; prompts and outputs may be stored for speculative decoding, on by default, unless zero data retention is switched on source

Three that need a closer look

Mistral AI serves two places from one API. Requests to its default global endpoint come with no commitment on where they run; its EU endpoint keeps inference in EU and EFTA data centres, for 1.1 times the list price, and Mistral notes that account data, keys and billing — the control plane — are not regional source. Mistral also hosts third-party open models, such as Z.ai's GLM-5.3, on both endpoints source.

Nebius is a Dutch company with a parent listed on Nasdaq, and its public endpoints carry “No Region Commitment”: the processing location “is decided dynamically, is not tied to any single data center, and can change at any time without prior notice” source. Its sub-processors for model processing include US entities — Nebius Inc., RunPod, BoostRun, Shadeform and Axe Compute source. Nebius passes the company test; on public endpoints, it does not pass the place test.

T-Systems is two services in one. Open models run on T-Cloud in Germany source; the closed models it also offers, from Azure and Google Cloud, “may be processed worldwide” source. Its plans start at €1,000 a month, billed as the minimum or your usage, whichever is higher source.

Who can sign up

Businesses only: IONOS source, AKI.IO source and Berget AI source.
A plan first: Berget AI has no pure pay-as-you-go, plans start at €25 a month source; T-Systems at €1,000 a month source.
A free start: 1,000,000 free tokens at Scaleway source; 1M free credits at Infomaniak, card required source; a 30-day trial without a card at Regolo source; a 30-day trial at STACKIT, after a manual check of each registration source.

Routers: one key, several European providers

A router owns no GPUs. It takes your request and forwards it to one of the providers above, so two companies handle each prompt, and two policies apply: the router's and the provider's. Three routers that sell access to European inference:

ROUTERS · READ ON 13 OCTOBER 2026
Router	Company	Where requests go	What it charges and keeps
cortecs	Cortecs GmbH, Vienna, Austria source	Sixteen providers, Amazon, Google and Microsoft among them source; the eu_native option keeps a request on providers located in the EU source	Upstream list prices plus a 5% service fee on credit purchases source; it stores no prompt, and each upstream's own retention policy applies source
Melious	Melious AI GmbH, Saarbrücken, Germany source	Twelve European inference sub-processors, Nebius among them source; “we don't keep a US or APAC fallback” source	Pay as you go, net prices, for businesses only source; it does not log prompt content, while a provider “may keep short-lived operational logs” source
Requesty	Requesty Ltd, London, United Kingdom, incorporated in 2023 source	An EU endpoint hosted on AWS in Frankfurt source, but “where inference happens follows the model you select” source	Plus 5% on what you spend source; on self-serve plans, prompt logging is on by default, up to 30 days, in the EU source

cortecs is the most explicit of the three about its route: its data processing agreement names each upstream company, country and processing location, and says the router itself is hosted by OVH in France source. Melious gives different counts of its partners on different pages — nine, eleven, twelve — so read its sub-processor list rather than its home page source. Requesty's EU endpoint governs Requesty's own processing; it does not change where the model runs, so you have to pin models whose hosts are in Europe.

Not every router with an “EU” setting is European. OrcaRouter is run by a Singapore company, and its EU workspace region tags compliance reports without moving inference: our OrcaRouter comparison reads its documentation line by line. OpenRouter, a US company, offers EU in-region routing on its paid plans; OpenRouter alternatives covers it.

European, but not a per-token API

Three European companies often appear in these lists, but sell GPU time rather than tokens. You deploy and pay for the model's hardware, not for each request.

Gcore (G-Core Labs S.A., Luxembourg) source. Its Everywhere Inference answers its own question “Do you have pay-per-token hosted models?” with dedicated models billed by GPU time source. Its catalogue can deploy DeepSeek V4 Flash or Kimi K2.6 on your own pods, and if a pod in one region fails, requests are routed to the next-closest region source.
Koyeb (Koyeb SAS, Boulogne-Billancourt, France) source. Serverless GPUs billed per second, in Washington, Frankfurt or Singapore source. In February 2026 it signed a definitive agreement to join Mistral AI source.
Verda, formerly DataCrunch (DataCrunch Oy, Finland, under Finnish law) source. GPU instances and serverless containers in Helsinki, priced per GPU-hour, with no public per-token price list source. It hosts TensorX's GPUs in Helsinki source.

Who serves DeepSeek, GLM and Kimi in Europe

The open-weight models from Chinese labs are where European hosting matters most: DeepSeek's own service stores personal data in China source, yet the weights are published, so a European company can run them. On 13 October 2026, these European providers sold them per token. Prices are per million tokens, input / output, before tax, as each provider publishes them; the last column says whether, and where, Evanesca serves the same model.

