LLM Hangar / EU hosting

EU-hosted LLM API providers compared (2026)

Checked from each vendor's public pages. Corrections are welcome at [email protected].

"EU-hosted" answers one question: in which country do the GPUs sit. It does not answer the question most privacy reviews actually ask: who, apart from you, processes the prompt. Most EU-hosted LLM APIs are either a shared gateway in front of other people's models or a fleet of GPUs the vendor owns. In both cases the request passes through the vendor's systems, and the vendor is a processor of its content, however good their retention policy is.

LLM Hangar is built around a different shape. It deploys an open model onto a GPU instance inside your own AWS, Nebius or RunPod account, in an EU region if you choose, and gives you a private endpoint on that instance. Prompts and responses travel between your client and your instance. They do not pass through LLM Hangar, so there is no second party to list under "prompt data". The table below puts that next to the options you are probably already comparing.

The comparison

Facts are taken from each vendor's public pricing, region and security pages on the date above. "Not stated" means the vendor's pages did not say; it is not a judgement.

Provider Where it runs Runs in your own cloud account? Prompts pass through the vendor? EU surcharge Pricing model
LLM Hangar Your AWS, Nebius or RunPod account; EU regions selectable, EU-only pinning available Yes No. Client to instance only None; EU pinning is a checkbox Flat platform subscription; GPU hours billed by your provider
Scaleway Generative APIs Scaleway infrastructure, France No Yes (vendor-hosted API) Not applicable, EU-native Per token; for example glm-5.2 at €1.80 input / €5.50 output per million tokens on its pricing page
OVHcloud AI Endpoints OVHcloud infrastructure, France No Yes (vendor-hosted API); zero retention stated Not applicable, EU-native Per token
IONOS AI Model Hub IONOS infrastructure, Germany No Yes (vendor-hosted API) Not applicable, EU-native Per token, $0.11 to $4.00 per million tokens depending on model
Mistral Studio Mistral infrastructure, France; self-hosting of Mistral models under a commercial licence Only via the licensed self-host option Yes for the API; no for licensed self-hosting Not applicable, EU-native Per token for the API; licence for self-hosting
Requesty EU Gateway in Frankfurt in front of proprietary model providers' EU endpoints No Yes (gateway), then the upstream model provider Not stated Per token; audit log on the enterprise tier
HostYourAI Vendor-owned GPUs in the EU, dedicated instances No Yes (vendor-hosted) Not applicable, EU-native Prepaid credits, pay as you go
Apertus Vendor-hosted, EU No Yes (vendor-hosted API) Not applicable, EU-native €20 per month base plus €0.11 to €0.95 per million tokens
Tessera Vendor-hosted dedicated GPU, EU and Latin America No Yes (vendor-hosted) Not stated Flat $20 to $30 per month for a dedicated GPU
Northflank BYOC Your own AWS, GCP, Azure or Kubernetes cluster Yes No, in BYOC mode None; "no added cost for running in your VPC" Pay as you go per vCPU, GB and GPU hour; setup is technical (clusters, node pools, images)
Modal Modal-managed infrastructure; EU regions available No Yes (vendor-hosted) EU region selection billed at 1.5x to 1.75x per its region documentation Usage-based compute; audit logs on the Enterprise plan
Fireworks AI Fireworks-managed infrastructure No, except large enterprise arrangements Yes (vendor-hosted) Region-restricted deployments billed at 1.5x per its pricing page Per token and on-demand GPU hours

Two things stand out. The EU-native vendors solve the location question well and are the right choice if you want a managed API and are comfortable with the vendor processing content. The US platforms charge for region pinning, between 50 and 75 percent on top. Running in your own account is the only row where the answer to "who else sees the prompt" is nobody, and where EU pinning costs nothing extra.

Residency is not sovereignty

A useful checklist for EU hosting claims comes from Mate iT: five points that a region name alone does not answer. Here is how each maps to what LLM Hangar actually does.

