Trust & Infrastructure
Security & Sovereignty
Where your workload runs, what we can and cannot see, and what we have not built yet.
Sovereign AI infrastructure
Every machine we run for you sits in an EU region, and your client talks to your GPU directly — nothing of ours sits between them. Your prompts and completions are never written to our systems, because the control plane is not in that path and has nothing to write. What we do hold is your account and billing data, and the subprocessors that touch it are listed by name, all of them in the EU.
Encryption
Our strongest guarantee here is not an encryption one: your prompts and completions are never written to our systems in the first place. Beyond that, this is exactly what is encrypted today and what is not.
- At rest: Our infrastructure provider encrypts the physical disks our volumes sit on and manages the lifecycle and rotation of those disk keys. We do not add volume-level encryption on top — under the provider's shared responsibility model that is ours to do, and we have not done it yet. What lives on those volumes is model weights and each instance's own endpoint credential. Your prompts and completions are not among them: they are never written to our disks at all.
- In transit: Model endpoints are not yet served over TLS. A request to a deployment reaches the machine's own address in the clear, guarded by a per-deployment credential rather than by encryption. The TLS-terminating ingress that replaces this is built and not yet deployed. Internal service-to-service traffic is not separately encrypted.
EU data residency
You choose the region, we name the company running the hardware, and neither changes behind your back.
- Where the GPU is: You pick an EU region when you deploy, and the card stays in it. Today the machines run in France, with our EU infrastructure provider — the subprocessor list says what each provider processes and where.
- What you choose: The region is yours to pick, not ours to assign, and a deployment is never moved out of it.
- Sovereign data path: Your client calls your GPU's own address. We register the hostname and step out, so no machine of ours is in the request path and none of it leaves the EU on our side.
Compliance posture and roadmap
We build to GDPR. We hold no certification: each needs an independent audit, and none is booked.
GDPR posture
We are the processor and you are the controller for anything you run. We do not store, log or monitor your prompts and responses — the control plane never receives them. The Data Processing Agreement sets out the rest in writing.
EU sovereignty levels (CADA)
The Cloud and AI Development Act is a European Commission proposal from June 2026, not yet law. It would let public buyers ask for one of four sovereignty levels. Level 1 asks that data stays in the EU, which every Ratel deployment does. Levels 3 and 4 ask for EU ownership and, generally, EU-citizen staff, and we do not claim them.
EU AI Act transparency
Each build in the catalog publishes its license, its context length, the GPU it runs on and its quantization, so you can name in your own documentation what you are actually running. We do not inspect your models or your inputs.
Related trust and compliance documents
The longer versions of everything above:
Trust Center
The same posture at length, section by section, including what we have not built.
Read posture →Data Processing Agreement
Draft terms governing data processing under GDPR.
View DPA →Subprocessor list
Every company that touches your data, what it does, and where. Draft, under legal review.
View subprocessors →