Which European Cloud Providers Keep AI Training Data AND Compute Inside the EU?

Seven EU-incorporated providers let you book GPUs for AI training and inference inside EU data centres today: Scaleway (Paris, Warsaw), OVHcloud (France), IONOS Cloud (Germany), StackIT (Germany), Open Telekom Cloud / T-Systems (Germany), UpCloud (Helsinki) and Nebius (Mäntsälä, Finland; Amsterdam-headquartered, Nasdaq-listed). Hetzner sells dedicated GPU servers but no cloud GPU instances. Exoscale is operated by a Swiss company and Nscale runs in Norway, so neither is an EU operator.

I have talked residency with more than 200 CTOs, and the conversation almost always stops at "our data is in Frankfurt". That answers the wrong question. Storage residency is table stakes. Where the GPU job runs, where the inference endpoint terminates, where the logs land and where the support engineer's SSH session originates decide whether your AI stack is actually European.

All prices and availability below were checked on 28 October 2026.

Key Points:

• Seven EU operators run GPU compute in the EU: Scaleway, OVHcloud, IONOS Cloud, StackIT, Open Telekom Cloud, UpCloud and Nebius. Nebius is the biggest by capacity, but it is listed on Nasdaq. Hetzner offers dedicated GPU servers only, Exoscale is run by a Swiss entity, and Nscale sits in Norway (EEA).
• Storage is the cheapest promise to make. Four other layers decide residency: where the training job executes, where inference terminates, where the control plane and telemetry land, and who can open a shell on production.
• A hyperscaler EU region puts your GPUs in Europe but keeps a US-controlled operator, unless you buy a partner-operated or separately incorporated offer: AWS European Sovereign Cloud, Delos Cloud for Microsoft, or T-Systems Sovereign Cloud and S3NS for Google Cloud.
• Verify with five artefacts: the sub-processor annex, a contractual region pin, the key-management model, the non-EU support access policy, and a written exclusion from model training. ISO 27001 and SOC 2 prove nothing about residency.
• Qovery deploys into your own cloud account on Scaleway, AWS, GCP or Azure, or into any existing Kubernetes cluster. The region, KMS keys, VPC, egress rules and cloud bill stay in your name.

Which European cloud providers actually run AI compute in the EU, not just storage?

Scaleway, OVHcloud, IONOS Cloud, StackIT, Open Telekom Cloud, UpCloud and Nebius all run bookable GPU compute in EU data centres under EU-incorporated companies. Hetzner sells dedicated GPU servers but no cloud GPU line-up. Exoscale and Nscale are European but not EU operators. For context, European-headquartered providers hold about 15% of their home cloud market, against 70% for Amazon, Microsoft and Google (Synergy Research Group, July 2025).

• Scaleway (Iliad group, France): H100 PCIe, H100 SXM, L40S and L4 instances in fr-par-2 (Paris), a smaller set in Warsaw, plus an 8x B300 node in Paris (Scaleway GPU pricing). It also runs OpenAI-compatible Generative APIs hosted in Paris. Missing: a Bedrock-sized model catalogue and a managed fine-tuning API.
• OVHcloud (France, listed on Euronext Paris): H200, H100, A100, L40S and L4 instances (OVHcloud prices), plus AI Training, AI Deploy and AI Endpoints. The group also owns OVH US LLC, "a wholly owned subsidiary of OVH Group". Your EU contract sits with the European entity, but the US subsidiary is something to ask about.
• IONOS Cloud (United Internet group, Germany): GPU VMs plus the AI Model Hub. IONOS states that "all inference is performed in data centers in Germany", that no third-party sub-processors are involved, and that customer content "is not used to train, fine-tune or otherwise improve any model" (IONOS docs).
• StackIT (Schwarz Digits, Schwarz Group, Germany): A100 80GB, L40S and 8x H100 HGX machine types (STACKIT docs), with SKE as its Kubernetes engine. Missing: a managed model catalogue at hyperscaler depth.
• Open Telekom Cloud / T-Systems (Deutsche Telekom, Germany): GPU-accelerated instances on its public cloud, and the separate Industrial AI Cloud in Munich, a one-billion-euro partnership with NVIDIA with "up to 10,000 NVIDIA Blackwell GPUs" (Deutsche Telekom).
• UpCloud (Finland): the outline I started from said UpCloud had no GPUs, and that is out of date. It now sells L4, L40S, RTX PRO 6000 and H100 servers from its Helsinki data centre, with zero egress fees (UpCloud GPU servers).
• Nebius (Amsterdam headquarters, Nasdaq: NBIS): the largest EU GPU specialist by capacity. Its Mäntsälä, Finland site is scaling to 75 MW and "up to 60,000 GPUs" (Nebius). It is a Dutch company with a US stock listing and US regions, so ask it the same jurisdiction questions you would ask anyone else.
• Hetzner (Germany): good value for CPU workloads. Its GEX dedicated servers use RTX PRO Blackwell cards and have no H100 option (Hetzner GPU servers). Fine for inference, small fine-tunes and embeddings. Not a training cluster.
• Exoscale (Akenes SA, Lausanne, part of the A1 Group): RTX Pro 6000 GPUs in Frankfurt, Vienna and Zagreb, but its B300 is in Geneva only (Exoscale GPU). Switzerland relies on an EU adequacy decision and is not an EU member.
• Nscale (Glomfjord, Norway): "30MW operational capacity expandable to 60MW" on hydropower (Nscale). Norway applies GDPR through the EEA Agreement, but any contract clause that says "EU only" excludes it.

