Sovereign AI

Your data stays in the EU.

For EU organizations whose prompts, documents, outputs, and logs must not leave the Union. We run the model and the workflow in an EU facility you control, or with an EU operator, and leave a system your people operate. On-prem in your own EU facility is how the strictest cases are built.

173 tok/s reference point on about $30K of hardware. Not a client result.

The boundary is the EU

A US cloud region labeled EU is still a US company processing the data. This engagement keeps the processing boundary inside the Union.

GDPR

Personal data is processed inside the EU. The architecture is evidence for your review. We do not give legal opinions.

DORA

The ICT dependency is an EU facility you control, or an EU operator, not a cloud region that fails a substitutability test.

NIS2

Essential and important entities get a deployment whose inference does not depend on a provider outside the Union.

Air-gap is the strict subset: a defense site or a disconnected facility that cannot connect. When the boundary is US healthcare, the same deployment is on the HIPAA data boundary page. We are not a compliance advisory firm.

UAE deployments use the same pattern: the model stays inside the country. UAE data boundary names the IA Regulation buyers still call NESA, Dubai DESC, and the federal PDPL. Those three bind different entities.

How a Sovereign deployment is shaped

Every architectural decision treats the data boundary as a hard constraint.

The boundary is the EU.

Prompts, documents, outputs, and logs stay in the EU. The facility is one you own, or the operator is an EU operator. Air-gap is available when the site itself cannot connect. It is not the default.

No single vendor sets the price.

Open-weights models on hardware you control. Predictable capital cost instead of an API invoice that moves every quarter. One provider cannot cut off access by changing a region.

Performance is a reference point, not a trophy.

173 tokens per second on about $30K of hardware is the reference we size against. We benchmark the actual workload before anyone buys GPUs.

Your people operate it.

Runbooks, versioning, and an owner on the client side. The engagement is designed to end. Ongoing work is a later fixed milestone, not a retainer.

What we deploy and how

A complete sovereign AI stack — from hardware spec to production monitoring.

Layer What we deliver Compliance value
Hardware Spec, procurement support, rack layout, GPU node configuration Physical data residency, no shared multi-tenant risk
Model selection Open-weight model evaluation, benchmarking for your use case, licensing review No proprietary cloud model dependency, model provenance documented
Inference stack Optimized inference server (vLLM, TGI, or custom), API gateway, load balancing No inference call to a provider outside the EU. Air-gap when the site cannot connect.
Data pipeline RAG architecture, vector store, embedding pipeline — all on-premise Regulated data stays inside your network boundary
Access control LDAP/AD integration, RBAC, API key management, audit logging Controls evidence for GDPR, DORA, and NIS2 reviews
Observability Inference logging, model version tracking, usage dashboards, alerting Audit trail of inference events, model change management
Hybrid routing EU facility for the sensitive path, EU operator for burst work that is not the core data The core data still does not leave the EU

Who is blocked by the boundary

Buyer situations, not AIPIVT client results.

Financial services

CISO, EU bank or insurer.

Before
Regulator scrutiny on cross-border flows, and the business still wants AI. Projects are blocked in legal review.
With us
The model and the workflow run inside the EU, with an audit trail.
Outcome
A documented boundary their people can operate.
Book a scoping call
Cost

Head of IT.

Before
The API invoice grew faster than the budget. The cost is a variable monthly bill.
With us
A sized EU deployment for the steady workloads, cloud only where the data is allowed to leave.
Outcome
A cost they can forecast.
Book a scoping call
Domain data

CDO.

Before
Decades of proprietary data that cannot be sent to an external provider. A generic model that misses internal terms.
With us
A model on that corpus, inside the EU, with an eval harness.
Outcome
A system that speaks the firm's language and stays inside the boundary.
Book a scoping call
Air-gap

Disconnected site.

Before
Defense, critical infrastructure, or a facility that cannot connect. No model inside the perimeter. This is the strictest case.
With us
An air-gapped install and an offline update path.
Outcome
Assistance under the site's own rules.
Book a scoping call

Want the compliance architecture overview?

