⚡ The Signal

Europe's AI-sovereignty debate is best understood as a continuity test.

In Today's Issue:

🎙️ PandaOS on the operational test for AI sovereignty

🇪🇺 Why a European postcode does not remove US jurisdiction

🧪 What companies can realistically own, switch and self-host

Dear Readers,

In June, Anthropic showed how quickly model access can become a policy decision. The company released Fable 5 and Mythos 5 on 9 June; three days later, a US Commerce Department export-control directive forced it to suspend both models for all users because it could not verify nationality in real time. The controls were lifted on 30 June, and Fable 5 returned globally on 1 July; Mythos 5 remained limited to approved organisations. The interruption lasted less than three weeks, but every company that had made an American frontier model load-bearing saw the dependency clearly.

Europe's political reflex arrived within weeks. Austria's state secretary for digitalisation, Alexander Pröll, wrote to EU tech commissioner Henna Virkkunen at the end of June proposing that the bloc explore "the strategic establishment and participation of Anthropic within the European Union": legal certainty, market access, capital, shared values. It is a revealing answer, because it treats an access problem as a real-estate problem.

That gap is where this week's guests build. The easy version of sovereignty is a data centre with a European postcode. The hard version is a question you can put a stopwatch on: can you change models, keys and providers without your operations breaking?

Marco Szeidenleder is a managing partner at Pandata, the Berlin data consultancy he has helped run for a decade, and a co-founder of PandaOS, the local-first AI workstation Pandata spun out six months ago. Philipp Türker co-founded PandaOS and works on it full time. They are worth listening to precisely because they refuse to sell the maximal version of their own story: not the chips, not the models, and for 99% of companies not on-premise either. What is left, they argue, is your data and your ability to walk away.

All the best,

Kim Isenberg

PandaOS is six months old. Pandata, the Berlin consultancy that spun it out, is ten years old and has spent that decade inside German companies arguing about their data. The product is a desktop-first AI workstation that connects an editor, terminal, browser, databases, deployment tools and an inbox into one workspace, runs on the user's own model keys, and does not train on the user's content. The first version was built by Riccardo Destratis, the third co-founder, after a conversation with Marco about what a computer should feel like once a model sits in the middle of it.

We put the harder questions to Marco and Philipp: what the Anthropic shutoff actually proved, what stays American inside a "European" stack, whether an orchestration layer increases sovereignty or merely decorates it, and what they would refuse to claim their own product can do. The conversation has been edited for length and clarity. Answers are marked Marco and Philipp.

Founders deploying AI agents across sales, marketing, and customer service are closing more with smaller teams. 65% more sales leads. Less headcount.

Download the free Practical Guide to Agentic GTM for Startups and start building the stack your competitors dream of.

Get the Free Playbook

Before we get into PandaOS: the Anthropic moment, and especially the Austrian Anthropic debate. What did that moment actually reveal about Europe's real dependency on US AI?

Philipp: It is quite remarkable, to say the least. The market has shifted, the way we work has shifted, and we are genuinely using AI to speed processes up. But we are also realising that dependencies are becoming dangerous in certain situations. I am not sure you could call it a proper crisis. What we saw was a proper warning shot. The question is not so much which model we are using, but how much access we have to models depending on the situation. Continuity is the question, far more than outage. Are we prepared for multiple scenarios? How do we route our setup so that we do not have a problem if something goes dark, for whatever reason?

Marco: If you are stuck with one country's provider and you have basically all your AI, for all your processes, running off that, and then suddenly that provider is not able to give you a very powerful model anymore, you are obviously in big trouble. The question is much less how to get this provider's best model, and much more how you mitigate the risk of being bound to one provider at all.

If you had to explain the AI sovereignty problem to a European founder or CTO in one sentence, what would you say?

Philipp: The issue is basically redundancy. Every CTO looking at AI today is looking for ways to build up their capabilities in a robust manner, because you do not want to build just for today and then have to rebuild tomorrow because the situation has changed, whether that is a price hike, the availability of a model or its capacity. We are convinced that model usage itself is commoditising. What was true for an Opus six months ago is also true for an open-source model today. So what a CTO has to look out for is whether whatever they are building can be future-proofed for whatever comes their way: access has changed, prices have hiked or dropped, compute capacity has changed. Building a resilient infrastructure for your workflow is the job today.