Checklist pointWhat to askLLM Hangar
Processing region Where is inference actually executed, not just where the company is registered? On a GPU instance in the region you picked. The EU-only option pins every resource of the deployment to EU member-state regions and keeps it there.
Backups and copies Where do snapshots, caches and prepared images live? In your account. A prepared stage kept after a destroy, so the next boot is fast, is storage in your account, listed with its own cost and deletable at any time.
Sub-processors Who else touches the content? For prompt and response content: none. The only party handling it is your cloud provider, under the agreement you already have with them. The dated list is at /sub-processors.
Maintenance access Can vendor staff reach the machine, and how is that bounded? On AWS, access is a cross-account role assumed for an hour at a time and scoped to resources we tagged; delete the CloudFormation stack and it ends. Nebius uses a project-scoped service account, RunPod an API key you can revoke. Details in security.
Deletion timeline When you delete, what proves it happened? Teardown is verified against your provider until the resources are confirmed gone, and the sweep is written to the audit log, which you can export as JSON.

What an EU region does not change

AWS and RunPod are US companies. Nebius is headquartered in the Netherlands and operates EU regions. An EU region of a US-headquartered provider is still run by that provider, and whether that places data outside the reach of US law is a live debate; the discussion around AWS's European Sovereign Cloud is a good summary of both sides. LLM Hangar does not claim to settle it.

What we can say precisely: using LLM Hangar adds no party to the list of who handles your prompts. The control plane that creates and destroys your instances runs in Germany and never sees prompt content. Your account data, deployment metadata and audit log are hosted in Germany. The inference itself happens in your own account, under your own provider agreement, in the region you chose.

The EU AI Act and self-hosting

The AI Act's obligations for high-risk systems have applied since 2 August 2026, and general-purpose model obligations since August 2025 (see the Linux Foundation Europe explainer). When you self-host a model you are the deployer, and record-keeping duties that come with that fall on you rather than on an API vendor. The audit log LLM Hangar keeps for every deployment, including the plan of what will be created before it exists, each resource as it is created, and the sweep after a destroy, is the kind of record those duties ask for. This is a description of what the product records, not legal advice on your obligations.

How to deploy EU-only

  1. Connect your cloud account: AWS, Nebius or RunPod.
  2. Pick a model from the catalog and a shape that fits it.
  3. Tick EU-only. The wizard then offers EU regions only, and every resource, including storage, is created there.
  4. Set a budget cap, confirm the estimate, and copy the endpoint URL and key into your code.

The whole flow is a few clicks; there is no command line, Terraform or Kubernetes involved. The deploy-on-AWS guide walks through it with instance shapes and observed prices.

Questions we get asked

Is EU data residency the same as GDPR compliance?

No. Residency says where data is processed and stored. GDPR compliance also depends on who processes it, under what agreement, how long it is kept, who can reach it for maintenance, and how it is deleted. A vendor can host in the EU and still be a processor of your prompts.

Does hosting in an EU region remove CLOUD Act risk?

Not on its own. An EU region of a US-headquartered provider is still operated by that provider, and LLM Hangar does not claim otherwise. What does change is that no additional party handles your prompts: they travel between your client and an instance in your own account, and the platform that created the instance never sees them.

Which EU regions can I pin a deployment to?

On AWS the EU-only option covers eu-central-1, eu-west-1, eu-west-3, eu-north-1, eu-south-1 and eu-south-2. Nebius and RunPod EU regions are offered where the provider has GPU capacity in them. The option pins every resource of the deployment, including storage, to EU member-state regions.

Do you sign a DPA?

Yes. A signed data processing agreement is available on request; the processing facts it records are summarised at /dpa.

Who are your sub-processors?

For prompt and response content: none. For account and platform data: Hetzner (Germany), Cloudflare, Scaleway (France) and Paddle. The dated list is at /sub-processors.

Start a 24-hour trial