While a German mid-sized company is still drafting the policy that bans ChatGPT, its American competitor ships its third AI feature. This is not an anecdote. This is the state of play. And the reason carries a noble name: sovereignty.<\/p>\n\n\n\n
Sovereignty is the most popular word in European boardrooms right now, and that alone should make us suspicious. Two crowds gather behind it: those who stoke fear of every American technology, and those who call for the state to sort it out. Both mean protection. Protection from competition, from speed, from their own lag. But sovereignty is not a shield you put on. It is the spoils of the one who fought and won. And winning today means using the best AI, not avoiding it.<\/p>\n\n\n\n
So take this as a call to courage, addressed to investors, founders and managers. Wanting change comes hard to us in Germany, Switzerland and Europe. Americans and Chinese find it easier, and that is exactly our disadvantage right now. The rest of this piece first takes apart the fears that are used to sell standstill, then makes the case for going to the edge, because that is the only place where you learn and win.<\/p>\n\n\n\n
On 10 June 2025, Anton Carniaux, Director of Public and Legal Affairs at Microsoft France, testified before the French Senate<\/a>, under oath. The question was simple: could he guarantee that French citizens’ data on Microsoft’s European servers would never reach US authorities without the French government’s consent? His answer was honest. “No, I cannot guarantee that.” The reason: the US CLOUD Act binds American companies wherever the data physically sits. He added that such access had never yet happened.<\/p>\n\n\n\n That single sentence has become the founding text of Europe’s cloud and AI policy. It shows up in every hearing, every strategy paper, every justification for a new procurement rule. Rightly so, because it is true. Microsoft’s own counsel confirmed under oath what privacy advocates have said for years.<\/p>\n\n\n\n The mistake is not in the sentence. It is in the lesson the preservers draw from it: therefore, no American AI. A correct statement about jurisdiction turns into an operational reflex that has nothing to do with what was actually said. And that leap is where the whole fear industry hangs its hat.<\/p>\n\n\n\n Here is the uncomfortable observation that appears in no sovereignty speech. If a contract sits on SharePoint and gets analyzed in Databricks, and Databricks runs on Azure or AWS, then that processing has been on US-controlled infrastructure for years. Reachable under the same CLOUD Act, signed off long ago by the very companies that now fear AI. Databricks is not a warehouse. It computes. It processes. The clean split you hear everywhere, that storage is harmless and processing is the new and dangerous case, does not survive the first question.<\/p>\n\n\n\n Whoever now bans AI because it processes would, to be consistent, also have to switch off Databricks, cancel Azure and rip out half the company’s IT. Nobody does that, and rightly so, because it would knock out every European company overnight. But then we should be honest: the AI ban is not the result of a risk assessment. It is selective.<\/p>\n\n\n\n That selectivity runs through every one of the fears below. AI is held to a standard no one ever applied to the cloud beneath it. Not because AI is more dangerous, but because it is new, visible, and because the word sovereignty is cheap to invoke right now. The dependency everyone warns about did not arrive with the language model. It arrived with SharePoint, and we signed for it years ago.<\/p>\n\n\n\n Take the arguments that come up in every compliance meeting and look at them one at a time.<\/p>\n\n\n\n The first fear: the US government reads our data. Two entirely different things get confused here all the time. The CLOUD Act is a law enforcement tool. It needs a court order and a specific target. FISA 702 is an intelligence tool, bulk collection of non-US persons, with no individual order. Microsoft’s own transparency report is clear: business customers are almost never targeted, individual consumer accounts regularly are. The fear is calibrated to the consumer and wrongly carried over to the company. I am not saying the risk is zero. You are not told, the provider may not even be allowed to tell you, and no contract removes the reach entirely as long as the provider is American. But a low probability plus a structural residual exposure is something other than the operational certainty implied in the room.<\/p>\n\n\n\n The second fear: the GDPR forbids it anyway. Simply outdated. The EU-US Data Privacy Framework<\/a> has been in force since 10 July 2023 and, for certified US firms, replaces the old standard contractual clauses. In 2025 it survived a challenge before the EU’s General Court. Standard contractual clauses plus a transfer impact assessment remain as a fallback. So there is a legal basis. Forbidden is false. The situation is fragile all the same, a third Schrems case is in preparation, and the FISA 702 intelligence power was up for renewal. But that fragility is an argument for the structural work, not against using the tools.<\/p>\n\n\n\n The third fear: the AI trains on our data. That was genuinely new, but it applied to the free tools and to the shadow IT where employees paste contracts into an open chat window. At the enterprise tier it is handled by contract. Azure OpenAI, Amazon Bedrock, Claude for enterprise, all with zero retention and no training on customer data. The danger was uncontrolled use, not the language model.