How would you define the word sovereignty? Some people listening know exactly what we mean, others will hear something completely different. For everyone from a CTO down to someone who goes to work every day and does not know a lot about code: does it mean ownership, control, or whether you can walk away if you want?

Marco: It is a combination, but it boils down to owning the systems that are absolutely crucial for you, and being in control of those systems. It does not mean you have to own it down to the tiniest part. It means you want a certain independence with those systems. Whether that allows you to walk away is more a symptom of sovereignty than the thing itself. It is much more about being in control of, and owning, the systems that are essential for you.

There is a line in your pitch: Europe often wants American AI with a European postcode. Where is that critique fair, and where is it too simplistic?

Philipp: It is obviously a bold claim, but it is a very true sentiment when you come at it from the sovereignty perspective. Not too long ago an Austrian politician wanted a movement to set up an Anthropic entity in Europe, to mitigate the shutoff. As a first step that sounds plausible. But in the end, without wanting to sound too obtrusive, it is a bit of window dressing. As long as you are not completely covered by the legal framework of a given jurisdiction, you are not really independent and not truly sovereign. You can get a US entity set up with a European postcode, but the CLOUD Act is still going to be able to intrude into that environment. It will never completely de-risk you.

And we have seen that movement happening. Last year Microsoft had an exchange with the French parliament and on the record could not guarantee that data stored on Microsoft servers in Europe would not be touched by US jurisdictions. That is very telling. Even with the moves that have happened, and AWS did set up a European entity, Airbus decided with its feet a couple of days ago and moved away from AWS's environment to a French competitor. It shows that even if you bring a European postcode to a US entity, it will not satisfy proper sovereignty in the eyes of the customer.

But there is a secondary element, which is figuring out the right balance for you and your use case. Not everybody is using AI with incredibly sensitive data and needs to gate everything. That becomes a false dichotomy too. The interesting question is how you balance full autonomy, full transparency and full accessibility against speed and commercial viability. Those two pull against each other, and we are seeing more and more companies become sensitised to it.

When you talk about sovereignty you are talking about the EU and the nations inside it. But individual operators make decisions to stay alive, and in a constantly shifting environment your competitors are not bounded by geography. How are the rules being defined, inside your company and around you?

Philipp: What we are looking at is not just the use of AI as building technology, but how you consume it properly. There is a dichotomy here too, because the point is not "we need independence from the US or anywhere else in the world". We should be able to cherish whatever is the most interesting positioning in the market. What is necessary is to understand how data is being used in the end, for the purposes you collected it for. That is where the tension builds.

So there are two aspects. One is European autonomy: do you build out European capacity, which brings in foundational models, compute, how much investment there is, and what kind of environment you curate for entrepreneurs like us who are trying to get a foothold. The second is that once a company has established it needs AI, where does it consume that AI? Does it run over a US model, a homegrown one? Do you do your own inference? Does it make sense to bring it on premise? And on the EU level, the important bit is that we move away from villainising regulation. Regulation is not a problem per se, it is a set of rules we want to abide by, and we are proud of the data-protection side of things. The question is what we need to do to keep that upheld in an age when inference can be bought abroad, and then how you integrate those systems into the European environment.

What are you seeing in your client relationships? Cycle times are shorter, leadership teams keep having to decide strategy and tactics, and who is even in the room: legal, the CEO, the CTO? You have to move fast, but also behave within norms. How are those norms changing?

Marco: At least from our consulting practice, looking around German companies, it is an absolute mess. People do not know, everyone is afraid of breaking something, and in the end everyone just defaults to Microsoft Copilot, because at some point someone decided Microsoft are the good guys and everyone else is evil. I am very sceptical that this is how we should be doing it.

