AI sovereignty is not a romantic ideal—it is a survival strategy

Why Germany will not win through larger models, but through smarter architecture. And why Dieter Schwarz and his Schwarz.digits are right.

Manager Magazin recently described it with brutal clarity: Dieter Schwarz is investing billions in a sovereign European cloud infrastructure — and is being mocked for it. Eleven billion euros in Lübbenau. A digital campus for one billion euros. STACKIT and Schwarz Digits as an answer to AWS, Azure, and Google Cloud. And the criticism is: Amazon AWS generates the same revenue in twelve hours that STACKIT generates in an entire year. Better to leave it to the professionals.

I say the opposite: This criticism is precisely the best proof that Schwarz is on the right path. And I am not saying this merely as an observer — I am saying it as a partner. neuland.ai AG works with STACKIT and Schwarz Digits because we are convinced: This infrastructure is the foundation on which Europe can build its AI sovereignty. What STACKIT provides as a sovereign cloud platform is complemented by the neuland.ai HUB as the application and orchestration layer — together, this creates a complete, data-sovereign enterprise AI platform that is ready for use today.

When hyperscalers start sending lobbyists, distributing dossiers, and influencing procurement procedures — as documented by the Manager Magazin article — then this is not a sign of weakness in the European initiative. It is a sign of strength. Companies that pose no threat are not fought. If Schwarz Digits’ strategy succeeds, Microsoft, Google, and Amazon will lose billions in European cloud revenues. That is the reason for the resistance. Not the technical superiority of the hyperscalers. Economic interest.

Dieter Schwarz should not give in. He should keep going.

The wrong race

There is one sentence that comes up again and again in the German AI debate: “That is just technology nationalism.” I consider this sentence dangerous. Not because it sounds wrong, but because it is so convenient.

The strategic mistake Germany is currently making: We are discussing which model is the largest, who is investing the next billion in GPU clusters. We are not discussing the truly decisive question: Whose knowledge is embedded in the AI that powers our companies and our public administration? The answer is sobering. Usually: everyone else’s knowledge.

The language model is not the intelligence. It is the tool. The intelligence lies in the context — in the data, the processes, the knowledge that an organization has built up over decades. An LLM knows everything about the world and nothing about your company.

The result is a silent homogenization. Accenture, PwC, EY, and Deloitte all draw on the same narrow pool of two to three foundation models. MIT Sloan Management Review put it bluntly in 2025: “Far from being a source of differentiation, artificial intelligence will be a source of homogenization.” Harvard Business Review warns of the “Agentic Convergence Trap”: Companies that run the same models on the same market data arrive at identical decisions. Companies pay premium consulting fees for recommendations that are barely distinguishable algorithmically. Consulting is expensive. Differentiation is gone.

Bigger is not smarter — architecture is smarter

The good news: Germany does not have to lose this race. But it must stop running the wrong one.

The cost of achieving a certain level of AI capability has fallen by a factor of 1,000 in recent years. Whoever builds the largest model pays exponentially more — only Google, Microsoft, Meta, and OpenAI can do that. Whoever uses AI most intelligently pays less and less — anyone can do that with the right architecture.

Gartner predicted in April 2025: By 2027, organizations will use small, task-specific AI models at a ratio of at least 3:1 compared with general-purpose models. The future is intelligent combination: Large Language Models for exploratory tasks, Small Language Models for repetitive and compliance-critical processes, ontologies and knowledge graphs as structured corporate memory, task graphs as encoded traces of work — and an orchestrator that holds all of this together. This is not theory. This is architecture. And this architecture needs an operating system.

AI needs an operating system — STACKIT provides the infrastructure, neuland.ai the intelligence

Many companies today have access to powerful models — but no operating system that integrates these models into their processes, protects their data, respects their roles and permissions, and ensures their compliance.

The neuland.ai HUB is this operating system. It runs on STACKIT’s sovereign cloud infrastructure — and precisely this combination is at the core of our partnership with Schwarz Digits. STACKIT provides the data-sovereign infrastructure operated in German data centers. The neuland.ai HUB places the orchestration layer on top: The LLM is not the knowledge repository, but the curator. The knowledge resides in structured, semantically indexed, permission-controlled knowledge databases. Assistants are persistently stored AI roles with defined task profiles. Workflows are encoded task graphs that make successful solution paths reproducible.

The result: input-token reduction by a factor of 50 to 100, query costs reduced by a factor of 10 — not by sacrificing quality, but through architecture. Data-sovereign, without a single line of corporate knowledge flowing into the training of American models.

What STACKIT builds as infrastructure needs precisely this application layer: the 320 preconfigured AI applications for public administration, the knowledge graphs, the governance layers, the agentic AI workflows. This is the gap that Aleph Alpha was unable to close. Together with Schwarz Digits and STACKIT, we are closing it — with European solutions that are available today.

France is acting. When will Germany act?

While Germany debates, France is systematically displacing Microsoft from millions of Office workplaces in public administration. Not out of technology nationalism. Out of strategic reason: Whoever controls the software controls the data. Whoever controls the data controls the knowledge. Whoever controls the knowledge controls the decisions.

In Germany, 77 percent of municipalities consider AI useful — but only 13 percent have introduced corresponding tools. The municipal record deficit in 2024 amounts to 24.8 billion euros. 570,000 positions in the public sector are vacant, and 1.39 million employees will retire over the next ten years due to age. The technology that could cushion this crisis is available — on STACKIT, with the neuland.ai HUB, today. What is missing is the political will that France has already mustered.

We pay twice — and do not notice

When a German company rolls out Microsoft Copilot, it pays license fees. That is the visible part. The invisible part is more expensive.

Since April 2026, Microsoft has activated “Flex Routing” by default for all EU/EFTA Copilot customers — without explicit consent. When European data centers are at capacity, prompts, emails, and documents are forwarded for processing to servers in the United States, Canada, or Australia. Data is stored in Europe — but no longer necessarily processed there. And processing is the relevant act under data protection law. The EDPS precedent is clear: The European Commission itself violated the GDPR through its use of Microsoft 365. This is not a configuration problem. This is a structural problem.

But this is about more than data protection. Every query to an American model is a data trail. Every uploaded document, every analyzed email, every summarized contract: proprietary corporate knowledge flowing into the infrastructure of an American corporation. We pay with money. And we pay with knowledge. Knowledge is the more valuable of the two.

The decision that must be made now

The world is dividing into two camps: companies that buy AI as a commodity and receive the same thing as their competitors — fast, efficient, but neither sovereign nor differentiated. And organizations that build AI as strategic infrastructure, combine models according to task and compliance, and retain control over their data.

Germany has a choice. But the window of opportunity is closing.

Those who invest in sovereign AI infrastructure today — in STACKIT as the foundation, in the neuland.ai HUB as the orchestration layer — build a knowledge advantage that, over time, will be harder to copy than access to a general-purpose model. Those who wait will continue to finance the prosperity of the U.S. hyperscalers — and will additionally pay with the most valuable asset a company has: its knowledge.

The question is not whether your company will use AI. The question is whose AI it will be. And whose knowledge it will contain.

Image generated with the neuland.ai HUB.