THE SIGNAL IN ONE SENTENCE
The Netherlands has chosen a surprisingly grown-up way to talk about AI sovereignty. It is not promising to manufacture every chip, own every cloud and train every model behind a very tall national fence. That would be expensive theater with excellent catering. The cabinet's new national AI vision uses a more selective idea called open strategic autonomy. The basic question is not whether every technology is Dutch. It is whether the country and Europe retain enough knowledge, capacity and access to make choices when a supplier changes the price, a government changes the rules, a service fails or a crisis arrives. The cabinet approved the vision on November 6. It organizes the work around three pillars: responsible AI development and use, access to advanced AI and safety, and public direction of the transition. It names healthcare, energy, food production, mobility and logistics, security and defense, and the technology industry as areas where the Netherlands should focus. The most useful part is buried inside the sovereignty language. The government describes an AI technology stack that runs from critical materials, machinery and semiconductors through digital infrastructure, compute, cloud, data, software, models and applications. At each layer, the country should decide what capacity must exist in the Netherlands or Europe, what can come from trusted partners and which concentrated dependencies should be reduced or diversified. That is not independence. It is an inventory with consequences. The vision also points toward a Groningen AI factory, more joint public procurement, fewer barriers for smaller companies, wider reuse of successful government applications and a possible Dutch AI safety institute. Detailed letters about AI infrastructure, computing power and data centers are promised for early 2027. Citizen participation is planned for 2027 and 2028. Those future dates matter. This is a cabinet vision, not a funded implementation plan. It does not yet tell a hospital exactly which cloud contract it should replace, how much national compute will cost, which dependencies count as unacceptable or how fast a public service must recover after a foreign supplier becomes unavailable. So here is the practical signal: publish the dependency map. For every important AI service, show the chips and machines underneath it, the data-center location, the cloud operator, the model supplier, the data rights, the administrator accounts, the encryption keys, the software licenses, the update channels, the people with operating knowledge, the energy and water demands, and the legal jurisdictions that can interrupt access. Then test the exit. Export the data. Move the model or application. Rotate the keys. Rebuild an account. Run the service in a degraded mode. Recover from a supplier outage. Give a person a way to challenge a decision. Sovereignty is not a sticker on the server rack. It is knowing which parts can stop the service, who can restart them and how long the public will be waiting while they try.
01
WHAT ACTUALLY CHANGED
The Dutch cabinet approved a national AI vision built around selective capability rather than complete technological independence. The government calls the approach open strategic autonomy. It wants the Netherlands to choose its own course from within a strong European Union, focus on areas where Dutch institutions can make a difference, preserve people's control over technology, use AI to strengthen human work and let society help direct the transition. That framing is more precise than a general promise to become sovereign. It admits that a small, highly connected economy will keep relying on international suppliers. The policy question becomes which reliance is tolerable, which needs a trusted partnership and which can place an essential service at unacceptable risk.
The vision treats AI as a stack of dependencies. The government document traces the stack across critical materials, production machinery, semiconductors, digital infrastructure, computing power, cloud services, data, software, models and applications. It proposes deciding at each layer what Dutch or European capacity is necessary and where access through trusted partners is sufficient. This matters because the visible model is often not the fragile part. A public agency may be able to change a model but remain trapped by proprietary data formats, identity systems, administrator access, a cloud contract, a chip shortage, a licensing server or a team whose knowledge left with a contractor.
The cabinet promises more detailed infrastructure decisions in early 2027. The official announcement says a broad data-center approach and a detailed plan for AI infrastructure and computing power will follow. The vision acknowledges that compute uses energy, water and physical space. The government's public explainer points to the Groningen AI factory and says waste heat should be reused for homes. Those are material choices, not footnotes. A compute plan competes for grid capacity, land, water, public money and technical workers. The later documents will need to show who receives scarce capacity, what public benefit is expected and what happens when environmental promises miss their targets.
The vision links sovereignty to public institutions and human control. The cabinet wants government organizations to share successful applications, procure more jointly and make adoption easier for smaller firms. It is studying a Dutch AI safety institute, with advice expected before a decision. It also plans citizen participation in 2027 and 2028. The government says people should know when AI plays an important role in a decision and have somewhere to turn when something goes wrong. That converts abstract autonomy into a useful test: can a resident understand the service, challenge its result and still receive help if the automated part is unavailable?
02
WHY THIS MATTERS
For hospitals, municipalities, utilities and security services, dependency is a continuity problem. An essential service can meet every ordinary performance target and still be brittle. A foreign court order, revoked account, expired license, failed update, cloud-region outage or missing specialist can interrupt the system. Mapping those dependencies makes it possible to decide which ones need redundancy, local capacity, contractual protection or a manual fallback.