The companies that excel with AI, in my opinion, are the ones that build small tech teams that are allowed to innovate and are not scared to try things. I had a conversation with a rather small publishing house and I was surprised by how well they do it. They have a good small tech team that iterates over different tools and tests them for the company, and if one proves successful they roll it out. They told me on the call: we want a workshop for this tool, but we do not know yet whether it will stick. We will try this one for now, and if in half a year we decide it is not the one, that is also fine.

On the other hand, I just gave a workshop for a very large German corporate, and what they were allowed to use was Copilot with half its features curtailed, because IT had deemed them too dangerous. Another large corporate, a multi-billion company: first day of the workshop, they run out of credits. And I am thinking, how is that even possible? That is not how we are going to win this. Sorry for the rambling, but you can tell this is an emotional topic. I am worried.

When a company tells you its stack is sovereign because the data sits in Europe, what dependencies are still sitting there that they are not counting? The model, the licence, the updates, the chips?

Marco: The chips we are not going to replace anytime soon, and I would not advise any company to even try. The same goes for the model, in my opinion. Just because you are a logistics company does not mean you have to build your own engine. What you need to know is how to operate your trucks in the best way possible. That is where I would advise companies to become sovereign and more independent, and it is usually not a matter of building an engine from scratch, but of making sure nobody can switch off your engine midway.

Where do European options like Mistral, Aleph Alpha's Pharia AI or Teuken-7B already make practical sense, and where do they still fall short in a real workflow, especially next to the latest open-source releases from China?

Philipp: It is difficult, because you cannot lump them together. The European offering is not equal across the board. There is evolution coming in, but I think the question is still the wrong one. It is not about what Mistral can do today, because as you said yourself the Chinese models have picked up, and maybe Mistral has triple the capacity tomorrow and Teuken is better again the day after. The question is how you set yourself up so that you do not lock in today's choices forever.

Where Mistral and Pharia and Teuken and the others have a good story is that they conform to what we as Europeans consider sovereign, because worst comes to worst we can default back to them. That does not make them the default for every use case. Aleph Alpha has a great on-premise story, but on-premise is not the default for every user. To be truly sovereign, in your operations or on a macro scale, is optionality: the capacity to switch without everything breaking. The true litmus test is whether you are capable of switching your inference on a random Tuesday without it breaking your operations and without it affecting your client base. That is the test, and everything else follows from it. As for the European models: today they are one foot behind, tomorrow they might be one foot ahead.

Is Europe realistically capable of catching up?

Philipp: Yes, if it creates the conditions for investment to become substantial enough to compete.

Marco: Commercial models still lead, but open-source models can close the gap quickly. The developments we are seeing with the Chinese models are a chance for Europe: as open-source models become more powerful, European models can catch up more easily.

Capital and talent also move faster than policy. What would make Europe easier to build from?

Philipp: Taxes matter, but so does the freedom to explore, fail and try again. The US and Singapore have built that environment better. Europe still operates as a fragmented market across taxation, company structures and fundraising. EU Inc could make cross-border cap tables, capital and operations easier, but it is still pending and untested.

Self-hosting sounds clean in theory. What does it require inside a normal company?

Marco: Hardware, maintenance and people who know how to run it. Bare-metal self-hosting is expensive and feels like a step backwards for most companies. A European infrastructure provider can reduce the burden, but even that needs operational capacity. Smaller organisations are often better served by a reputable hosted model with credible data-security controls.

Philipp: 99% of the companies we deal with today do not need on-premise setups. The exceptions involve extreme IP sensitivity, government contracts or enough scale to absorb compute costs. Edge inference may become cheaper, and PandaOS can connect to models wherever they run, but on-premise is not the default.

You built PandaOS in precisely this gap. In that process, what did you learn about the missing layer between a capable model and the work that actually has to get done?

Marco: I learned a great deal. PandaOS came out of a conversation between our lead Riccardo, the third co-founder, and me. I told him I thought we needed a new form factor for devices and operating systems. I feel like we are currently riding horse carriages with an engine in them, and not cars. I thought the model had to come first. Then about a month later he came back and said: look, I think I built what you want, have a look. And I was very happy about that.

