Private LLM cost for EU companies (2026): EU API tiers vs open models vs your own GPU

"Can we use AI on our own documents without the data leaving the EU?" is the question that stalls a lot of company AI projects. The answer is yes, several ways. What nobody lays out is what each way costs, because the prices sit on a dozen vendors' pages in different units, and some of the most important ones are only in machine-readable price feeds.

So we read them all on 27 September 2026. For AWS and Microsoft, whose price pages load in JavaScript, we pulled the numbers from their own official price APIs. The GPU rental rates come from aliteq's own tracker, which has logged Runpod's and Vast's prices every hour since 21 July 2026. We didn't deploy or benchmark any of these routes, and none of these vendors pay us.

What a private LLM costs a month, route by route

At 10 million tokens a workday, the cheapest way to keep prompts in the EU is an open model on a French cloud: gpt-oss-120b on OVHcloud costs about $29 a month. The same volume on a frontier model's EU tier runs $350 for Claude Haiku 4.5 up to $763 for GPT-6 on Azure's EU Data Zone.

How big is 10 million tokens a day? A 25-person team asking an internal assistant 20 questions a day each, with about 4,000 tokens of documents and question going in and 500 tokens coming back, uses about 2.25 million. So 1 million a day is a small team's light use, 10 million is heavy company-wide use, and 100 million is a product serving customers.

The formula is monthly tokens times the blended price: eight parts input at the input price plus one part output at the output price, divided by nine. Two caveats change the table in practice. Claude's newer models count about 30% more tokens for the same text than other models do, so Claude's real bill per word is higher than a per-token table suggests. And Gemini 3.8 Flash has an introductory price until the end of 2026 that is half the 2027 price shown here; don't budget a year on the intro price.

Try your own volume:

The EU premium isn't one number

Keeping inference in the EU costs anywhere from nothing to 75% more than the same model's global price. For the big frontier APIs it's usually 10%, and for GPT-6 on Azure it's 20%.

What each vendor's own pages say:

• OpenAI charges a 10% uplift for regional processing on models released since 5 March 2026. EU residency isn't self-serve: you contact sales and must be approved for abuse-monitoring controls. For gpt-6-sol and gpt-6-luna it's available only with Standard processing, so the 50% Batch discount doesn't apply.
• Azure's EU Data Zone adds 20% on GPT-6 but 10% on GPT-5.6 and on gpt-oss-120b, according to Microsoft's own price feed. It's self-serve inside Azure.
• Claude's own API has no EU option. Its inference can run globally or in the US only. To keep Claude in the EU you go through AWS Bedrock in an EU region or Google Cloud's EU multi-region, both at a 10% premium. Sonnet 5 costs $2.20 in and $11 out per million tokens there, against $2 and $10 globally.
• Mistral hosts data in the EU by default at its list price. Its Enterprise APIs, which add regional data processing controls and SLAs, cost 75% above list.
• Open models are the surprise. gpt-oss-120b on OVHcloud in France is 34% cheaper than the common US price for the same model. On Bedrock in Frankfurt it's 33% dearer than on Bedrock in the US. There's no single EU premium for open weights.

One more catch: Bedrock's region table tags Zurich and London as "EU" regions, and neither is in the EU. If you need EU-only, pin a region in an EU member state.

When your own GPU pays off

Running your own GPU only beats an API when the alternative is a frontier model. Against gpt-6-sol's EU tier, one rented GPU running gpt-oss-120b in business hours breaks even at about 5 million tokens a workday, and an owned server at about 13 million. Against an EU open-model API, it never breaks even at a volume one GPU can serve.

The reason is capacity. One 80–96 GB GPU such as an H100 or an RTX PRO 6000 fits gpt-oss-120b, and from published throughput benchmarks it can serve about 99 million tokens in an 8-hour day in theory. Against OVHcloud's price for the same model, a rented GPU would need 124 million tokens a day to break even, beyond what it can serve. So the real choice isn't "API or server". It's which model you need:

• If an open model is good enough, an EU open-model API is the cheapest private route at every volume a small company has.
• If you need GPT-6 or Claude quality, you pay the EU tier, and self-hosting only helps if you're willing to swap to an open model at high volume.
• If the data must never leave your building, you own the hardware, and you're buying control, not savings. Our three-year own-vs-rent calculator prices that choice in full.

The own-server figure here is $914 a month: one RTX PRO 6000 workstation at $26,979 over three years, with four hours of admin a month and 20% resale, using Danish business power and wage rates as the EU example. Which models fit which card is on our cost to run pages, and live rental prices are on our GPU price table. Our tracker stores one price per GPU type, not per region, so check the EU data-centre price when you deploy.

Who keeps your prompts, and where

OpenAI, Azure, Google Cloud and Scaleway say in their docs that they don't train on your API data by default, and OVHcloud says data isn't stored or shared. Mistral's pay-as-you-go wording doesn't say which way its training setting starts, so switch it off explicitly. Zero data retention usually needs an application or a sales conversation.

Three details from the docs matter more than the marketing:

• Residency isn't total. OpenAI's EU residency excludes "system data" such as account, billing and usage metadata, which may be processed outside the region. Mistral says data can be temporarily transferred outside the EU for some features, to listed sub-processors.
• Bedrock keeps Claude's maker out of the loop. AWS's docs say model providers have no access to Bedrock logs or to customers' prompts and completions.
• Vendor slogans aren't facts. Scaleway says it isn't subject to the US CLOUD Act, IONOS calls itself 100% GDPR-compliant, and Runpod says it is fully GDPR-compliant in its EU regions. Those are the vendors' own claims. We haven't verified them.

What GDPR actually requires

GDPR doesn't require EU hosting. It requires two things when an AI vendor processes personal data for you: a written processor contract, and a lawful basis for any transfer outside the EU.

• A processor contract. Article 28 says you may use only processors that give "sufficient guarantees", under a written contract that covers what's processed, for how long and on whose instructions, including transfers. That's the vendor's Data Processing Agreement, and you need it whether the servers are in Paris or Oregon.
• A transfer basis. Sending personal data outside the EU needs a basis under Chapter V. For US companies that participate in the EU–US Data Privacy Framework, the Commission's adequacy decision of 10 July 2023 lets data flow freely. Check each vendor on the official Data Privacy Framework list; we couldn't read it on 27 September because the list only loads in a browser.
• Why companies pay for EU residency anyway. The Data Privacy Framework has a live legal challenge: an appeal, case C-703/25 P, was lodged on 31 October 2025, and we found no ruling as of September 2026. EU residency removes the transfer question if the framework falls. It doesn't remove the processor contract.

The EU AI Act adds duties by role. If you put a chatbot in front of customers under your own name, you may be its provider and must tell people they're talking to an AI, a duty that applies since 2 August 2026. An internal document assistant used by staff is usually not high-risk; a system that screens CVs may be, and those rules start on 2 December 2027. The timeline is on our security and compliance hub.

This is a summary of the law and official guidance, not legal advice.

How to choose

The full model-by-model table of EU and global prices, and how to switch residency on with each vendor, is in EU data residency for AI APIs.