Verdict by use case. Training from scratch: Nebius, OVHcloud or Scaleway H100 SXM. Fine-tuning: Scaleway, OVHcloud, StackIT or UpCloud H100. Production inference: Scaleway or OVHcloud L40S, or IONOS AI Model Hub when you want the no-training guarantee in writing. CPU-only RAG: Hetzner or UpCloud.

Why does "EU data residency" usually only mean storage?

Because data at rest is the only layer that is a configuration flag. Picking a storage region is a dropdown. GPU capacity, control planes, telemetry pipelines and 24/7 support rotations are global by design and almost never pinned in the contract.

Five layers have to stay in the EU, not one: data at rest, compute execution, the control plane, telemetry, and human access. GPU scarcity is the structural reason training leaves the EU. Capacity follows the accelerators, and when the Frankfurt queue is full, someone launches the job in Virginia.

The leak nobody audits is the API call. An app hosted in Frankfurt that calls a US-hosted LLM endpoint ships every prompt, document and fine-tuning sample to the US. The second leak is telemetry. Datadog lists US1 first and makes EU1 (Germany) a separate site with its own URL (Datadog). Sentry makes you pick US or EU (Frankfurt) when you create the org, and you cannot change it later (Sentry).

Jurisdiction is the layer engineers skip. The US CLOUD Act (18 U.S.C. § 2713) lets US authorities compel a provider under US jurisdiction to disclose data in its "possession, custody, or control", wherever that data sits. Schrems II (CJEU C-311/18) struck down Privacy Shield in 2020. Its replacement, the EU-US Data Privacy Framework, was adopted as Decision (EU) 2023/1795 in July 2023.

Then there is human access. EDPB Recommendations 01/2020 state that "remote access by an entity from a third country to data located in the EEA is also considered a transfer" (EDPB, footnote 23). A support engineer in another country opening a shell on your Frankfurt node is a GDPR Chapter V transfer. You are the controller, the cloud is your processor, and the sub-processor annex lists everyone who can touch the data.

The AI-specific wrinkle: when the source data is personal, the training set, the fine-tuned weights, the embeddings and the vector index are all in scope.

Do AWS, Google Cloud and Azure EU regions count as keeping AI compute in Europe?

On geography, yes: the GPUs run in Europe. Legally, not fully, because the operating entity stays US-controlled unless you buy a separately operated offer: AWS European Sovereign Cloud, Microsoft's Delos Cloud (an SAP subsidiary), or T-Systems Sovereign Cloud and S3NS for Google Cloud. One catch most teams find late: AWS's price list shows no on-demand P5 (H100) instances in Frankfurt or Paris. In the EU, the single-H100 p5.4xlarge is offered in Stockholm and p5en in Spain (AWS EC2 pricing).

• AWS European Sovereign Cloud reached general availability on 14 January 2026 in Brandenburg, Germany (eusc-de-east-1). It is backed by €7.8 billion, run through "dedicated European legal entities established under German law", and "operated exclusively by EU residents located in the EU". Bedrock and SageMaker were available at launch (AWS).
• Delos Cloud is a Microsoft-technology cloud run by an SAP subsidiary for German public administration (Microsoft). Check service availability with Delos before you plan AI workloads there.
• Google Cloud has T-Systems Sovereign Cloud in Germany and S3NS in France. S3NS is a Thales and Google Cloud alliance with three data centres in France (S3NS). Its PREMI3NS offer received SecNumCloud 3.2 qualification from ANSSI, announced on 17 December 2025 (S3NS).

My honest split: for most private-sector teams, a hyperscaler EU region plus SCCs, EU telemetry endpoints and contractual controls is enough. For buyers bound by SecNumCloud or BSI C5, it is not. Sovereign offers cost you features: a smaller catalogue, later service launches, fewer foundation models in-region, and higher unit prices.

The rules are moving. The EU Data Act (Regulation (EU) 2023/2854) has applied since 12 September 2025 and abolishes switching charges from 12 January 2027. GPAI obligations under the AI Act have applied since 2 August 2025, with enforcement from 2 August 2026. After the Digital Omnibus, Annex III high-risk rules apply from 2 December 2027 (AI Act Service Desk). EUCS is still not adopted (ENISA).

How do I verify that a provider really keeps AI training inside the EU?

Treat every residency claim as unverified until you hold five artefacts: the DPA sub-processor annex, a contractual region-pinning guarantee, the key-management model, the written non-EU support access policy, and a written exclusion of your data from model training. Everything else is marketing.

