EU sovereign cloud & data residency
Where can you run these models on European-hosted, GDPR-compliant infrastructure — and is that even possible for the models that matter?
The catch: only a handful of open models are SOTA-competitive on coding
Of the open-weight models, only GLM 5.1+, Kimi K2.6+, DeepSeek V4 Pro, MiniMax M2.7+ (and arguably Xiaomi MiMo-V2.5-Pro) are genuinely competitive with the leading US closed labs (Claude Opus, GPT-5.x) on coding. Everything else an EU host typically offers — gpt-oss-120b, Llama 3.x, Mistral, Qwen3 — is a clear step below on the coding benchmarks tracked in this app. So the relevant question isn't “is there an EU inference provider?” (there are many) but “is there an EU provider that hosts the SOTA models?”
✅ EU-hosted and company-approved equivalent offers for SOTA models
| Model | EU-hosted / approved-equivalent offers (10:1 blended $/1M) |
|---|---|
| GLM 5.2 | Inceptron $0.900Inceptron/OpenRouter $0.900Mistral $1.67Mistral/OpenRouter $1.67TensorX $1.77Scaleway $2.49T-Systems LLM Hub $5.09 |
| GLM 5.1 | Nebius $1.67TensorX $1.67 |
| Kimi K2.6 | Inceptron $0.817Inceptron/OpenRouter $0.817TensorX $1.27 |
| Kimi K2.7 Coding | Inceptron $0.918Inceptron/OpenRouter $0.918Azure AI Foundry $1.23EU equivalentTensorX $1.55 |
| DeepSeek V4 Pro | Azure AI Foundry $1.90EU equivalentTensorX $1.91 |
| MiniMax M3 | TensorX $0.545 |
| Xiaomi MiMo-V2.5-Pro | no EU-hosted or approved-equivalent route within the active global filters — self-host the open weights in an EU region |
7 model families within the active global model/provider filters. Active catalog listings without a public per-token rate remain visible as “price not public”.
TensorX (Ireland; 3 EU data-centre regions, 100% EU-sovereign / isolated from US hyperscalers, zero data retention) is now the broadest EU-sovereign option — a self-serve per-token API carrying GLM 5.2/5.1/5, Kimi K2.7 Code/K2.6/K2.5, DeepSeek V4 Pro/Flash/V3.2, MiniMax M3/M2.5 and Qwen, all in-EU. Inceptron (Swedish HQ, Finnish datacenter; zero-retention) serves GLM 5.2, Kimi K2.6/K2.7 Code and MiniMax M2.5 from the EU. Scaleway now also serves GLM 5.2 from Paris. Nebius (Netherlands; Finland/France; ZDR + no-training) lists GLM 5.1/5.2, Kimi K2.6/K2.7 Code, DeepSeek V4 Pro and MiniMax M2.5/M3 — but serves several of them from US/UK regions, so its only SOTA model currently running in an EU region is GLM 5.1. The table normally lists a provider only where that specific model runs in-EU — a provider being EU-capable in general isn't enough. There are exactly two deliberate company-policy exceptions: the native Azure Direct Globaloffers for DeepSeek V4 Pro and Kimi K2.7 Code are included as company-approved EU-hosted equivalents. This is a legal/business classification for this application, not a technical EU-residency guarantee; those Global deployments may process inference outside the EU. ⚠️ Azure's Fireworks-hosted alternatives, including GLM/MiniMax, remain US-served and excluded from the EU Data Boundary. Thanks to TensorX, Kimi K2.7 Code and MiniMax M3 also have a genuinely EU-hosted managed route; NextBit additionally serves DeepSeek V4 Flash from Spain. MiniMax M2.7 and Xiaomi MiMo-V2.5-Pro still have no managed EU route.
🟡 Coming soon
TrustedRouter (api.trustedrouter.eu) — an EU-based, security-focused router. Still being launched / not production-ready yet; tracked here so it shows up once it goes live.
