European AI Clouds 2026: Model Prices and Where Data Stays

Six European cloud providers publish what they charge, model by model, to run AI for a company: STACKIT, IONOS, OVHcloud, Scaleway, Mistral and Infomaniak. On October 26, 2026 the same open model, gpt-oss-120b, cost €0.08 per million input tokens at OVHcloud and €0.45 at STACKIT. Where requests are processed ranges from one named German region to, at Mistral's standard rate, no location commitment at all. This reference lists every general text model the six providers price publicly, in their own currency and with their tax status, where each one processes requests, and what it does with what you send. It is written for companies choosing where their own AI system runs.

On October 26, 2026, gpt-oss-120b cost €0.08 per million input tokens at OVHcloud and €0.45 at STACKIT: one model, two price lists, a factor of more than five. Price lists like these exist at six European clouds, model by model.

Two things matter more than the price: the place where each request is processed, and what the provider does with what your system sends. This reference puts both next to every published price.

Which European clouds host AI models for companies?

As of October 26, 2026, six European providers publish per-model prices for running AI models on their own servers: STACKIT and IONOS in Germany, OVHcloud, Scaleway and Mistral in France, and Infomaniak in Switzerland. A company pays per token, the unit these providers bill by, and its own software connects to the model over the internet. Four more European providers sell the same kind of service without a public per-model price list: T-Systems, mittwald, Aleph Alpha and Nebius.

These are services a company's own systems call: a quoting tool, an inbox assistant, a document reader. They are not chat assistants that staff log into; those seats are compared in our guide to European alternatives to ChatGPT for business teams. STACKIT and T-Systems describe their offers as sovereign; what that means in practice shows in the processing location and data handling further down.^[1][12]

The counts cover general text and chat models only. Most of the providers also price models for searching documents, transcribing speech, reading scans or creating images; those are left out here. Smaller resellers that pass requests on to other providers are not covered.

Every price in this article sits in one spreadsheet, a row per provider and model with the source page and check date: CSV file of European AI cloud prices.

What does a million tokens cost at European AI clouds?

Per million tokens, input prices run from €0.04 (gpt-oss-20b at OVHcloud) to €1.80 (glm-5.3 at Scaleway) on October 26, 2026, and output prices from under €0.10 to €7.50 (Mistral Medium 3.5 at Scaleway). The model decides where in that range a price falls. Input is what your system sends; output is what the AI writes back, and it usually costs more. The same model often has different prices. Open models, whose files are published so that any provider can run them, make the comparison possible: gpt-oss-120b, an open model released by OpenAI, costs €0.08 per million input tokens at OVHcloud, €0.15 at Scaleway and IONOS, and €0.45 at STACKIT.^[5][7][3][1]

The table lists every model that at least two of the six providers price publicly. Prices are per million tokens, as each provider lists them, without conversion.

IONOS also lists a "Mistral Small 24B Instruct" at €0.10 and €0.30 without naming the release, so it is not grouped with version 3.2.^[3] Mistral's figures are the euro prices its own page shows for EU processing; its standard rate is 10% lower.^[9]

Three models in the table show what a "same model" comparison can look like. Qwen3.5-397B-A17B costs exactly €0.60 and €3.60 at IONOS, OVHcloud and Scaleway. Llama 3.3 70B is cheapest at STACKIT. Qwen3.8-27B costs €0.65 per million output tokens at STACKIT and €2.70 or more at three other providers, so the cheapest provider depends on the model you need.

What does a million tokens mean for a company's AI bill?

OpenAI's rule of thumb is that 100 tokens make about 75 English words, so an email thread of 1,500 words with its instructions comes to about 2,000 tokens.^[16] Budget extra for German: the same content splits into more tokens than in English, as our token count of German and English text shows. Because sending and receiving are billed separately, a monthly bill depends on both volumes and on the model.

As an illustration, picture a system answering 10,000 requests a month. Every request carries about 2,000 tokens, an email thread with its instructions, and returns about 500. Over the month that adds up to 20 million tokens in and 5 million out. Priced at the October 26, 2026 lists, before tax (Mistral states no tax treatment):

The model moves this bill more than the provider does. Between the cheapest and the most expensive provider of gpt-oss-120b, the month differs by €8.65. Moving to a larger model costs €30.00 or €63.11, between two and seventeen times as much. Free allowances and discounts shift small bills further: Scaleway does not charge for the first 1,000,000 tokens, Infomaniak gives 1 million free credits, and Scaleway and Mistral halve the price for work sent in batches that can wait for an answer.^[7][11][9]

The volumes are an assumption for illustration, not a measured workload. They show the order of magnitude: for a system of this size, tokens cost tens of euros a month, not thousands. The larger costs sit in building the system, connecting it to your tools and keeping it running, which our breakdown of what an AI implementation costs in Europe sets out.

