AI Data Residency (training/inference, prompts/outputs in EU)
Why important?
AI sovereignty (SOV-3 "Data & AI") requires EU control over all AI-related data flows.
How measured?
Scale 0–5 + N/A:
- 0 = AI data (training/inference data, prompts, outputs) not EU-based or unclear.
- 1 = EU residency only partial or announced; AI telemetry/logging non-EU.
- 2 = Inference data EU-based, but training/fine-tuning data or logs non-EU or unclear.
- 3 = Training and inference data EU-based; no non-EU replication of prompts/outputs.
- 4 = Broad EU residency incl. prompts/outputs; few documented exceptions.
- 5 = Full EU residency incl. training, inference, prompts, outputs and AI telemetry + secured contractually/technically.
- N/A = no reliable evidence
Sources / Evidence
Validation questions (RFP)
- Where exactly are prompts, outputs, context data and AI telemetry processed and stored, for each model offered and also during load peaks and failover? Are inputs used for training or fine-tuning, and is the exclusion contractually agreed? For how long, and where, are prompts retained for abuse monitoring or troubleshooting?
Scores comparison
| Provider | Score | |
|---|---|---|
| AWS European Sovereign Cloud | 4.0 | GA: Amazon Bedrock, SageMaker and Amazon Q are available in the AWS ESC (operated in the EU); according to AWS, Bedrock content is not used to train the foundation models, inference/training data remains in the EU region. (evidenced) |
| IONOS Cloud | 4.0 | IONOS AI Model Hub: managed inference, 100% in German data centres; prompts/outputs in the EU; not used for model retraining; geofencing. (stated) |
| STACKIT | 4.0 | STACKIT AI Model Serving: managed inference in DE/AT data centres; data/requests neither stored nor used for training. (stated) |
| OVHcloud Public Cloud (inkl. SecNumCloud) | 4.0 | OVHcloud AI Endpoints: serverless inference, data entirely in EU data centres (Gravelines); not used for training; GDPR. (stated) |
| pluscloud open | 4.0 | Operated in GDPR-compliant, BSI C5-attested data centres in Germany; no data outflow from the private plusKI instance. (evidenced) |
| Exoscale | 4.0 | Fully sovereign European hosting; prompts and responses remain in Europe. (evidenced) |
| Scaleway | 4.0 | Scaleway Generative APIs / Managed Inference: 100% in the EU (Paris); stateless, no storage/no training; prompts/outputs only in the EU. (stated) |
| noris Sovereign Cloud | 4.0 | Operated in noris high-security data centres in Germany. (evidenced) |
| Cloud Temple Trusted Cloud | 4.0 | Inference on SecNumCloud-qualified, HDS-certified infrastructure in France. (evidenced) |
| Infomaniak Public Cloud | 4.0 | Infomaniak AI (Euria + AI services for developers): processing/storage exclusively in CH data centres, no data transfer abroad, no training on customer data, ephemeral mode. (stated) |
| T Cloud Public | 3.0 | Open models hosted in German data centres; proprietary models run via third-party platforms outside the EU. (evidenced) |
| Oracle EU Sovereign Cloud | 3.0 | OCI Generative AI available in the EU sovereign realm (region eu-frankfurt-2, since Sep 2025); EU-operated, isolated realm (200+ services); model catalogue Cohere (proprietary) and Meta Llama; EU residency for inference. (evidenced) |
| Delos Cloud | 3.0 | Azure OpenAI within the sovereign boundary (DE); additionally 'OpenAI for Germany' (SAP/OpenAI/Microsoft, rollout 2026) on Delos Cloud. Processing within the DE boundary. |
| Microsoft Sovereign Cloud | 3.0 | Azure OpenAI / AI Foundry; EU Data Boundary; Azure OpenAI data remains in the region, according to Microsoft not used for training. (stated) |
| UpCloud | N/A | No managed model serving; GPU servers for self-hosting, managed Kubernetes announced. Criterion not applicable. (N/A) |
| Hetzner Cloud | N/A | No managed AI or inference offering; GPU servers for self-hosting only. Criterion not applicable to this offering. (N/A) |
| SysEleven OpenStack Cloud | N/A | No managed model serving; AI via self-hosting on MetaKube / OpenStack GPU. Criterion not applicable. (N/A) |