Europe is not caught between two powers. It is being squeezed out by both.

For the past two years, the working assumption among European AI buyers was simple: if US models get too expensive or too restricted, switch to a Chinese alternative. DeepSeek made that feel credible. Qwen made it feel practical. The escape hatch was real, and a lot of procurement strategies quietly depended on it.

That hatch is closing.

China is now weighing export controls on its most capable AI models, according to Reuters reporting in June 2025. Meanwhile, US controls on advanced chips and model access have been tightening steadily since 2023. The two largest AI producers in the world are, at roughly the same moment, pulling their best work out of reach for European buyers. This is not a geopolitical coincidence. It is a structural condition. And agencies working with AI need to treat it as one.

The squeeze is not symmetric, but it is real on both sides

US restrictions have been well-documented. Export controls on H100 and successor chips limit what European cloud providers can offer. API access to frontier models like GPT-4o and Claude 3.5 Sonnet remains available for now, but the terms of access, the data residency requirements, and the compliance overhead have all increased. For agencies doing work for clients in regulated sectors, finance, health, public sector, the friction is already material.

The China side has been less discussed in Amsterdam agency circles, partly because the open-weight story was so appealing. DeepSeek R1 dropping in January 2025 genuinely shifted how people thought about model access. If a Chinese lab could release weights freely and match frontier performance, the geopolitical risk felt manageable. You could run the model yourself. No dependency.

But the models that matter most, the closed, proprietary ones sitting above what is publicly available, are precisely the ones that may now face export restrictions. The open-weight story only holds if the open weights are competitive. As the capability gap between closed and open models persists, the escape route narrows.

What this means for agencies specifically

Agencies sit in an awkward position in this structure. You are not buying AI for yourself. You are building pipelines, products, and workflows on behalf of clients who have their own compliance requirements, their own procurement rules, and their own risk tolerance. When a model disappears from a pricing page, or when terms change mid-project, it is not just your problem. It is your client's problem, delivered by you.

The practical risks break into three categories.

Dependency concentration. Most agencies currently run on one or two model providers. If access to either narrows, the fallback options are worse, more expensive, or both. A multi-model strategy is not a luxury. It is basic risk management, and most teams have not built it yet.

Client contracts. If you have committed to a deliverable that assumes continued access to a specific model or capability, and that access changes, you have a gap in your contract. This is especially sharp for retainer relationships where AI is baked into the delivery model.

Talent and tooling. Teams that have built deep expertise in a single model family, its quirks, its prompting patterns, its API structure, carry transition costs when they need to move. That cost is real but hard to quantify until you are already in the middle of paying it.

Europe does have options, but they are smaller and slower

The honest case for European alternatives is that they exist, are improving, and are meaningfully less exposed to US or Chinese export policy. Mistral is the obvious name. Aleph Alpha, though it has shifted strategy, has contributed to the infrastructure conversation. There are serious research groups at universities across the continent building on open foundations.

The honest case against them is capability. On most benchmarks that matter for agency work, writing quality, code generation, multimodal understanding, European models are not yet competitive with the top US and Chinese offerings. That gap may close. It will not close this quarter.

So the practical strategy for now is portfolio thinking rather than replacement thinking. You keep using the best available models for the work that warrants them. You build your workflows so that the model layer is swappable, not load-bearing in the architectural sense. And you track European model development seriously rather than as an afterthought.

This is not a nationalist argument. It is a supply chain argument. When two suppliers coordinate, even unintentionally, to reduce your access at the same time, you need more suppliers.

The strategic move is architectural, not political

Lobby for AI sovereignty if you want. Support European model development if you can. Those are worthwhile positions. But the thing that protects your agency and your clients in the next 18 months is not a policy outcome. It is how you build.

Concretely, that means three things.

First, abstract the model from the application. If the name of the model is hardcoded into your product or your process, you are carrying risk you do not need to carry. Orchestration layers, whether you use LangChain, a custom wrapper, or something else entirely, exist precisely to decouple capability from dependency.

Second, document your model choices like you document your other technical decisions. Why this model for this task. What the fallback is. What the performance threshold is that would trigger a switch. This is basic, and almost no agency does it.

Third, run experiments on alternatives before you need them. Not because they are better today, but because switching under pressure is expensive. The agency that has already run Mistral Large on its core use cases, knows where it underperforms, and has a view on acceptable tradeoffs, is in a fundamentally different position than the one that starts that work when a pricing email lands in their inbox.

The squeeze is real. The response to it is not panic, and it is not waiting for Brussels to sort it out. It is building systems that can move.