A lot came out of the process since. One thing is that I really like seeing what the tool does. That is what I miss in most other tools, that you actually get visible feedback on what the model is doing right now. I did not even realise it was one of PandaOS's USPs until I started working with other tools and thought, oh wait, that is not how this works, you usually do not see what is going on under the hood. I do not want to miss that anymore.

PandaOS is described as desktop-first and local: your keys, no data training. Architecturally, what genuinely stays on the machine, what can still reach out to external services, and how should a user understand that boundary?

Philipp: Imagine every level of customisation you can think of at your fingertips. Desktop-first means the software can interact with whatever is running on your machine, and if what is running on your machine is also connected to the internet, you can connect that too. Here comes the tricky bit that most people forget: where you can really leverage your own sovereignty is permissions. You give the model the right level of detail to execute your request without it actually reading the data, because maybe it does not need to.

That is what we typically do with databases. Instead of going into the raw data, the schema is more than enough. You look at the schema, you plop that into the LLM, it works out what it needs to write as an SQL query, and then the query runs on your device. There is no need for it to go all the way down the pipeline to see what is in the tables. That is where we see a lack of transparency in a lot of other products, and where we want to cut through and reveal it to the user, so that depending on their level of confidence or necessity they can really dial it in for their specific use case.

PandaOS is a Swiss Army knife. How do you keep it useful rather than overwhelming?

Marco: We are building the part that holds the individual tools. Customers choose the blades. Our clearest groups are European enterprises with higher data-security requirements, developers who benefit from deep deployment integration, and office teams. We start with one department and one defined scope, then expand. You cannot tackle an entire organisation at once.

Philipp: Our design partners found uses we had not imagined. That is why we call it an OS: the environment adapts to the user's rules, data and recurring workflows. It becomes a personal setup rather than the same interface for everyone.

Marco: The comparison with a friendlier SAP is flattering, if premature. I am convinced companies will eventually run an AI operating system that orchestrates their data and models; PandaOS is one early attempt at that layer.

Can a workspace or orchestration layer actually increase sovereignty, or does it risk becoming a nicer interface sitting on top of exactly the same dependencies?

Philipp: It depends on how you set it up. An orchestration layer is necessary to create proper workflows, because you need to be able to switch from one tool to the next, and that is the reality of things. But the more powerful a tool becomes, the more risk is involved, and you cannot disassociate those two. The way we see it, by creating more transparency around that orchestration layer, the sovereignty comes with it, and so does trust in the tool, because you can see what changes are being made, what triggers are firing, what permissions are being asked for. Not only do you feel more in control, you factually are more in control, because you can switch that on and off depending on your situation.

How you deal with it is another matter. If you are happy to let it run wild, you are giving up a level of sovereignty over how you work with your data or your inference, but that is your choice to make. Sovereignty is not an on-off switch. It is a scale. Depending on the layers you add, or where you draw the line, that is where you make the trade-offs.

Let me steelman the other side. What is the strongest argument against the whole European AI sovereignty narrative, the case that it is a distraction?

Philipp: The case against it is basically nobody got time for that. We have been waiting and waiting, and we do not want to wait. It is a land grab. Everybody wants to be on the edge, ahead of the curve if they can, so they take whatever is cutting edge right now. That is the biggest argument: why bother building our own if others are already available?

But this is where the argument flips and becomes moot, because it is true until it is not true anymore. That is when there is an outage, or no access, or a policy shift, or a person who decides prices need to go up. And then it is too late to have an exit.

How do PandaOS and Pandata relate?

Marco: They are independent sister companies. Pandata has existed for ten years; PandaOS is six months old. I split my time across both, while the other founders work exclusively in one company. The overlap is useful, especially because companies still need solid data infrastructure before they can use AI well, but we do not expect PandaOS consulting to replace Pandata's core business in the short term.

You have competitors claiming to solve this sovereignty problem. If I am a buyer, how do I tell a pretty screen that makes me feel comfortable from something that is actually doing the work?

Marco: There are many tools at the moment that can do the work effectively. The question is much more what fits your profile and your company. The reason for PandaOS could be that you want something that integrates more deeply into your company. Most of our competitors do not offer desktop solutions, and because we do, we can do a lot more. We can use inference on the edge once it really becomes a thing, but we can already use your computer, and that is enormously powerful.

