Sovereign AI 🚀: Control Model Weights—Now14 min read2,770 words

Sovereign AI 🚀: Control Model Weights—Now

Learn sovereign AI and AI model weights—how AI data sovereignty works, when open-source AI models fit, and assess value with business process intelligence.

sovereign AIAI model weightsAI data sovereigntyopen-source AI modelsbusiness process intelligence
Sovereign AI 🚀: Control Model Weights—Now

Sovereign AI Is About Business Control, Not Going It Alone

AI sovereignty is often treated like a geopolitical question - should a country rely on models built elsewhere? For businesses, though, the day-to-day question is simpler and more urgent: where does sensitive data go, who controls the model that touches it, and who captures the value when AI changes how work gets done.

Those questions came up in a September 2026 TNW Podcast interview with Celonis co-CEO Alexander Rinke. His argument is that Europe should build more local AI companies - not because foreign providers are inherently untrustworthy, but because the region should take part in the economic upside. For companies and public agencies, he made it more concrete: retaining control over data, and in some cases, over the model weights themselves.

And that distinction matters. Owning a model isn't automatically safer, cheaper, or more effective than using a hosted frontier model. The right approach depends on the task, the data, the operational environment, and the organisation's ability to run and maintain the system. This article breaks down what sovereign AI can mean in practice, when model weights actually matter, how process data and specialised models fit in, and how leaders can pressure-test an AI investment before scaling.

Why does Europe need local AI champions?

Rinke's case for local AI companies is primarily economic, not defensive. In the interview, he argued that Europe needs to build companies that can participate in the next wave of AI value. The issue isn't just where a model was trained. It's whether European businesses, workers, and institutions will help build the products and infrastructure that shape the market - or mostly pay to use what others built.

That's different from saying every country needs a fully domestically built AI stack. Recreating everything - from chips and data centres to foundation models and applications - would be expensive and, in many cases, unrealistic. Local capability can still matter in specific places: building models for regional needs, creating products that understand local industries, and giving customers real choices about where and how their systems run.

Participation is not the same as isolation

A local provider isn't automatically better than an international one. They may have less capacity, a smaller product, or a less mature support operation. A global platform, meanwhile, can be very capable and still leave customers with limited control over deployment, updates, or data handling.

So the useful question isn't whether a provider is local. It's whether your organisation has options and leverage. If a critical workflow depends on a single supplier, switching costs can creep up even when the product works well. Having credible alternatives can strengthen negotiations on price, deployment terms, and access to your own information.

Celonis is an example Rinke used to explain how European companies can operate in a global market. The company has headquarters in New York and Munich, and it also has offices across Europe and beyond. Rinke's point was that building "from Europe" doesn't require operating only in Europe.

A practical test for local capability

For a business or public agency, assess local AI capability by looking at what it enables - not treating geography as a stand-in for quality.

  • Can it meet a real local need? This could include data residency, language support, sector-specific rules, or a deployment model that fits public-sector infrastructure.
  • Does it increase supplier choice? A second viable provider reduces dependence only if the product can meet real operational requirements.
  • Can the organisation influence how the system evolves? Support, deployment options, and a clear roadmap may matter more than the provider's address.
  • Is the capability sustainable? A model that can't be updated or supported isn't a strategic asset just because it was built locally.

The trade-off is straightforward: local champions can improve choice and keep more value in a region, but they still have to earn adoption through performance, reliability, and fit. Policy support can help an ecosystem grow - but it can't replace software that actually solves a problem.

Key takeaways

  • Sovereignty isn't the same thing as refusing foreign technology.
  • Local providers matter when they create genuinely useful alternatives and let customers participate in the value created.
  • Judge capability by outcomes, support, and control - not by geography alone.

What does owning AI model weights actually mean?

Model weights are the learned parameters that shape how a trained model responds. When an organisation can access those weights, it may have more options for running, adapting, or hosting the model than it would with a closed service that you can use only through a provider's interface.

Rinke's description of sovereign AI emphasized organisations owning their AI - and ultimately, their model weights. He also raised a related concern: companies don't want to hand over all their data to an external model for training. Those issues overlap, but they aren't identical. Keeping data under your control doesn't automatically mean you own the model, and owning weights doesn't by itself control what happens to prompts, logs, or other information you send to a service.

