Digital Dependency: Why Europe Is Asking the Wrong Question On AI Sovereignty

Contents

1. Frontier models solve the wrong problems
2. Open Weight models as an alternative
3. The Cost of Waiting
4. What This Means for Decision-Makers

Frontier models solve the wrong problems

The billion-dollar foundation models from major U.S. providers position themselves as “generalists” in the body of human knowledge. General-purpose everyday use, open-ended conversations, complex coding tasks, and vaguely defined knowledge work—this is precisely where they demonstrate their strength and justify their price and market power.

For the specific use cases in large parts of the predominantly industrial German economy—such as process automation, company-specific expertise, or integration into physical systems and existing software—specialized domain knowledge is what matters. For this class of tasks, there is no need for further optimization of knowledge breadth and eloquence; rather, what is required is an architecture with a high degree of integration that specifically supplies existing, smaller models with the relevant context drawn from domain knowledge and system data and tailors them to their own specific field of expertise.

The capabilities of leading models have evolved in a direction that misses the mark for industrial B2B applications. They demonstrate performance to investors—not in addressing most business problems.

Open Weight Models as an Alternative

For a growing number of typical enterprise applications, open-weight models have therefore long since become a serious alternative to the large frontier models. Their advantage lies not only in their independence from individual model providers and the associated geopolitical interests, but above all in predictable, significantly lower usage costs. With comparable coding benchmarks, the cost difference between individual models can be as much as fiftyfold.

Cost per medium-sized coding agent task: DeepSeek V4 Pro is up to 98× cheaper than premium models^¹

[1] Assumption per task: 600,000 input tokens and 80,000 output tokens. Calculation based on public API list prices, excluding prompt caching, batch discounts, infrastructure, sandbox/CI, and human review costs. Actual costs per successfully completed task also depend on the solution rate, retries, and workflow design.

Sovereignty is therefore a deliberate architectural decision for which new answers can already be found today. Instead of waiting for future European alternatives to ChatGPT & Co. or changes to government regulations, AI decision-makers in this country should begin to actively shape sovereignty within their organizations. Four aspects are crucial in this regard:

• Modular integration instead of vendor lock-in. Models are integrated via interchangeable interfaces, not hard-wired—switching vendors or models becomes a configuration change, not a new development project.
• Proprietary AI infrastructure creates independence. Even open models require computing power—if this is sourced from a single provider, another dependency is created. Those who operate their own GPU capacity or source it through interchangeable infrastructure partners remain independent.
• Own data as the true competitive advantage. German industry possesses process knowledge and data assets that no hyperscaler has. It is this substance—not the size of the model—that determines the added value of an AI solution in operation. And this is where the line between dependency and true data sovereignty lies.
• Domain-specific refinement rather than technical generalization. Value is created when a model is specifically tailored to a particular business use case—a task that no model provider will undertake for a single company.

The Cost of Waiting

Such a shift in thinking still seems to be virtually nonexistent. The current trend is much more reminiscent of the early days of the cloud: companies enter into long-term contracts with individual providers, expecting that this dependence will pay off—and in doing so, trap themselves in a cost trap from which it becomes increasingly uneconomical to exit.

Then as now, two behavioral patterns are emerging, both of which carry high risks. Some companies are taking a wait-and-see approach—hoping for a sovereign, equally high-quality, and equally powerful alternative—and in the meantime are steadily losing their competitive edge. Others, anticipating economies of scale and access to the most powerful generation of models available, rely on individual hyperscalers, thereby exposing themselves to precisely that long-term dependence and those cost risks.

What this means for decision-makers

The question for companies is therefore no longer when sovereign infrastructure will arrive, but rather: Which of their own AI use cases can already be implemented today in a modular way, using freely available models and without creating new dependencies?

• Sort use cases by dependency risk. Very few tasks actually require a state-of-the-art model—many can be solved with smaller, interchangeable models.
• Decide on the architecture before choosing a model. By first building the integration layer in a modular way, you keep your options open regarding model selection—and thus retain your bargaining power.
• Don’t wait for the next generation of models. Their added value will largely lie outside your own use cases.
• Be aware of your own strengths. The real competitive advantage lies not in the model, but in access to the company’s own process knowledge. Models are interchangeable; company knowledge is not. Those who recognize this negotiate from a stronger position than the “sovereignty” debate has suggested so far.

How sovereign is your architecture?

Models are interchangeable, but your expertise is not. Talk to us about how you can set up your AI use cases in a modular way, independent of any single vendor.