<\/p>\n\n\n\n The fourth fear: more US AI makes us more dependent. The jurisdiction risk is not the marginal case that arrives with AI, it already sits in the stack. What does arrive is deeper lock-in, because every added layer raises switching costs. That is real, and it is serious. But the answer is not to ban the top layer, it is forced replaceability. More on that shortly.<\/p>\n\n\n\n The fifth fear: we have to make everything sovereign now. Taken to its end, that would mean ripping out Azure, AWS and SharePoint along with it. Nobody does. And even where the sovereign alternative already stands, today it is more expensive and thinner. The AWS European Sovereign Cloud<\/a> went live on 15 January 2026, a separate German legal entity, operated entirely in the EU, but it costs about 15 percent more, offers roughly 90 services instead of more than 240, has no GPU instances and is missing most AI models. Sovereign right now means fewer services, no compute for AI, higher cost. That is the measurable price of refusal, and the one already behind pays it first.<\/p>\n\n\n\n Five fears, one shared pattern. AI is held to a standard no one ever applied to the cloud beneath it.<\/p>\n\n\n\n This is not about pretending none of it matters and closing our eyes. The skeptics can hold on for one more round.<\/p>\n\n\n\n Two things are genuinely new. The first is training on your own inputs, and at the enterprise tier that is settled by contract. The second is behavior. A language model invites people to type the crown jewels into an input box, faster and more sensitively than any software before it. That is a matter for access control and data loss prevention, not for jurisdiction. The pipe is the same as the cloud beneath, only more confidential material now flows through it, faster. It belongs governed, internally, with clear guardrails, not banned.<\/p>\n\n\n\n One point deserves precision, precisely because it gets used as a conversation stopper. For German professionals bound to secrecy, lawyers, tax advisers, auditors, section 203 of the criminal code lets them bring in outside providers without committing an offense. But only on one condition: the provider must be bound to confidentiality in writing and expressly informed of the criminal consequences. A data processing agreement under the GDPR does not cover this, it handles data protection, not the criminal duty of secrecy. Both must be in place. The German Federal Chamber of Tax Advisers published its own AI guidance<\/a> in February 2026, and a plain ChatGPT account does not meet the bar, while a properly contracted enterprise tier does. The same pattern with confidentiality agreements: an old secrecy clause can be breached if a counterparty’s confidential data runs through a third party’s AI tool without consent. Modern clauses require exactly that prior consent and the disclosure of the tools used. None of this is an argument against AI. It is an argument for the right contract, not for going without. Settle it once, cleanly, then work.<\/p>\n\n\n\n What remains is the structural residual: as long as the provider is American, a residual reach via FISA and the CLOUD Act stays, low probability, opaque, not fully negotiable away. That is exactly why the political answer, building your own European legal entities and sovereign infrastructure, is right. But that residual does not justify banning AI while the entire American stack beneath it keeps running. You manage this risk, you do not ban it away. The difference between the two is the difference between an entrepreneur and a worrier.<\/p>\n\n\n\n Now the part closest to my heart, and the only one where I genuinely push. Whoever does not drive the newest AI to the edge today falls behind. Not a little, but structurally, and for a reason most underrate: competence compounds.<\/p>\n\n\n\n You do not learn AI competence by reading, by strategy papers, or by waiting for the safe, mature, European tool. You learn it by using it, on real cases, at the edge of what the tools can do today. A company that trains two hundred people on real agent workflows today builds a lead a competitor cannot buy later, because that knowledge is distributed, practical, and sits in people’s heads. And the gap grows every month, because the tools get better every month and the learning curve compounds. Whoever waits two years does not start the curve two years later, they start it against someone who has been climbing it for two years.<\/p>\n\n\n\n Here is the observation that has to stand as the uncomfortable truth: not deciding is also deciding. Whoever waits has chosen to fall behind, they just do not notice, because caution feels responsible. The cautious reflex is the most expensive position in the room, and it wears the face of seriousness.<\/p>\n\n\n\n And so no one mistakes this for recklessness: the bold move is the safe one here. The risk of using AI is manageable today, by contract and by architecture, as shown above. The risk of falling behind is not manageable, in the end it is final. Courage here is not bravado, it is the sober calculation. There is also a quieter point everyone who competes for the best people knows: talent follows the best tools. Whoever refuses drives away the very people they will need tomorrow, and notices only once they are gone.<\/p>\n\n\n\n Look at the reality in most German and Swiss companies, and the AI policies point the wrong way. They mostly try to stop employees from trying the newest tools, by forbidding them. That is hostile to innovation, and at heart it is the IT department’s fear of losing even more control. AI has to leave the IT tower. It has to go into use, it has to be tried.