Separate the control questions

Before picking a deployment approach, spell out what you need and why. "We need sovereign AI" is too vague for procurement or architecture decisions.

  1. Data: Where is data stored and processed? Is it used to train or improve a provider's model? How are prompts, outputs, and logs retained?
  2. Model: Can you inspect, select, or adapt the model? Do you have access to its weights, and what does the licence allow?
  3. Deployment: Can the system run in your own environment or in a specified cloud region? Who controls updates and availability?
  4. Operations: Who handles security patches, monitoring, backups, and incident response? What happens if the provider changes terms or stops supporting the model?

This is where the trade-off shows up. Hosted models can reduce the burden of maintaining infrastructure and keeping software current. Self-hosted or more open models can give more control, but they shift work onto the customer or a service partner. This isn't only a privacy vs convenience decision - it's also about cost, technical capacity, and responsibility.

Weights are one part of sovereignty

Weights access can be useful when you need to adapt a model for a narrow task or run it under tight data constraints. But access doesn't mean the model will work for every purpose. Licensing terms, hardware needs, performance, and whether you have people who can run it all matter.

Running locally also doesn't automatically make a system safe or compliant. You still need access controls, testing, processes for sensitive data, and clear accountability when the system gives a bad recommendation. In Europe, data protection obligations like the GDPR apply regardless of whether a model is open, hosted, or self-managed.

A sensible approach is to map each use case to the minimum level of control it requires. If the task uses low-sensitivity information and a managed service meets the requirements, owning weights may be extra cost without meaningful risk reduction. If the task depends on confidential operational data, strict deployment rules, or specialised adaptation, more control may justify the extra effort.

How can process intelligence make AI more useful?

A model can produce a fluent, plausible answer and still do little for your actual business process. The missing piece is usually context: how work really moves through the organisation, which systems are involved, where exceptions happen, and what "success" looks like.

Rinke described Celonis as starting with a visibility problem. The founders wanted models that could simulate how a business works - but they first needed a reliable view of how it works today. Companies couldn't easily provide that picture, so they built what he called an "x-ray machine" to examine the data businesses generate and reveal their processes. Celonis now calls that work process intelligence.

Start with the actual workflow

In practice, your documented process map may show the intended route. Operational data shows the route people and systems actually take. An order gets re-entered. A claim waits for a manual approval. A purchase skips the expected sequence. That gap - between the process on paper and the process in real life - is often where automation projects run into friction.

Process data can help teams spot where AI could add value, but it shouldn't be treated as a command to automate everything. First, figure out whether the delay or exception has meaningful cost. Then ask what's causing it: a repetitive decision, bad data, unclear policy, or a broken handoff. A model won't solve an organisational problem just because you plug it into the workflow.

Match the model to the task

Rinke also discussed Celonis's acquisition of Ikigai Labs, an MIT spin-off focused on specialised models for tabular and time-series data and for what-if simulation. He described these models as tools that a large language model can use - for example, to support supply-chain forecasting or decisions about settling an insurance claim.

The broader lesson is simple: don't assume a general-purpose language model needs to perform every calculation or prediction itself. A language model can help interpret a request, pick a tool, or explain results. A specialised model may be a better fit for structured data, forecasting, or simulation. Evaluate the system as a whole, including how components pass information around and how people review what comes out.

Rinke also said that for specific business tasks, adapted open-source models could match or beat frontier models. That's a claim about particular jobs, not a universal rule that open models always outperform larger systems. The practical takeaway is to test models against the work they'll actually do, using relevant business data and a clear success metric.

A grounded evaluation might include:

  • Task quality: Does the system improve accuracy, speed, or consistency compared with the current process?
  • Failure handling: Can it identify uncertainty, route exceptions, and keep a human in the loop when needed?
  • Data boundaries: Does sensitive information stay inside the approved environment?
  • Total operating cost: What do hosting, integration, monitoring, and maintenance cost over time?
  • Change impact: Does the workflow need to change, and who owns that change?

The goal isn't to add AI to a process just because the tech exists. It's to use process evidence to pick a small set of high-value tasks, then confirm that the full system improves the outcome.