Developers notice it earlier because they are closer to the hardware, but even as an office worker you will notice. I was annotating PDFs a while ago with PandaOS, which is really hard when you are bound to your browser, and Panda just used my macOS tools to annotate them for me, which I found absolutely insane.

Give me two examples.

Marco: I plugged in an e-ink device whose wallpapers no longer fit and asked PandaOS to fix them. It found the folder, detected the screen size and cropped every image correctly, then overwrote the originals with my permission. In another project it recognised that ADB was installed, asked to use it, and read an Android display and captured screenshots within minutes. Both worked because the desktop app could see and use local tools rather than staying inside a browser tab.

What is something people would expect you to claim your product can do, maybe to win a sale, that you would not claim yet?

Marco: The one that comes up a lot for me is anything design-related. We do have design features, and an artifacts feature that is very advanced and can build slide decks based on your company design. But as soon as it comes to creating images, video, classical design jobs, that is not what PandaOS is for, and I am not even sure it ever will be. We do not really cater to the creative industry.

Philipp: One other thing that is not central to today's discussion but very central to how we think about building PandaOS: proper context. One of the elements we focused on from the beginning is a feature we call PandaAtlas, a kind of neural brain that lives inside PandaOS and creates nodes based on your usage, the tools you invoke, the skills you use. The more you use it, the more it customises itself. That is what I meant about personifying my interface: it genuinely understands what is necessary for an easy workflow.

What we want to build, and have not set up yet, is the next layer of that network: on an organisational level, what do you learn, and what can you share? Say I have found a solution to a problem, and Marco happens to be working on the same problem on his end. His setup could look into mine, pick up that solution and spare him the headache. That network effect can become a superpower if you share the knowledge base across an organisation. We are actively working on it because we know it will be very powerful and necessary. But it is not something we can promise today, because there are problems around it we have not solved: infrastructure, making sure the right data is in the right place, that it does not get corrupted. There are a lot of question marks and we are not quite there yet.

If you were starting from scratch right now, what stack decisions would keep you from boxing yourself in?

Marco: Own your own data. That is number one. Build a data infrastructure that you own, that lets you move to a different provider at any point, and build everything else on top of it. When it comes to data sovereignty and being in control, it means getting control of your data first, and only then seeing what AI can do on top of it and which parts of AI you want to own. It is not feasible to own everything in terms of AI at the moment, I do not believe that. But it is absolutely feasible to own your data, and then in the next steps you see which parts of AI you can feasibly own.

Over the next 18 months, what would have to happen for European alternatives to feel genuinely competitive in day-to-day company workflows, and not just in a press release?

Philipp: My impression is that first they need to be commercially viable and available. It is fine to have a very good foundational model built in house, but it does not help much if the masses cannot use it or do not know how to set it up. If I had something on my wish list, it would be to make those models more accessible. I would love to have a far deeper go at Mistral and the others than I currently do. They are available to me, sure, but I cannot benchmark them as well as I would like, because there is not the same kind of user base.

Marco: One bit I would add: Europe still has a chance when it comes to more specialised models. We had it a while ago with Black Forest Labs, where we were leading in a specific part. And there is another company, from Freiburg, working with SAP to deliver one specific model trained on tabular data. I think we will see more specialised models, and that is a huge chance, because the general models are becoming commoditised anyway over time.

We started with the Anthropic moment and the Austrian Anthropic debate. What is the one uncomfortable lesson Europe should take from it?

Philipp: Wake up. That is the very short version. It is not sufficient to rest on the fact that you have something commercially available to you today. It might be taken away from you tomorrow. So it makes more sense to invest properly and build out your capacity here, at the very least to have a fallback if it is not at the frontier. That is the big lesson from that episode.

Marco: Diversify. Do not rely on one model from one provider. Think of it like a stock portfolio.

Where can people follow what you and the PandaOS team are building?

Philipp: You can find us at pandaos.ai. Riccardo also has a very active X profile where he shares builds, struggles and ideas from the product's development.