Why does the same AI model cost different amounts?

Each European provider sets its own price for the same open model, and the gaps follow each provider's price structure more than the model. STACKIT sorts its text models into three price levels on its price list of October 6, 2026: €0.15 and €0.25, €0.45 and €0.65, or €1.50 and €1.75 per million input and output tokens.^[2] Five of its seven text models sit on the middle level, which is how Qwen3.8-27B costs €0.65 per million output tokens at STACKIT and €2.70 at IONOS and OVHcloud.^[1][3][5]

Four other differences sit behind the same model name:

• Version. IONOS lists "Mistral Small 24B Instruct" without a release number; OVHcloud and Scaleway name version 3.2 with the release code 2506.^[3][5][7]
• Format. OVHcloud lists the number format each model is stored in, for example "fp4" for gpt-oss-120b and "fp8" for Llama 3.3 70B; STACKIT names "FP8" in its Llama and Qwen3 VL listings.^[5][1] Ask which format you get when you compare two listings.
• Discounts and allowances. Scaleway gives the first 1,000,000 tokens free and charges half for batch work; Mistral halves batch prices; Infomaniak starts accounts with 1 million free credits.^[7][9][11]
• Lifetime. OVHcloud ends Mistral Small 3.2 and Mistral Nemo on November 4, 2026.^[5] IONOS retired Llama 3.1 8B on October 15 and Mistral Nemo on October 22, 2026, naming a replacement for each.^[18] STACKIT marks Qwen3.6 27B as deprecated in its documentation.^[17]

A cheap listing that ends next week is not cheap. Before comparing prices, write down the exact model name, its version and its end-of-life date at each provider.

Where do European AI clouds process your data?

Five of the six providers name a country for every request as of October 2026: STACKIT and IONOS process in Germany, OVHcloud in Gravelines and Scaleway in Paris, both in France, and Infomaniak in Switzerland. Mistral processes in EU and EFTA countries only when a company sends requests to its separate EU access point, which costs 10% more; at the standard rate, "Mistral does not commit to a specific inference location", meaning where the model runs, its documentation says.^[10]

Two more details change decisions. Mistral states that its EU access point serves only the models hosted in that region, so a model on its global price list may not be available there.^[10] And Switzerland is not in the EU: if a client contract requires processing inside the EU, Infomaniak's Swiss location needs checking against that wording. For the US clouds, the same two questions are answered in whether AI data residency keeps your data in Europe and which AI providers train on business data.

What does each European AI cloud offer?

Each of the six providers sells pay-per-token access to open models from its own data centres, and each prices it in its own way: STACKIT in three price levels, IONOS and OVHcloud per model, Scaleway with a free first million tokens, Mistral with a 10% charge for EU processing, and Infomaniak in Swiss francs. The prices below are the providers' own, per million tokens, as listed on October 26, 2026.

STACKIT AI Model Serving (Germany)

STACKIT, run by Schwarz Digits Cloud, processes requests in its Germany South region (EU01) and states that "your data and queries are neither stored nor used to train models".^[1][2] Its price list gives net prices plus taxes.^[2]

STACKIT also prices two models for searching documents at €0.02 and €0.08 per million tokens.^[1] Qwen3.6 27B is marked deprecated in STACKIT's documentation, without a date.^[17]

IONOS AI Model Hub (Germany)

IONOS processes prompts and outputs in German data centres "and nowhere else", never logs their content and does not use it to train any model.^[4] Its data centres are certified to ISO 27001, and the AI Model Hub falls within the scope of BSI C5, the German federal cloud security catalogue.^[4] The offer is for business customers, and all prices are plus statutory VAT.^[3]

IONOS's English site lists the same service in US dollars; the euro prices above come from its German page.^[3] In October 2026 IONOS retired Llama 3.1 8B (October 15) and Mistral Nemo (October 22) and named Qwen3.5-9B as the replacement for both.^[18]

OVHcloud AI Endpoints (France)

OVHcloud runs AI Endpoints on infrastructure in Gravelines, France, states that data "is not stored or shared during or after model use", and keeps only what it needs for billing.^[19][6] It names ISO 27001, SOC 2 Type II and HDS (French health data hosting) certification for the service.^[6]

The catalogue shows prices without a tax note. OVHcloud's German price list shows the same items with 19% VAT added, so the catalogue figures are net.^[22] Models can be tried without an account at two requests per minute; with a key the limit is 400 requests per minute per project and model.^[19]

