--- title: "Your Competitor No Longer Comes From Your Own Street" author: Marc Diks date: 2026-08-17 modified: 2026-08-17 category: AI & Strategy reading_time: 10 min url: https://www.marcdiks.nl/en/blog/ai-competition-rival-not-from-your-street canonical: https://www.marcdiks.nl/en/blog/ai-competition-rival-not-from-your-street language: en --- # Your Competitor No Longer Comes From Your Own Street *The world's best AI models come from the US or China. Europe doesn't have them, and that changes who ends up competing in your market.* This summer I read a scenario set in 2028 in which a law firm in Milan loses out to an American firm. Not on knowledge of Italian law, but on access to an AI model. And that scenario is less far off than it looks. It turned out that way just a few days after publication. It comes from [Europe 2031](https://europe2031.ai/), a piece by Daan Juijn and Stan van Baarsen, among others, about where Europe stands if the current line is extended. The authors are honest about what it is: everything up to June 2026 actually happened; after that the speculation begins. The Milanese firm sits on the speculative side. But the mechanics underneath do not. That firm charged premium rates for years for something scarce: deep knowledge of Italian commercial law, in Italian, delivered by people who knew the local practice. In the scenario it competes with an American firm whose AI handles Italian, French and German law at the same time. Faster and cheaper. And the bitter part of that passage: the Italian lawyers aren't allowed to use the best models themselves, blocked by internal rules. The authors don't leave it there. They add that the same pattern plays out in consultancy, software, marketing and finance. And that's where this piece suddenly becomes about you. ## This isn't about cost Before you read on, I want to clear up one misconception. I regularly hit the usage limit of my AI tools. Then I'm out for a few hours, sometimes a day. Honestly, that's not even a problem. I go work out, I do something else, and the next morning it just works again. You could call that an annoying side effect of too much dependency, something you solve with a bigger subscription. That conclusion is wrong, and that's why I mention it. This piece isn't about what your AI costs, but about whether you can reach the best models at all. Those are two different questions, and the consequences diverge widely. In one case you simply pay more. In the other, you can't deliver something your competitor can, and your client notices. ## Eighteen days without a model That the second question isn't theory, we've known since this summer. On 12 June 2026 Anthropic received a directive from the US Department of Commerce ordering it to suspend access to its two most powerful models, Fable 5, for all non-Americans, anywhere in the world. Because nationality couldn't be verified per user, the models went dark worldwide for everyone. On 30 June the export controls were withdrawn, and on 1 July access was restored. Eighteen days. Not an outage and not a vendor's choice, but a decision by a government you have no vote in. On what that means for the continuity of your own processes I wrote earlier in [AI vendor lock-in: why the kill switch is your problem](/en/blog/ai-vendor-lock-in-kill-switch). That was about how to keep your business running when a model falls away. That question still stands, but it isn't today's question. Because there's a second consequence that gets far less attention. If access to the best models becomes politically divided, then that access has itself become a competitive advantage. And advantages you don't control yourself are the most unpleasant kind. ## What used to make entry hard Here comes the marketing side, and for you as an entrepreneur it matters more than the geopolitics. In his work on competitive strategy, Michael Porter described the barriers that make it hard for a new party to enter a market. The higher those barriers, the better you sleep. Capital, scale, brand and distribution belong on the list too, but for a Dutch service provider two other barriers were the strongest protection for years: language and local knowledge. A foreign party wanting in had to learn Dutch, get to know Dutch practice and build trust. That took years, and usually they didn't bother. Those two barriers are evaporating. A good AI writes flawless Dutch, knows Dutch legislation and produces subject-matter texts you can't tell apart from those of a local specialist. The knowledge barrier has become a licensing question. Let me make it concrete, because in the legal profession it may feel far from your world. A translation agency sold two things for years that weren't available separately: language mastery and expert knowledge of legal or medical terminology. Both are now purchasable, and the client who once had no