Contact us for a one-pager covering our reference architecture for GDPR, DORA, and NIS2 deployments — including data flow diagrams, access control models, and audit logging specifications.

Who designs and deploys Sovereign AI

A senior practitioner leads every Sovereign engagement. The people below have built platforms for banks, where a data breach is a regulatory event as well as a technical one.

Alexey Kichin — Senior AI Platforms Practitioner. AWS Certified Solutions Architect with 20+ years across Tier-1 banking, FinTech, and crypto. Prior work: six years as Enterprise Architect for Emerging Markets at Deutsche Bank, owning ~100 applications across currency derivatives, fixed income, and core banking. He anchors the AI Architect seat. Full profile
AIPIVT — Sovereign AI delivery. We connect deployment architecture to data sovereignty, operational requirements, and business outcomes. Engagements define the deployment boundary, access controls, integration requirements, and ownership of the operating environment. Production delivery and handover are structured around the constraints of your institution and the capabilities your internal team needs to run the system. About the practice

Meet the full team

Private cloud, inside the EU

Three boundaries, in this order. Your own EU facility. An EU operator. A US cloud region with an EU label. The third is not this offer. GDPR and DORA are what your counsel checks the deployment against. We do not issue a certificate. Air-gap is the strict subset, for a site that cannot connect. A US healthcare boundary is the HIPAA data boundary.

Common questions

Air-gap means inference with no outbound connectivity. It is the strict subset, for a site that cannot connect. The default Sovereign deployment is an EU facility or an EU operator, which may have a network, and which still does not send prompts, documents, outputs, or logs outside the Union. Model updates on an air-gapped site move by a controlled offline transfer.

The reference point is 173 tokens per second on about $30K of hardware. We benchmark your workload before sizing GPUs.

GDPR, DORA, and NIS2. We document data flows, model provenance, access control, and audit logging so your review can use them. We do not give legal opinions or issue certificates. The same deployment, with a different boundary, is on the HIPAA page for US healthcare and on the UAE page, where the IA Regulation buyers call NESA, DESC, and the PDPL each bind different entities.

No. Sizing is part of the assessment. You can buy the hardware through us or on your own, or skip it when an EU operator is the right boundary. The specification comes after we benchmark your workload, not before.

Open-weight models under permissive licenses, such as Llama, Mistral, Mixtral, Qwen, Phi, and Falcon, and enterprise-licensed models whose agreement allows self-hosted inference. The choice depends on accuracy on your tasks, your GPU memory and compute budget, your compliance constraints, and the speed you need. We benchmark candidate models on samples of your real workload before recommending one. A contract-analysis tool and a document summarizer often end up on different models.

We leave runbooks, versioning, and a named owner. We do not sell an open-ended retainer or a retained on-call SLA. Further operation is a later fixed milestone, or an AI Ops seat booked separately.

Yes, and we deploy it. The model runs inside a facility you control or with an EU operator, and we leave a system your people operate. Hardware procurement is part of the engagement, not a prerequisite. A site that cannot connect gets an air-gapped install.

The boundary is the EU. That is your own EU facility, or an EU operator. A US cloud region with an EU label is not this offer. GDPR and DORA are the frameworks the deployment is built to give your counsel evidence for. We do not issue a certificate. Counsel decides scope. Air-gap is available for a site that cannot connect.

Yes. Open-weight models run on your hardware, or a contracted model that is allowed to run inside the boundary. Prompts, documents, outputs, and logs do not go to the OpenAI API. If you already call a provider API and the problem is the bill, that product is TokenGuard, not this engagement.

Ready to deploy AI on your terms?

Tell us the boundary you need, the workflow, and who will operate it afterward. We will scope the hardware, select the model, and show you what a Sovereign deployment looks like for your specific environment.

Free one-page checklist: Your Data Stays in the EU →