Our Take
US export controls are working as intended: they're making it economically rational for well-funded European firms to build their own models instead of buying American ones.
Why it matters
If European AI labs succeed at building competitive models on local infrastructure, the US loses pricing power and market access in a region worth billions. If they fail, the controls simply delay European adoption without creating an independent capability.
Do this week
Enterprise buyers: audit your model licensing agreements now to understand what happens to your inference costs if US vendors raise prices in response to reduced European demand.
European companies are funding domestic AI development to avoid US export restrictions
Reuters reports that American curbs on advanced chip exports to Europe are prompting European firms to invest in building their own AI models rather than licensing from US vendors like OpenAI or Anthropic. The reporting does not name specific companies or funding amounts, but the pattern is clear: restricted access to US inference infrastructure creates an economic incentive for European firms with capital to develop local alternatives.
This is not new industrial policy. It is basic substitution economics. When a critical input becomes scarce or expensive, buyers either find a workaround or build it themselves. US export controls on advanced AI accelerators (particularly Nvidia H100 and H200 chips) have already driven up costs and lead times for European cloud providers and AI labs. For large, well-capitalized firms, training a model in-house on available European hardware becomes a rational choice.
The specific mechanism matters. Europe has access to older-generation chips and some domestic GPU capacity. Training a state-of-the-art LLM on this hardware takes longer and costs more than training in the US with unrestricted access to the latest silicon. But if the alternative is paying a premium to license inference from an American vendor, and if you have the engineering talent and capital, you build.
This reshapes the vendor moat and the geopolitical leverage of US AI firms
For OpenAI, Anthropic, and other US-based model labs, the loss of European licensing revenue is material. Europe represents significant enterprise demand. If customers can access a locally-trained alternative that meets their compliance, data residency, and latency requirements, price becomes the tie-breaker. US vendors lose the ability to capture value from a region where they once held monopoly pricing power.
For US policymakers, this is the intended outcome of export controls but also the risk. The goal is to slow China's AI capability. The side effect is accelerating Europe's independence, which may eventually reduce US influence over global AI standards and deployment. Europe building its own models is not the same as China or Russia doing so, but it still fragments the US-dominated AI market.
For European firms, the calculus is simpler: build now while chips are constrained, or be locked into US vendor dependency forever. The window to establish an independent stack is open but will not stay open indefinitely. Once US export controls loosen (either through policy change or technological obsolescence), the economic case for European in-house development weakens.
Expect fragmentation and higher prices for US inference in Europe
Enterprise customers should anticipate two outcomes over the next 12 to 24 months: first, European-trained models will appear with feature parity to US baselines on tasks where data residency or compliance matters. Second, US vendors will raise prices in Europe to capture value before that happens.
If you are deployed on US inference infrastructure in Europe, lock multi-year pricing now before rate increases. If you are evaluating new model deployments, run benchmarks on both US and European alternatives. By 2026, the default assumption that US models are cheaper will no longer hold in Europe.