Integration / Sovereign cloud
STACKIT, routed by jurisdiction.
Route German-hosted inference to STACKIT AI Model Serving in region eu01, the Schwarz Group cloud, with YTT Secunet policy enforcement and per-request residency receipts.
STACKIT is the cloud platform of the Schwarz Group, the German retail group behind Lidl and Kaufland. Its AI Model Serving product hosts a shared catalogue of open-weight chat and embedding models behind an OpenAI-compatible REST API, with chat completions and embeddings endpoints, in region eu01 on STACKIT infrastructure in Germany. STACKIT states that it does not store request data or train models on customer requests, and authentication uses an auth token issued in the STACKIT portal.
For a German Mittelstand manufacturer or a public body that wants inference in Germany from a German-owned provider, STACKIT is a natural primary endpoint. YTT Secunet adds what a single provider cannot: a tenant-level policy that also covers the other models your teams use, PII masking before the request leaves your data plane, and a signed receipt per call for the Article 30 record. Many customers pair STACKIT with a US-parented model in an EU region and let the policy pack decide by sensitivity class: personal data and HR documents to STACKIT, public marketing copy to whatever model is cheapest within the EU.
Since STACKIT's catalogue is shared and models are deprecated as new versions arrive, YTT Secunet's registry polls the model list and warns before an alias points at a retired model. If you need a model STACKIT does not offer, run it on vLLM in your own STACKIT project and register that endpoint as a Self-hosted runtime with facility eu01, so it is governed by the same policy.
| Region or deployment | Location | Jurisdiction |
|---|---|---|
| eu01 (AI Model Serving) | Germany | EU / European Union |
| Self-hosted runtime in a STACKIT project | Customer project, eu01, Germany | EU / European Union |
- 01
Activate AI Model Serving in your STACKIT project and create an auth token for YTT Secunet.
- 02
Review the shared model catalogue and pick the chat and embedding models you need.
- 03
Register the eu01 OpenAI-compatible base URL in YTT Secunet as an EU endpoint with facility eu01 and store the token in the YTT Secunet vault or your KMS.
- 04
Create aliases for the chosen models and attach the endpoint to policy packs, for example as the only allowed target for personal data in eu-gdpr-strict.
- 05
Point your SDK at https://eu.yttsecunet.com/v1 and verify a receipt names facility eu01.
endpoints:
- id: stackit-eu01
provider: openai-compatible
upstream: https://api.openai-compat.model-serving.eu01.onstackit.cloud/v1
jurisdiction: EU
facility: eu01
credentials: vault://stackit/model-serving
routes:
- policy: eu-gdpr-strict
when:
sensitivity: [personal, special-category]
allow: [stackit-eu01]
- policy: eu-gdpr-strict
when:
sensitivity: [public]
allow: [stackit-eu01, aoai-gwc-gpt41]- STACKIT endpoints are tagged German-hosted and EU-headquartered, which a Germany-only policy can require.
- Only chat completions and embeddings are routed to STACKIT; other API surfaces are refused rather than forwarded.
- Deprecated catalogue models trigger an alert seven days before YTT Secunet stops routing to them, configurable per alias.
Is STACKIT available outside Germany for AI Model Serving?
STACKIT documents eu01 in Germany for this service. Check STACKIT's current region list if you need another location.
Can I use STACKIT for embeddings in a RAG pipeline?
Yes. STACKIT exposes an embeddings endpoint, and YTT Secunet applies the same policy and receipts to embedding calls as to chat calls, which matters because embeddings are derived from personal data too.
Does YTT Secunet store my STACKIT token?
It holds it encrypted with AES-256 in the YTT Secunet secret vault, or references it in your own KMS on Enterprise and Sovereign Dedicated. Prompts and completions are never stored.
Sovereign route ready
Stop risking 4% global turnover GDPR fines on AI inference.
Deploy sovereign, cryptographically proven AI routing in 15 minutes.