EU-sovereign per-token providers and their catalogs
These sovereign clouds expose proper per-token APIs with strong residency and certifications. Most catalogs center on Western open models (gpt-oss / Llama / Mistral / Qwen), while Scaleway now also hosts GLM 5.2in Paris. Their current model listings:
| Provider | Residency / notes | Models offered | Listing |
|---|---|---|---|
| STACKIT AI Model Serving Schwarz Group (DE) | BSI C5, ISO 27001 | gpt-oss-120b/20b, Qwen3-VL-235B, Llama 3.3 70B, Gemma 3 | models ↗ |
| T-Systems LLMHub (Telekom) Deutsche Telekom (DE) | sovereign T-Cloud · €1,000/mo min | gpt-oss-120b, Llama 3.3, Mistral, Qwen3 | models ↗ |
| IONOS AI Model Hub IONOS (DE) | BSI C5, Gaia-X | gpt-oss-120b, Llama 3.1/3.3, Mistral, Qwen3-Coder-Next-80B | models ↗ |
| OVHcloud AI Endpoints OVHcloud (FR) | SecNumCloud/ANSSI, ZDR | gpt-oss-120b, Llama 3.3, Qwen3.x, Mistral | models ↗ |
| Scaleway Generative APIs Scaleway (FR) | GDPR, ZDR by default | GLM 5.2, Qwen3.x, Mistral, Llama, Gemma, gpt-oss-120b, Holo2 | models ↗ |
| SAP Generative AI Hub SAP (DE) | enterprise · opaque CU billing | mostly proprietary frontier; open weights deprecated/BYOM | models ↗ |
⛔ Proprietary-only & GPU-rental EU providers (not in the comparison)
These EU vendors appear in residency trackers but are excluded from the price comparison: they either serve only their own proprietary models that aren't in any public benchmark (e.g. Aleph Alpha's Pharia, IBM Granite), or they only rent GPUs/VMs with no managed per-token API.
| Provider | Where | Why it's not listed |
|---|---|---|
| Aleph Alpha | DE | Own proprietary Pharia models only — not in public benchmarks; sovereign on-prem/Gaia-X focus. |
| IBM watsonx | US/EU regions | Pushes own Granite models; not benchmark-competitive for SOTA coding. |
| GPU / dedicated rental only | Hetzner, OVHcloud (bare GPU), Open Telekom Cloud, Exoscale, Scaleway GPU, IONOS GPU, CoreWeave EU, Aruba, elastx, Genesis, gridscale, Seeweb, UpCloud, Cloudiax | Rent GPUs / VMs and self-host any model — no managed per-token API, so no comparable price. |
| US-law platforms (EU region only) | Hugging Face, Together AI | EU data residency only via enterprise/dedicated; operating company under US law. |
📊 For a continuously-maintained map of which LLMs actually run on EU soil (and which providers are under the US CLOUD Act), see the community EU LLM Hosting Tracker (llm-tracker.eu) ↗.
Hyperscaler EU regions & confidentiality
- • AWS Bedrock (EU Geo profile) — strongest turnkey EU posture: EU cross-region inference + zero-data-retention by default. Western open models only (Llama/Mistral/DeepSeek V3.2/Nova/gpt-oss).
- • Azure AI Foundry EU Data Zone — documented EU routes remain the technical residency standard. Separately, this company treats only the native Azure Direct Global offers for DeepSeek V4 Pro and Kimi K2.7 Code as EU-hosted equivalents for filtering; inference may still occur outside the EU. Fireworks-hosted alternatives remain outside the EU boundary.
- • Google Vertex AI europe-west / EU multi-region — Model-Garden open models (Llama/Gemma/DeepSeek V3.2); never use the global endpoint.
- • Confidentiality (TEE) for GLM/Kimi/MiniMax: Chutes (Intel TDX + NVIDIA CC, 13 TEE models) gives confidentiality but no EU pinning; Privatemode (Edgeless Systems, Germany) is EU-pinned + confidential but narrow.
- • “EU available” but dedicated/enterprise only — not self-serve serverless (so excluded as price sources): Together AI (Sweden GPU clusters / dedicated endpoints, Sep 2025), Fireworks (Frankfurt/Iceland dedicated/BYOC), Cloudflare (Data Localization Suite, Enterprise add-on), AtlasCloud (eu-west dedicated). Their public per-token APIs have no EU region selector and run from US capacity. DigitalOcean has EU serverless on its roadmap only. All are US-HQ (CLOUD Act).
Use the global “EU-hosted / approved equivalent only” and “Non-US provider only” filters (top bar) to restrict the interactive comparison views and linked model details to matching offers. Full research with sources lives in data/research/ in the repository.