Paste this into your next vendor email:

• Send your current sub-processor list with the country of incorporation for each entity, and mark the ones that touch compute, storage or support.
• Confirm in the contract that training jobs, inference endpoints and backups run only in the EU regions we select, with no automatic failover outside the EU.
• Tell us where your inference endpoints terminate, and whether any model call is proxied to a third-party API outside the EU.
• Describe your break-glass process: can staff outside the EU access production, who approves it, and can we export the session logs?
• State the key-management model (provider-managed, BYOK or HYOK) and which legal entity could be compelled to hand over keys.
• Confirm in writing that our prompts, training data and outputs are never used to train or improve any model.
• List the certifications that cover the exact services we use (SecNumCloud, BSI C5), with scope and expiry dates.

SecNumCloud (ANSSI) and C5 (BSI) touch sovereignty. ISO 27001 and SOC 2 say nothing about residency.

Then enforce it technically instead of trusting it. Set region allow-lists through AWS SCPs, GCP organization policies or Azure Policy, add Terraform guardrails, and deny VPC egress by default with an allow-list of EU endpoints. Keep the evidence: audit logs showing the region of every training job, Kubernetes node labels such as topology.kubernetes.io/region, and signed provider attestations.

What does it cost to run AI workloads on EU providers versus a hyperscaler?

On published list prices checked on 28 October 2026, EU providers are cheaper per H100-hour and much cheaper on egress. The gap shrinks once you price the managed services you would have to rebuild yourself. One H100 costs €2.87/h on Scaleway (PCIe), €2.80/h on OVHcloud and €2.17/h on UpCloud. AWS charges $7.36/h for p5.4xlarge in Stockholm and Azure charges $9.07/h for NC40ads_H100_v5 in Germany West Central.

A worked month, with every assumption stated: 200 H100 GPU-hours of fine-tuning, one L40S serving inference 24/7 (730 hours) and 5,000 GB of egress. Storage is left out because it is close on both. On Scaleway that comes to €573 + €1,073 + €0, about €1,646. On AWS Stockholm it comes to $1,472 (p5.4xlarge) + $1,441 (g6e.xlarge at $1.974/h) + $441 egress, about $3,354.

Now the hidden cost. In my experience, a self-hosted vector database, a model registry and an autoscaled inference path take a few engineer-weeks to build and an ongoing slice of someone's time to run. If you would use Bedrock or Vertex heavily, include that in the comparison. And because committed-use discounts and Savings Plans only pay off when the account is yours, a platform that runs in your account keeps the discounts. A platform that resells compute keeps them for itself.

The hybrid split often wins: train on an EU GPU specialist, then serve inference from your primary cloud's EU region, close to your data.

How do I keep AI workloads EU-only across several clouds without a platform team?

Run the deployment layer inside your own cloud accounts. When the platform orchestrates but your account holds the compute, the region pin, the keys, the egress rules and the bill never leave your control. That is also the only setup an auditor can verify from your own logs.

BYOC in two lines: the control plane calls your cloud account's API, and the data plane runs entirely on your infrastructure. Qovery is an internal developer platform that works this way. It deploys into your own Scaleway, AWS, GCP or Azure account, or into an existing Kubernetes cluster of any distribution, self-managed or on-prem. Scaleway is the EU-native provider we support natively, alongside the others.

That settles the residency question, because you choose the region, the KMS keys, the VPC, the subnets and the egress rules. No part of the EU-only decision is delegated to us. The features that matter here are git-push deploys, preview environments per pull request, auto-stop for non-production environments (useful when the idle environment holds a GPU), managed cluster upgrades, per-environment RBAC, and databases backed by your cloud's managed services.

To be fair to the alternatives: Heroku, Vercel and Render run workloads on their own infrastructure. Some offer European regions, but the compute does not live in your account and the operator is a US company. That is a reasonable trade for speed, and it does not meet a strict EU-compute requirement. Self-managed Kubernetes gives you the same residency control as BYOC, and it costs you a platform team. Raw Terraform gives you control with no developer self-service layer at all.

The multi-cloud setup I see work: one EU GPU provider such as Scaleway for training, your primary cloud's EU region for serving, and the same deployment workflow on both.

How do I set this up with Qovery from Claude Code or Cursor?

Install the Qovery Agent Skills once, then ask your coding agent in plain language to deploy into the EU region of your own cloud account. The agent proposes a plan and you approve each step before anything runs.

Install the skills globally for every compatible agent (Claude Code, Cursor, Codex, OpenCode, Gemini CLI):

curl -fsSL https://skill.qovery.com/install.sh | bash

Or scope the install to the current project only:

curl -fsSL https://skill.qovery.com/install.sh | bash -s -- --project

To let the agent inspect what already exists, add the Qovery MCP Server to Claude Code over OAuth. It is read-only by default:

claude mcp add --transport http qovery https://mcp.qovery.com/mcp --callback-port 4242

Add write access only when you want it. This also requires enabling write access in the Qovery Console:

claude mcp add --transport http qovery "https://mcp.qovery.com/mcp?read_write=true" --callback-port 4242

Then ask: "deploy this inference service to my Scaleway Paris account with Qovery." The available skills are qovery-onboard, qovery-deploy, qovery-troubleshoot, qovery-optimize, qovery-speedup, qovery-preview and qovery-terraform. The full walkthrough is in the AI agent quickstart and on the Claude Code integration page.