Editor's note: PandaOS was not yet downloadable when this interview was recorded in July 2026; the app is available now.

Watch Fin and Plaid now on demand to see how AI can resolve issues directly in customer conversations. This includes everything from reducing bank-linking friction, enabling proactive support, and helping customers complete real financial tasks.

Watch now

The interview's most useful claim is also its most modest: sovereignty means staying able to leave. Can you switch inference on a random Tuesday without breaking operations or affecting customers? That is easier to test than a flag on a data centre.

The founders do not tell companies to replace chips, frontier models or every cloud dependency. They tell them to own their data and preserve the switch. Philipp estimates that 99% of their customers do not need on-premise AI; the serious cases are government contracts and unusually sensitive IP.

The harder enterprise lesson is Marco's account of companies defaulting to a half-disabled Copilot and exhausting workshop credits on day one. That is a procurement problem, not a jurisdiction problem.

PandaOS also creates the question every sovereignty vendor should answer: does the layer that reduces one lock-in create another? Own keys, visible permissions and no training on user content help. PandaAtlas cuts the other way: a private knowledge graph of individual and organisational work can make the product more useful and more expensive to leave. The team is candid that the organisational layer remains unsolved.

On Europe, their case is narrower than the political slogan. An American provider with a European address still carries US jurisdiction; specialised European models may be a more realistic opening than trying to replace the frontier outright. Black Forest Labs and Prior Labs provide evidence for that route.

For the record: Anton Carniaux gave the Microsoft testimony to the French Senate in June 2025; Airbus confirmed its Scaleway migration in July 2026; SAP completed its acquisition of Prior Labs on 17 July 2026; and EU Inc remains pending. The workshop anecdotes are Marco's accounts and are not independently documented. The line worth keeping is Philipp's: sovereignty is not an on-off switch. It is a scale.

PandaOS is a local-first AI workstation that connects an editor, terminal, browser, databases, deployment tools, business apps and an inbox into a single desktop workspace, running on the user's own model keys from OpenAI, Anthropic, Google and others, with no vendor lock-in and no training on user content. It was spun out of Pandata in early 2026 and is at pandaos.ai. Pandata has been advising organisations on data strategy, business intelligence and AI infrastructure from Berlin for ten years.

Sources:

🔗 The full video interview on our YouTube channel: https://www.youtube.com/watch?v=NMN-_Le2Am8

🔗 Anthropic, "Redeploying Claude Fable 5" (30 June 2026): https://www.anthropic.com/news/redeploying-fable-5

🔗 PandaOS: https://pandaos.ai and Pandata: https://www.pandata.de

🔗 Trending Topics, "Austria Wants to Bring Anthropic to Europe" (29 June 2026): https://www.trendingtopics.eu/austria-wants-to-bring-anthropic-to-europe/

🔗 Bloomberg, "Austria Lobbies EU to Host Anthropic After US Access Curbs" (28 June 2026): https://www.bloomberg.com/news/articles/2026-06-28/austria-lobbies-eu-to-host-anthropic-after-us-access-curbs

🔗 The Register, "Microsoft exec admits it 'cannot guarantee' data sovereignty" (25 July 2025): https://www.theregister.com/2025/07/25/microsoft_admits_it_cannot_guarantee/

🔗 The Register, "Airbus migrating 70 critical apps from AWS to France's Scaleway amid digital sovereignty push" (16 July 2026): https://www.theregister.com/paas-and-iaas/2026/07/16/airbus-migrating-70-critical-apps-from-aws-to-frances-scaleway-amid-digital-sovereignty-push/5272373

🔗 SAP, "SAP to Acquire Prior Labs to Establish a Globally Leading Frontier AI Lab in Europe" (May 2026): https://news.sap.com/2026/05/sap-to-acquire-prior-labs-establish-frontier-ai-lab-europe/

🔗 European Parliament, "EU Inc.: what is the 28th regime?": https://www.europarl.europa.eu/topics/en/article/20260506STO42807/eu-inc-what-is-the-28th-regime

🔗 Epoch AI, "Most large-scale models are developed by US companies": https://epoch.ai/data-insights/large-scale-models-by-country