Sovereign operations
Data residency, supply chain and access control are decisions. We think they belong to the operator — and we think the only credible way to say that is to run it, every day, ourselves.
The stance
Useful AI does not require handing your data to a platform. Models can run where you choose, behind boundaries you set, on infrastructure you can audit, switch off, and move. That is what sovereignty means in practice — not a slogan, three properties: auditable, stoppable, portable.
The cost of those properties is engineering. Someone has to run the edge, the mesh, the identity, the inference — and keep it boring and reliable. That is the work we do for ourselves first, then for you. We don't resell anyone's managed layer and call it sovereignty.
Our reference architecture
Own edge
Dedicated servers at European providers, our own TLS termination, firewall posture reviewed to a written policy. Admin surfaces are never public.
Private mesh
A self-operated WireGuard-based control plane joins every machine we run. Management reaches infrastructure only through the mesh — zero public SSH.
Passkey SSO
A self-hosted identity provider issues OIDC to every admin surface. Phishable passwords: gone. The keys live on our devices, not in a vendor's cloud.
Self-hosted inference
Models run on our own GPUs where the workload allows it — including a 27-billion-parameter model serving from a single 8 GB card.
Measured, not claimed
35.1 tok/s
— decode speed of a 27B-parameter model on a single 8 GB GPU, quantised and KV-compressed, measured on our bench and recorded in our engineering journals.
1 gateway
— in production. Our AI gateway routes model traffic with budgets, permissions and PII redaction, and has been running our own workloads daily.
0 third-party
— requests. That is what this page makes. No analytics, no web fonts fetched from others, no trackers. View source; it holds.
The argument, written down
We put the supply-chain case in writing — why the default agent frameworks are structurally wrong for sovereign environments: Why LangChain and Composio are unsuitable for sovereign AI inference clouds