Souveräne KI · engineered in Munich
Run AI inside your own perimeter.
Your teams are already wiring agents to your analytics, ad platforms and CRM. Every prompt that leaves your network is a regulated event — and a model you don’t host can be switched off by someone else’s government. We build the sovereign stack that keeps your data, your models and your control on your side of the line.
# IN-PLAIN-TERMS
New to this? Here’s the whole idea in three steps.
Today, your data leaves
Every time your team uses a public AI tool, customer data is sent to servers you don’t control — often outside the EU. Most companies don’t even know it’s happening.
The fix: run AI in-house
The AI runs on computers you own, inside your own network. Sensitive data never leaves the building — you get modern AI without shipping your secrets to a US cloud.
Where we come in
This is engineering, not a download. We install it, connect it to your tools safely, prove it’s compliant, and keep it running — so you get the upside without the risk.
# THE-SHIFT
Sovereignty stopped being a compromise.
Two years ago "sovereign AI" meant "the weaker option you take when you can’t get the good models." That’s over. For European buyers it’s now a hard requirement on the shortlist — and the demand is concentrated exactly where you operate.
- 62%
of European organizations are actively seeking sovereign AI — regulated sectors leading, strongest intent in Germany & Switzerland.
// Accenture, 2026
- <5% → 75%
share of EMEA enterprises expected to move workloads onto sovereign infrastructure by 2030. Today it’s under 5%.
// Gartner, Predicts 2026
- 2×
European sovereign-cloud spend, ~$7B to ~$12B in a single year, as data control becomes a budget line, not a debate.
// Gartner IaaS forecast
// And in June 2026 a foreign export directive pulled two frontier models offline worldwide overnight. The risk is no longer hypothetical.
# WHERE-IT-STARTS
Your AI is already leaking. Quietly.
What’s happening now
- Agents connected to your analytics, ad platforms and CRM are sending customer data to external endpoints on every call.
- Staff paste contracts, tickets and personal data into consumer chatbots — shadow AI nobody logged or approved.
- Each prompt with personal data crossing to a non-EU model is a regulated processing event under GDPR.
- Your model provider, and their government, can change terms, raise prices, or cut access without your consent.
- No audit trail. When the EU AI Act asks what your systems did and why, there’s no answer.
What we put in its place
- A drawn perimeter: sensitive inference runs on models you host, on hardware you control.
- A governed gateway that classifies every request, masks personal data, and decides what may cross the line.
- Explicit, minimal, auditable connectors to your tools — regulated data never rides them.
- Open-weight models you keep, so no single vendor or jurisdiction can switch you off.
- A complete log of prompts, routing and outputs — the evidence your DPO and auditors ask for.
# THE-ARCHITECTURE
One perimeter. Three tiers of data.
Not everything needs to run air-gapped — that’s expensive theatre. The engineering discipline is classifying data by sensitivity and routing each request to the right place. You decide the rules once; the gateway enforces them on every call.
Tier 1 · public
Send freely
Marketing copy, public translation, non-sensitive drafting. External or EU-hosted models are fine, under clear data-processing terms.
Tier 2 · internal
Keep in the EU
Internal notes, anonymised documents, business analysis. Routed to EU-resident models with zero retention and full logging.
Tier 3 · regulated
Never leaves
Personal data, health, financial, trade secrets, source IP. Local or air-gapped inference on weights you own. Nothing crosses the boundary.
# SCOPE
What we actually build.
// local inference
On-prem or air-gapped serving of open-weight models — Mistral, Llama, Qwen, Gemma, or German-first Teuken / Aleph Alpha — sized to your hardware and latency needs.
// governance gateway
The router on the boundary: data classification, PII masking, policy-based routing, provider failover, and a full audit log for the AI Act.
// tool connectors (MCP)
Controlled MCP connections to your analytics, ad platforms, CRM, ERP and internal systems — explicit, least-privilege, and observable end to end.
// agents & RAG
Retrieval and agent workflows grounded in your own documents and data, tuned to your domain, running entirely inside the perimeter.
// hardware & deployment
GPU sizing, procurement guidance and EU-colocation options. We spec and stand it up; you own the iron. No reseller markup.
// security & access
mTLS, role-based access, secrets handling, no outbound calls where none are allowed. Air-gapped means air-gapped, not "phones home."
# HOW-TO-WORK-WITH-US
Three engagements. Assess, build, run.
01 · assess
Sovereignty Assessment
fixed price · 2–3 weeks
- Map current AI usage, including shadow AI
- Classify your data into the three tiers
- Target architecture & hardware model
- Compliance gap: GDPR, AI Act, NIS2, DORA
- Costed roadmap you own, vendor or not
02 · build
Perimeter Build
project · from the assessment
- Local / air-gapped model serving
- Governance gateway & audit logging
- MCP connectors to your tools
- Agents & RAG on your data
- Security, access, deployment, handover
03 · run
Managed Sovereign AI
retainer · co-managed
- Model updates & evaluation
- Cost & usage observability
- Ongoing compliance evidence
- Agent tuning as needs change
- Clear boundaries with your IT team
Straight talk
What sovereign does — and doesn’t — mean.
Sovereignty is an architecture, not a slogan. A deployment that quietly calls out for license checks or telemetry isn’t air-gapped, whatever the brochure says. Every serious European stack still runs on NVIDIA silicon — nobody has removed that dependency, and we won’t pretend otherwise; we design around it with honest documentation. And local inference alone doesn’t make you compliant: it’s the classification, the gateway, the access controls and the audit trail that hold up in front of a regulator. We’d rather tell you that on the first call than in the post-mortem. Two senior engineers, no juniors, no hand-waving — the same way we run every engagement.
# QUESTIONS
Before you book.
If it’s air-gapped, how does it use analytics or ad platforms?
It doesn’t — not with sensitive data. Those are cloud tools, so connecting to them is a deliberate crossing of the boundary. The gateway makes each connection explicit, least-privilege and logged, and Tier 3 data never travels on it. Local inference handles what must stay in; controlled connectors handle what’s allowed out. That separation is the whole design.Do we need to buy a rack of GPUs?
Not necessarily. The assessment sizes real hardware against your actual workloads — small models on modest hardware cover a surprising range. We spec it and advise on procurement or EU colocation; you own the hardware. We don’t resell iron or mark it up.Which models can we run?
Open-weight families you can keep and audit — Mistral, Llama, Qwen, Gemma, and German-first options like Teuken (Fraunhofer/OpenGPT-X) or Aleph Alpha for regulated public-sector work. Open weights are what make the sovereignty real: no single vendor or jurisdiction can revoke them.Is this only for large enterprises?
No. It’s built for DACH mid-market — the companies with real regulated data but without a 40-person platform team. That’s precisely the gap a senior, two-person shop is built to close.How does this help with the EU AI Act?
Architectures that keep data local, minimise retention and log every interaction are far easier to prove compliant. The gateway’s audit trail is designed to answer the questions the Act, plus NIS2 and DORA, put to your systems. We’re engineers, not your lawyers — but we build so your compliance team has something to stand on.
Draw the line before your data crosses it.
Start with a fixed-price Sovereignty Assessment. You walk away with a data map, a target architecture and a costed roadmap — yours to keep, whether you build it with us or not.
book-assessment// or write: hello@saloid.com · gräfelfing · de