Scaleway Generative APIs (France)

Scaleway processes requests in its Paris region and states that it does "not collect, read, reuse, or analyze the content" of prompts or outputs. It keeps request content for up to two weeks only when a request causes errors or harms the service, and does not use it for training.^[8] Prices are before tax; the first 1,000,000 tokens are free, and batch work costs half.^[7]

Scaleway also lists lower prices for repeated input it has already processed (cached input) on three models; the table shows the standard price.^[7]

Mistral AI (France)

Mistral sells its own models and two from Z.ai. Its price page shows US dollars by default and euros on request; the euro figures equal 85% of the dollar figures, and the page states no tax treatment.^[9] EU processing runs through a separate access point (api.eu.mistral.ai) at 1.1 times the standard price.^[10]

Two Mistral settings need a decision. Training: paying customers can opt out, and the free mode may train on what it receives.^[20] Retention: Mistral can keep nothing after answering ("zero data retention"), but only on paid plans, on request and after its own review.^[21]

Infomaniak AI Services (Switzerland)

Infomaniak runs its AI services in its own data centres in Switzerland and states that prompts are neither stored nor used to train models or improve its services.^[11] Prices are in Swiss francs, excluding VAT, and new accounts start with 1 million free credits.^[11]

Apertus, listed under the name swiss-ai, is offered by none of the other five providers.^[11]

What does AWS charge for the same model in Europe?

Processing in Europe costs more at the large US clouds too. On Amazon Bedrock, AWS's AI model service, gpt-oss-120b costs $0.20 per million input tokens and $0.79 per million output tokens in Frankfurt, against $0.15 and $0.60 in North Virginia, according to AWS's price list of October 16, 2026: about a third more for the same model in Europe.^[23]

The four European price lists for the same model run from €0.08 to €0.45 for input and €0.40 to €0.65 for output. Dollars and euros are not converted here, so compare the ranges, not exact amounts.

Mistral's EU access point follows the same pattern at 1.1 times its standard price.^[10] OpenAI, Microsoft and Google handle European processing through settings and separate offers with their own conditions; our guide to where Azure, Google and OpenAI process AI requests by default covers each one.

For a US company, a European location matters when a European client's contract or the company's own policy requires it. Without such a requirement, the same model on AWS costs less in a US region, as the price list shows.

Which European AI providers publish no per-model price?

Four European providers sell access to AI models without a public per-model price list as of October 26, 2026. T-Systems shows one example price, €0.20 per million input and €0.65 per million output tokens for models on Telekom's own cloud, with plans from a €1,000 monthly minimum. mittwald sells monthly bundles from €9 for 5 million tokens. Aleph Alpha and Nebius ask for a sales contact, a sign-up or a log-in.^{[12][13][14][15]}

A minimum commitment changes the comparison more than the token price. At T-Systems' example rate, the illustrative month above would use €7.25 of tokens, but the Essential plan bills at least €1,000.^[12] A bundle works the other way: mittwald charges €9 a month for up to 5 million tokens, so the price per token depends on how much of the bundle a company uses.^[13]

What should you check before choosing a European AI cloud?

Before choosing a European AI cloud, confirm seven things in the provider's own documents: the processing location, prompt handling, the exact model and its end date, currency and tax, rate limits, commitments, and the contract paperwork. Price is only one line of that list. On October 26, 2026, the six providers listed 52 general text models between them, in three currencies and with four different tax notes, so like-for-like takes some care.

The one-page price sheet for STACKIT, IONOS, OVHcloud, Scaleway, Mistral and Infomaniak (PDF) repeats the same-model table, the processing locations and these checks, dated October 26, 2026.

A provider's promises cover its own part. The obligations that stay with your company, whichever provider you choose, are set out in our guide to GDPR obligations for companies deploying AI agents.

When is choosing a provider part of building a system?

Choosing a European AI cloud is one decision inside a working system, and it does not stay made: in October 2026, IONOS retired two models, and OVHcloud will end two more on November 4.^[18][5] A system built on one of them has to move to another model, be tested again and keep running, while prices and new models keep arriving every month.

That work sits outside the price list. The instructions a system gives the model, the checks on its answers and the connections to email, files or an ERP system all have to survive a model change. Our guide to what switching AI providers involves walks through that move.

A shared connection format makes the move smaller. IONOS, OVHcloud and Infomaniak each state that their service follows OpenAI's interface, and OVHcloud says this lets a company "switch providers without rewriting a line of code".^[3][6][11] The answers still change when the model changes, so a switch needs testing even when the connection stays the same.