alternative now has ten. At a marketing agency that depends on Dutch copy and campaigns, it's subtler. The argument was always that you need to sense the Dutch consumer, and that argument still holds. It just gets a little thinner every quarter. For an accountancy firm, in-depth knowledge of Dutch tax rules has long been the core of the offering. That knowledge now sits in models that get better at it every month. What remains is the contact, the responsibility and the judgment, and that's a narrower base than the one your business was built on. Fill in your own profession. The question that counts: can someone with a better model soon deliver what you deliver, at a price you can't match? Regulation does hold up, by the way, and that's an important nuance. A lawyer has to be registered, a financial adviser needs a Wft licence. A party in Vietnam can't simply advise here. But don't get rich on that. A licence lets you into the market; it doesn't protect you from who's already in it. Whoever is based here and uses better models than you has exactly the same papers and better tools. On what a licence is still worth I wrote earlier in [the future of the insurance intermediary](/en/blog/insurance-broker-license-right-to-exist). So a new barrier has been added that appears in no classic model: access to the best models. The difference with the rest is that you have no say over it. You can work on capital, on a licence too, and you build up language and local knowledge if you put in the time. On an export decision in Washington, you can't. ## What does Europe have, then Then the logical question: don't we just use our own models? You can, but then you're not working with the best tool there is. Mistral is the only European player of any size and does fine work, but the company trails the absolute top. You can also run open-weight models yourself, but almost all of those come from America or China too. You then move where the model runs, not who made it. For a lot of work that's more than enough. I wrote earlier why [open source AI has shifted from a smart alternative to a governance necessity](/en/blog/open-source-ai-governance-necessity). But it doesn't solve your competition problem. If you work on a good-enough model and your competitor on the best one, you see that difference in what you deliver. ## Someone in Groningen just started There's a countermovement too, and it began closer to home than you'd think. While Brussels talks about billions and gigafactories, Jantine Doornbos of Groningen-based HostYourAI built a Dutch AI model you can use today. It's called [Loes](https://loes.ai/), is trained and hosted on GPUs inside the EU, and the weights and the complete training set are public on Hugging Face. You can see exactly what the model was trained on. It has an API compatible with OpenAI's, which means you only need to point your existing code at a different URL. Doornbos added something herself that stuck with me: one person, one week, without millions of euros is enough to make a start. Now for the honesty, because otherwise I'm selling you something the numbers don't support. Loes is no replacement for the best American models, and doesn't pretend to be. The project deliberately compares itself only with other European models. The European track scores strongly on language in the Dutch benchmark, but clearly lower on knowledge and reasoning than what you're used to from a frontier model. And the most uncomfortable detail: Loes runs on two tracks, the second of which uses a Chinese base. Even a sovereign Dutch project doesn't fully break free from non-European building blocks. That's where we stand. There's a start, and that start is real. But we're far from there, and anyone claiming otherwise is engaging in wishful thinking. ## I put the same expert questions to four models Claiming that the difference at the top matters is easy, so I tested it. I put three questions from my own field to four models. From America, Claude Opus by Anthropic and Gemini with 3.1 Pro by Google took part; from Europe, Vibe by France's Mistral and Loes by Groningen-based HostYourAI, running the Super Loes model. - **The first question:** from what date do insurers' risk models fall under the high-risk requirements of the [EU AI Act](/en/eu-ai-act)? - **The second:** how far does an insurance intermediary's duty of care reach in cases of underinsurance? - **And the third:** when does a liability policy cover pure financial loss? Three topics whose answers I can judge myself. Click through the models and, per question, compare the prompt with the answer. [model-vergelijking] On the factual question I found no difference that matters. All four cited 2 December 2027, including the delay via the Digital Omnibus, some with more caveats than others. Loes actually gave one of the sharpest answers there. It pointed out that the heavy