<\/p>\n\n\n\n The future of AI is not the employee typing sensitive data into a chat window. The future is agentic workflows, built by ordinary employees with skill who never learned to code and now do it anyway. That future must not be blocked. Enabling it is the job of entrepreneurs and investors, not of a policy that smothers every attempt in the cradle.<\/p>\n\n\n\n Now the how, because courage without architecture is just noise. You do not have to be sovereign at the model. You have to be sovereign at the layer that holds the knowledge: data, orchestration, memory. The model is a consumer. It can be swapped when the layer beneath it stands. The orchestration, which steps a workflow takes and which tool gets called when, is yours. The memory, the gathered context of documents, decisions, histories, is yours. Whoever controls those two layers can swap the model on top without losing their knowledge. That is the later sovereignty, operational rather than declared. It is the spoils of the bold who build today.<\/p>\n\n\n\n A concrete case, the way it already works. A mid-sized machine builder, a few hundred people. They want better and faster quotes, better and more efficient service for their customers, they want to be faster, better and cheaper. Instead of chaining themselves to one vendor, they build the architecture in layers. The documents and their index sit on European infrastructure. The workflows run through an orchestration they control themselves. The memory, the semantic index over the company’s documents, sits in a database they run themselves. And on top, for the actual language understanding, the system calls the best available model today, an American one included, because today it is simply the best. The decisive point: that top piece is the only one not owned in house, and it is exactly the piece that can be swapped with one line of configuration. The company learns at the frontier today and keeps the choice tomorrow.<\/p>\n\n\n\n And here is the real point. Whether the base infrastructure beneath is still American today or not matters less than the debate suggests. The new value sits in the orchestration, where AI meets governance. That knowledge does not belong locked inside American tools. It belongs to you.<\/p>\n\n\n\n One honest caveat, so this does not sound too smooth. Swapping later is a bet that European models will close the gap to the frontier, and that gap is real. So replaceability has to be built into the architecture from day one, not as a good intention for later. Whoever applies hard today but without that architecture wins the first round and loses the second. The two belong together.<\/p>\n\n\n\n I am no champion of the sovereignty debate that waves Mistral and Cohere around. Cohere is Canadian, even after the combined company took in Aleph Alpha. Yes, we have DeepL and Black Forest Labs, and both are good. But against the range and the speed of the American players, that is still far too little. And above all, the name-dropping clouds the view. It makes us look richer than we are.<\/p>\n\n\n\n The sober reality: StackIT from the Schwarz Group, IONOS, OVHcloud, Scaleway all offer only a fraction of the services of AWS, Google or Microsoft. Too little to build your own stack on at high speed and quality. The application layer around Microsoft 365 or the Google suite, plus mature databases and developer services, is almost entirely missing. Whoever demands in the sovereignty debate that we use our mediocre tools now makes us slow and widens the gap. That is the opposite of what they claim to want.<\/p>\n\n\n\n But in one place we have serious players, and it is exactly the place that counts now for the new AI world: the new orchestration and knowledge layer. n8n from Berlin, one of Europe’s most valuable AI firms, is exactly the automation and workflow layer I mean. deepset, also Berlin, builds an orchestration framework for agents with the open Haystack and calls itself, not by accident, a sovereign AI platform<\/a>. Langfuse from Berlin delivers the observation and control layer, where AI meets governance, the very point from before. ZenML from Munich orchestrates the pipelines beneath. Cognee builds the memory for AI systems, Qdrant the vector database, both from Germany, Weaviate from the Netherlands. A warning shot stands next to them: Jina AI from Berlin was strong and got bought by America’s Elastic in autumn 2025. That is what happens when we do not keep our winners.<\/p>\n\n\n\n The honest part: at the frontier models the gap to the American labs is huge. That does not lead to resignation, but to a clear order. First, whoever tries to build their future on second-rate models today loses. So we take the best on top, the American one included. Second, we learn along the way that the frontier models are too expensive for many use cases, and deliberately fall back to simpler, cheaper, often open models where they suffice. Third, we choose the architecture so we can swap the infrastructure beneath later too, once there are serious European alternatives. Only those three steps together are real work on sovereignty. For the AI itself, though, one rule holds without exception: work at the limit, with the newest there is. Here not a single one of the fearmongers’ arguments counts.<\/p>\n\n\n\nThe dependency arrived with SharePoint<\/h2>\n\n\n\n
Five fears, examined one by one<\/h2>\n\n\n\n
What is genuinely new<\/h2>\n\n\n\n
To the edge, or backward<\/h2>\n\n\n\n
The model is replaceable, your knowledge is not<\/h2>\n\n\n\n
Europe’s layer, measured honestly<\/h2>\n\n\n\n
Beside the EU, not behind it<\/h2>\n\n\n\n