How should companies assess AI products before scaling?

AI makes it easy to produce a convincing demo. Proving that customers get durable value - prove it's hard to replace for good reasons, and prove you can deliver it reliably over time - is much harder. Rinke's approach to product-market fit is a solid starting point, especially in a market where early interest can be mistaken for real demand.

He asks whether the product is valuable, uniquely valuable, repeatable, and likely to keep earning customer payments. Those questions help separate a compelling pilot from something that can sustain a business. They also apply inside an enterprise: a proof of concept isn't enough if each deployment requires a fresh custom build, or if no team is accountable for using the result.

Use four questions before expansion

1. Is it valuable? Identify what outcome the customer actually gets. "Uses AI" isn't an outcome. Reduced processing time, fewer avoidable errors, or better forecasts might be - if you can measure them.

2. Is it uniquely valuable? If the feature is easy to replicate, the product may still be useful, but you need to understand what keeps customers choosing it. That could be workflow integration, data context, reliability, or better fit for a specific type of task.

3. Is it repeatable? If it only works after extensive customer-specific engineering, it's closer to consulting than software. Customisation can make sense commercially, but the cost and delivery limits should be obvious - not hidden.

4. Will customers keep paying? Ongoing use and renewal are stronger evidence than a launch announcement or a one-off pilot. For internal tools, look for continued adoption and a budget owner who will fund ongoing operation.

Check whether the organisation can scale it

Rinke calls the next step product-market-people fit: having the organisation required to scale the product. It's a useful correction to the idea that more demand automatically justifies faster growth. A company can have a viable product and still lack the support, implementation, security, or engineering capacity to serve customers consistently.

This matters even more for AI, because early demos can hide the ongoing work. Models and data pipelines need monitoring. Integrations change. Users need guidance on when to trust outputs and when to escalate. If those responsibilities aren't assigned, a promising pilot can turn into a fragile dependency.

Capital can help a company move faster, as Rinke noted. But fundraising isn't a substitute for customer proof. He described Celonis as bootstrapping for five years and said it was difficult to win early enterprise customers without references. His principle was direct: build for customers, not for the next funding round.

For a team deciding whether to scale, use a simple gate:

  • Evidence: Is there a measurable result from real use, not just a demo?
  • Repeatability: Can the team deliver the result without disproportionate custom work?
  • Ownership: Is someone accountable for operations, risk, and customer outcomes?
  • Economics: Do expected revenue or savings exceed the full cost of deployment and support?
  • Exit options: Can the organisation switch providers, models, or workflows if assumptions change?

If several answers are unclear, the next step is usually a narrower test - not a bigger rollout. That may feel slower, but it prevents teams from scaling a product whose value depends on conditions that won't hold across customers or departments.

Conclusion: choose control where it changes the outcome

Sovereign AI is most useful when it turns a broad fear of dependence into specific decisions about data, models, deployment, and operations. Rinke's case for local champions is ultimately about participation and choice. His discussion of model weights points to the business question underneath it: which parts of an AI system does an organisation need to control to protect its data, serve customers, and capture value?

Start with the workflow and the risk, not the label. Test whether a model improves a real outcome, account for the cost of running it, and make ownership responsibilities explicit. Use more control where it actually changes the result - and avoid paying for control where it doesn't.

Frequently Asked Questions

Sovereign AI is an approach to building or using AI that gives an organisation or jurisdiction greater control over relevant parts of the system. Depending on the use case, that can include data, model choice, deployment location, model weights and operational decisions. It does not necessarily mean building every component locally.
Access to model weights can give a company more options for adapting or hosting a model, especially when it has specialised tasks or strict deployment requirements. It does not automatically provide data protection, compliance or lower costs. Licensing, infrastructure and ongoing maintenance still need to be assessed.
No. Privacy depends on how the system is deployed and operated, including what data is sent, retained or used for training. A self-hosted model can offer more control, but it also creates operational responsibilities. A managed service may meet privacy requirements if its data handling and deployment terms are suitable.
No. Many organisations will get better results from using existing models or services. Building or adapting a model makes sense when a specific task, data constraint or operating requirement justifies the additional cost and responsibility. Compare options against real workloads rather than assuming ownership is inherently better.

Related articles