requirements apply only to life and health insurance, and that the transparency duty has been running since early August. On how shaky that date still looked earlier this year I wrote in [AI Act delayed to December 2027](/en/blog/eu-ai-act-delay-failed-may-2026): on 28 April the negotiation was still deadlocked; only on 7 May did the agreement come. On the two reasoning questions the picture diverged, but not as neatly as you'd want. Opus answered at the level of a specialised lawyer: the legal framework, the burden of proof, a string of rulings with citations. When I checked those rulings, the core was right, but there was also an error in it. Opus attached a Supreme Court ruling (Sauna Peize/Rabobank, 6 September 2019) to the wrong court, as if it were a decision by the District Court of Midden-Nederland. And one ECLI number it cited turned out to belong to a completely different, tax-related case. Loes stayed much shorter and missed the deeper layers, but at first glance seemed to have nothing wrong. It cited a real ruling by the District Court of Gelderland dated 30 July 2025, about an intermediary who failed to warn about the expiry of a guarantee against underinsurance. That case exists, and the core of what Loes wrote around it is right. But Loes did invent a party name. It named the claimant "Blaak & Partners B.V.", whereas Blaak is the name of the intermediary and the claimant was in reality a homeowners' association. The damages figure it mentioned I couldn't find anywhere in the ruling either. Vibe, Mistral's model, made a different kind of error: it mixed up two abbreviations for insurance types, calling one a PAV and then, in the summary, suddenly meaning an AVP, which is a completely different policy. Three models, three different kinds of error, hidden in answers that each came across as self-assured. That's the real lesson of this test, and it isn't about which continent builds the best model. The most confident answer isn't the most reliable, from any model. Check the source. Where there was a clear pattern: on the factual question everyone did well. On the questions that really came down to reasoning, Opus went the deepest, despite its own error. And that's precisely the model of the company that in June had to switch off another model, Fable, for eighteen days after a letter from Washington. Same company, same export regime, different model. The difference at the top exists, but whether you keep access to it, and whether you can trust the answer without checking, are two separate questions. Both important. ## What you do with this on Monday This isn't a problem you as an entrepreneur will solve. You're not going to change the European model landscape. What you can do is avoid being caught off guard. First map which models run in your company and who owns them. Not just the tools you consciously chose, but also the AI hidden in software you've used for years. That list is almost always longer than you think. Then test once whether you can really switch: run for a week on an alternative model and see what gets worse. Then you know what's coming, instead of having to find out at the moment you must. Next, look at your own offering through the eyes of someone just walking in. Which part of what you sell is knowledge that sits in a model, and which part is judgment, responsibility and relationship? That first part is under pressure. The second part is where your margin needs to go. And build the habit of always checking a concrete answer against the source, regardless of which model gave it. My own test showed that both the most expensive and the cheapest model could confidently cite something wrong. That risk doesn't disappear by picking a better model. And ask yourself the question this whole piece raises: what will my competitor soon deliver that I can't, and is that because he has better tools? That question is more uncomfortable than the question of whether your AI is too expensive. It's also more important. For years you knew where your competition came from. From your own city, your own industry, with the same constraints as you. That's over. And the party entering your market doesn't need to speak your language. Its model does that for it. ## Sources - The scenario about the Milanese law firm and Europe's AI uptake: [Europe 2031](https://europe2031.ai/) - The Dutch sovereign AI model, the open weights and the benchmark figures: [loes.ai](https://loes.ai/) - The District Court of Gelderland ruling Loes cited in the test, with a party name and amount that differed from the model's answer: [ECLI:NL:RBGEL:2025:7254](https://uitspraken.rechtspraak.nl/details?id=ECLI:NL:RBGEL:2025:7254) - The Supreme Court ruling Sauna Peize/Rabobank that Opus wrongly attributed to the District Court of Midden-Nederland: [ECLI:NL:HR:2019:1298](https://uitspraken.rechtspraak.nl/details?id=ECLI:NL:HR:2019:1298)