The provider and model are therefore chosen together with the system around them: the process it takes over, the data it touches and where that data has to stay. That is the shape of a custom AI agent built for one operational function, and it is the work Yardwork does: we choose the provider and model with you, build the system around the process, and move it when the model underneath is retired.

Frequently asked questions

Which European cloud providers offer AI models? As of October 26, 2026, six European providers publish per-model prices for hosted AI models: STACKIT and IONOS in Germany, OVHcloud, Scaleway and Mistral in France, and Infomaniak in Switzerland. T-Systems, mittwald, Aleph Alpha and Nebius also sell access to AI models, but without a public per-model price list.

How much does an AI model cost per million tokens in Europe? At the six European providers that publish prices, a million input tokens cost from €0.04 to €1.80 on October 26, 2026, and a million output tokens from under €0.10 to €7.50, depending on the model. The same model can cost more than five times as much at one provider as at another: gpt-oss-120b input was €0.08 at OVHcloud and €0.45 at STACKIT.

Which AI platform is GDPR compliant? No AI platform makes a company GDPR compliant on its own; compliance depends on how the company uses it. What a provider controls is where it processes requests, whether it stores prompts and whether it trains on them. IONOS, STACKIT, OVHcloud, Scaleway and Infomaniak each state that they do not train on customer prompts, and each names the country where requests are processed.

Is there a European alternative to AWS for AI models? Yes. STACKIT, IONOS, OVHcloud, Scaleway, Mistral and Infomaniak sell pay-per-token access to AI models from their own European data centres. Several host the same open models as Amazon Bedrock: gpt-oss-120b costs $0.20 per million input tokens on Bedrock in Frankfurt and between €0.08 and €0.45 at the four European providers that list it.

Notes

1. STACKIT, STACKIT AI Model Serving, accessed October 26, 2026. Used for model prices (region EU01), the document-search models and the data statement.
2. STACKIT, Price list, version 1.0.44, last updated October 6, 2026, accessed October 26, 2026. Used for "Germany South" as region EU01, net prices plus taxes, the three price levels and the operator.
3. IONOS, AI Model Hub (German page), accessed October 26, 2026. Used for model prices in euros, VAT and business-customer terms.
4. IONOS Docs, AI Model Hub: Data handling, accessed October 26, 2026. Used for processing location, logging, training, ISO 27001 and BSI C5.
5. OVHcloud, AI Endpoints catalogue, accessed October 26, 2026. Used for model prices, number formats and end-of-life dates.
6. OVHcloud, AI Endpoints, accessed October 26, 2026. Used for EU hosting, training, billing data and certifications.
7. Scaleway, Generative APIs pricing, accessed October 26, 2026. Used for model prices, tax, Paris region, the free tier, cached input and batch discount.
8. Scaleway Docs, Generative APIs data privacy, accessed October 26, 2026.
9. Mistral Docs, API pricing, accessed October 26, 2026. Used for dollar and euro prices, standard and regional rates, and batch pricing.
10. Mistral Docs, Regional inference, accessed October 26, 2026.
11. Infomaniak, AI Services prices and AI Services (German page), accessed October 26, 2026. Used for prices in Swiss francs, VAT, free credits, location and data use.
12. T-Systems, AI Foundation Services: Plans and pricing, accessed October 26, 2026.
13. mittwald, AI Hosting, accessed October 26, 2026.
14. Aleph Alpha, home page, accessed October 26, 2026.
15. Nebius, Token Factory prices, accessed October 26, 2026; the page redirects to a log-in.
16. OpenAI Help Center, Understanding and counting tokens, accessed October 26, 2026.
17. STACKIT Docs, Available shared models, accessed October 26, 2026.
18. IONOS Docs, AI Model Hub: Models, accessed October 26, 2026. Used for retirement dates and replacements.
19. OVHcloud Docs, AI Endpoints: Features, capabilities and limitations, last updated February 3, 2026, accessed October 26, 2026. Used for Gravelines, data storage and rate limits.
20. Mistral Help Center, Do you use my user data to train your artificial intelligence models?, accessed October 26, 2026.
21. Mistral Docs, Zero data retention, accessed October 26, 2026.
22. OVHcloud, Public Cloud prices (German page), accessed October 26, 2026. Used for the net price and VAT display of AI Endpoints items.
23. AWS, Amazon Bedrock price list files, published October 16, 2026, accessed October 26, 2026. Used for gpt-oss-120b and Llama 3.3 70B on-demand prices in EU (Frankfurt) and US East (N. Virginia).