For smaller companies, selective sovereignty could widen access without pretending every firm can build its own infrastructure. Shared European compute, portable standards and clearer public procurement could give local companies a fairer route into health, energy and government markets. Joint buying can also create one enormous lock-in if every agency adopts the same closed platform. Procurement should include export tests, exit prices, transition support and proof that another qualified operator can take over.
For workers, the cabinet's language about human agency deserves a measurable version. AI should not be judged only by how many tasks it automates. Ask whether workers retain the knowledge to spot an error, stop the system, explain a decision and keep the service running. A tool that removes every experienced reviewer may look efficient until the first unusual case turns up and nobody remembers where the override lives.
For residents, the dependency map should include rights as well as hardware. Who can see the data? Where is it stored? Which law governs access? Can a person correct it? Can an agency explain a result after the model or vendor changes? Is there a staffed appeal route? Technical portability without evidentiary continuity can leave a public body able to move the software but unable to defend the decisions it already made.
For Europe and trusted partners, the Dutch approach could become a more honest model of cooperation. No European country controls the entire AI stack. Shared capacity can reduce concentrated risk, but trusted should not become a diplomatic adjective that excuses missing evidence. Partnerships need criteria, reciprocal obligations, incident reporting, continuity commitments and an exit plan. Trust is stronger when it can survive an audit.
03
WHERE IT COULD HELP
- Name the essential service first. Record who relies on it, the harm caused by interruption and the longest acceptable outage before discussing models or vendors.
- Trace the full stack. List critical materials, machines, chips, data centers, networks, cloud regions, identity systems, data stores, software, models, applications and support teams.
- Record ownership and jurisdiction. Show who operates each layer, which legal system can compel action and where administrator access, encryption keys and backups reside.
- Classify each dependency. Mark what must be Dutch or European, what can come from a trusted partner and what should be diversified across independent suppliers. Explain the reason for every classification.
- Test data portability. Export a complete, usable copy with metadata, permissions, audit history and documentation. Time the process and confirm that another system can read it.
- Test model and provider switching. Run a representative workload with an alternative model or operator and record accuracy, latency, cost, safety behavior and the engineering work required.
- Protect operating knowledge. Maintain current runbooks, named deputies, recovery credentials and training for public staff rather than leaving essential knowledge inside one contractor's chat history.
- Publish an infrastructure ledger. Report computing capacity, energy source, grid constraints, water use, heat reuse, emissions and the public services receiving priority access.
- Write exit terms into procurement. Include export formats, transition assistance, deletion evidence, capped exit charges, audit access and a deadline for returning keys and administrator control.
- Rehearse failure with people in the room. Simulate a cloud outage, revoked account, bad update, compromised model and unavailable vendor. Keep the service running in a reduced mode, record recovery time and fix the gaps.
KEEP A HAND ON THE WHEEL
This is a cabinet vision. The official material does not yet provide a funded delivery schedule for every promise, a complete list of unacceptable dependencies or outcome measures for each focus sector. Watch the early 2027 infrastructure, computing and data-center documents for budgets, owners, deadlines and public reporting. Open strategic autonomy is deliberately selective. That makes the classification rules crucial. Watch for clear criteria explaining why one layer requires Dutch or European capacity while another can rely on a trusted partner. Without criteria, selective can become a polite word for whatever the current supplier already sells. The Groningen AI factory is a concrete infrastructure project, not proof that the Netherlands controls the full stack. Compute still depends on chips, networking, software, energy, water, skilled operators and supply chains. Watch for access rules, capacity allocation, costs, environmental reporting and tested continuity arrangements. The cabinet is studying a possible AI safety institute. A study is not an institution, and an institution is not automatically independent or effective. Watch for its legal authority, budget, technical access, public reporting, complaint route and relationship with European regulators. Citizen participation is planned for 2027 and 2028. Watch who can participate, which decisions remain open, whether people are paid for their time, how disability and language access are handled and whether the government publishes what changed because of the process. Joint procurement can reduce duplication and improve bargaining power. It can also spread one bad dependency across the entire public sector. Watch for multiple qualified suppliers, open interfaces, portable evidence, exit tests and recovery drills before contracts are renewed. Finally, infrastructure efficiency claims need measured results. Data centers consume energy, water and land even when waste heat is reused. Watch for public baselines, local grid effects, actual heat customers, seasonal usefulness, water accounting and a plan for equipment that becomes obsolete faster than the building around it.
04
TERMS WORTH KEEPING
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on November 7, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US