OPEN MODELS FROM CHINESE LABS, RUN IN EUROPE
Model	European hosts, price per million tokens	Caveat	On Evanesca
DeepSeek V4 Flash (0731)	Scaleway (Paris) €0.40 / €0.80 source · TensorX (Dublin and Helsinki) $0.25 / $0.30 source	Nebius lists it; its page showed a US processing region (us-central1) source	DeepSeek V4 Flash, Scaleway in Paris
DeepSeek V4.1 Flash	AKI.IO (Germany) €0.25 / €1.00 source · Scaleway (Paris) €0.60 / €1.80 source · TensorX (Dublin and Helsinki) $0.50 / $1.50 source	Nebius lists it; its page showed a US processing region (us-north1) source	DeepSeek V4.1 Flash, AKI.IO in Germany
DeepSeek V4 Pro (0813)	TensorX (Dublin and Helsinki) $2.00 / $4.00 source	Nebius lists it; its page showed a UK processing region (uk-south1) source	Not served
GLM-5.3	AKI.IO (Germany) €1.00 / €3.50 source · Scaleway (Paris) €1.80 / €5.50 source · Mistral AI (EU endpoint, at 1.1× this list price) $1.40 / $4.40 source · TensorX (Dublin and Helsinki) $1.75 / $4.50 source	Nebius lists it; its page showed a US processing region (us-north1) source	GLM-5.3, AKI.IO in Germany
GLM-5.3 Flash	Berget AI (Sweden) €0.25 / €0.50 source · T-Systems (Germany) €0.20 / €0.60 source · TensorX (Dublin and Helsinki) $0.20 / $0.50 source	—	Not served
GLM-5.2	Scaleway (Paris) €1.80 / €5.50 source · T-Systems (Germany) €1.50 / €3.50 source · Regolo (Italy) €2.00 / €5.20 source · TensorX (Dublin and Helsinki) $1.50 / $4.50 source	—	GLM-5.2, Scaleway in Paris
Kimi K3	Berget AI (Sweden) €3.00 / €15.00 source · TensorX (Dublin and Helsinki) $3.00 / $15.00 source · Nebius (France, eu-west2, as its page showed; no region commitment) $3.00 / $15.00 source	—	Not served
Kimi K2.6	Infomaniak (Switzerland) CHF 0.60 / CHF 3.00 source	Nebius lists it; its page showed a US processing region (us-central1) source	Not served

Scaleway now lists both generations of DeepSeek Flash and GLM-5.3, all in Paris source. AKI.IO is the cheapest of the first-party hosts on DeepSeek V4.1 Flash and GLM-5.3 source. OVHcloud's catalogue has none of these models source. On Nebius, the region printed on a model page is the current one, not a promise: its public endpoints can move source.

Which of them run Evanesca's models

Evanesca is not a router: each model in its catalogue has a main operator chosen in advance, and a few have a second European operator as a fallback — never one outside Europe. Those operators are OVHcloud, Scaleway, Mistral AI, AKI.IO and Infomaniak, all from the first table:

EVANESCA'S OPERATORS, FROM ITS CATALOGUE
Operator	Where	Models on Evanesca
OVHcloud	Gravelines, France	as a fallback: Qwen3.8 27B, GPT-OSS 120B, Mistral Small 3.2, Llama 3.3 70B
Scaleway	Paris, France	Qwen3.8 27B, DeepSeek V4 Flash, GLM-5.2, Qwen3.6 35B, Qwen3.5 397B, GPT-OSS 120B, Qwen3 Coder 30B, Qwen3 235B, Mistral Small 3.2, Llama 3.3 70B, Pixtral 12B · as a fallback: DeepSeek V4.1 Flash, GLM-5.3
Mistral AI	EU region of Mistral AI	Mistral Large 4, Mistral Small 4
AKI.IO	Germany	DeepSeek V4.1 Flash, GLM-5.3, Gemma 4 26B
Infomaniak	Switzerland	Apertus v1.5 70B

Mistral AI lets its customers object to training on API data, and Evanesca's account objects, except for what is sent to Mistral Large 4: it is a preview, and Mistral's commercial terms let it use the data of preview models for training, whatever the objection source source.

The models page gives each model's price and checks, and the sovereignty page what each operator's ownership means. The passport under every answer names the operator that ran it.

How to choose: a checklist

Read the contracting entity and the governing law in the terms, not the marketing page. A European brand can contract through a US entity for US customers, and a US brand through a European one.
Find who owns it. A listed parent, a group or unpublished shareholders are three different answers; write down which one you have.
Check the commitment for the endpoint you will call: Mistral's EU endpoint and its global one differ, and Nebius's public endpoints commit to no region.
Read what is kept, and when: error logs at Scaleway, 30 days of abuse monitoring at Mistral AI, speculative decoding at Nebius, logging on by default at Requesty.
List the sub-processors: the GPU host behind AKI.IO or TensorX, the US entities behind Nebius, the AWS region behind Requesty's EU endpoint.
With a router, read two policies: the router's, and the upstream provider's for the model you call.
Check who may sign up: several providers serve businesses only, and some start at a monthly minimum.
Make a real call and keep the model name the API returns. A provider can retire a model faster than a list like this one can follow.

Frequently asked questions

What is a European inference provider?

A company incorporated in the EU, the EEA or Switzerland that runs AI models on GPUs in European data centres, sells access through an API, and states in writing what it does with prompts. The EU region of a US cloud is not one: the company remains under US jurisdiction.

Which European inference providers serve DeepSeek?

On 13 October 2026: Scaleway (Paris) and TensorX (Dublin and Helsinki) for DeepSeek V4 Flash (0731); AKI.IO (Germany), Scaleway (Paris) and TensorX (Dublin and Helsinki) for DeepSeek V4.1 Flash; TensorX (Dublin and Helsinki) for DeepSeek V4 Pro (0813). Nebius lists DeepSeek models too, but its pages showed processing regions in the US or the UK, and its public endpoints commit to no region.

Does Mistral AI run its API in Europe?

On its EU endpoint, yes: inference runs in data centres in EU and EFTA countries, at 1.1 times the list price. Its default global endpoint makes no commitment on where a request runs, and the control plane — keys, billing, account data — is not regional.

Is a US cloud's EU region a European inference provider?

Not by this definition. The data can stay in Europe, but the company stays under US law, and the CLOUD Act lets US authorities require it to disclose data it controls, wherever that data is stored.

Which European providers does Evanesca use?

Evanesca's models are run by OVHcloud, Scaleway, Mistral AI, AKI.IO and Infomaniak. Each model has one of them as its main operator, a few have a second one as a fallback, and the passport under each answer names the one that answered.