Fable Follies Sharpen Europe’s Sovereignty Conundrum
- 22 hours ago
- 5 min read
Although the US government lifted an export control order on Anthropic’s artificial intelligence model, the damage to transatlantic ties remains.
The episode has sharpened Europe’s dilemma over digital sovereignty. Instead of accepting a false choice between depending on US AI software or building an independent European AI capability, policymakers on the continent should follow the example that is already being set by industry and mix and match different AI models to reduce dependency on any single supplier.
The Trump administration told Anthropic to cut off access to Fable after it grew worried the model could be manipulated into sharing dangerous cyber capabilities. Anthropic restored access to Fable after the US government signed off on new safeguards.
Although everyone shares an interest in making sure dangerous cyber capabilities don’t fall into the wrong hands, the Fable episode demonstrates that the US can block access to a powerful American AI system. European officials must ask what might have happened if European companies had spent months building critical systems on top of Fable before Washington flipped the switch. The fear that such a kill switch could be fired again has strengthened European calls to reduce dependence on American suppliers.
But the growing clamor for digital sovereignty will not change the fact that Europe is poorly placed to build its own frontier AI. The US produced 59 notable AI models in 2025, according to Stanford’s 2026 AI Index, well ahead of China’s 35. European labs produced only a handful. France’s Mistral, seen as Europe’s most capable model developer, trails the leading American and Chinese models.
The gap is widest in advanced reasoning and the ability to autonomously perform multi-step tasks, which can find hidden cyber-vulnerabilities and string together the steps needed to exploit them. Mistral’s models score well behind the leading American and Chinese models on graduate-level reasoning benchmarks like GPQA Diamond, for example. Anthropic’s and OpenAI’s models currently lead the field in working through complex, multi-step tasks, solving general problems with minimal guidance and simulated hacking exercises.
Europe’s disadvantages are structural. The continent hosts just under 5% of the world’s large GPU clusters — used to train and run advanced models — compared to roughly 74% for the US. European companies’ electricity costs are roughly double those of their US counterparts.
The European Commission’s answer to this conundrum is the Cloud and AI Development Act (CADA), proposed in June and yet to be negotiated with the European Parliament and national governments. CADA sorts workloads into tiers with different sovereignty requirements: the most sensitive applications would have to run on a full sovereign European stack, while less sensitive ones would face fewer limits. Defining the different tiers and who can participate in them will be the subject of lively debate, but Henna Virkkunen, the Commission Executive Vice-President overseeing the effort, has said the aim is to ensure that no provider holds a “kill switch” over European services.
CADA contains a paradox, however. If European AI models continue to lag behind American ones in cyber capabilities, which seems likely, the most sensitive workloads facing the strictest buy-European rules would end up running on weaker models, with weaker cyber defenses. Europe would be trading insulation against a potential US AI “kill switch” for another vulnerability.
Away from Brussels, the AI market may have stumbled onto a partial solution. The industry has spent the past two years reducing single-model dependency. Companies deploying AI are increasingly turning to multi-model systems, in which an orchestrating program routes each task to whichever model handles it best. A system running inside a law firm might send document summaries to a small, cheap model; more complicated searches to a capable open-source model; and the final legal brief to a frontier model, for example. The system design relies on numerous models rather than one.
Companies started adopting this approach to save money. In 2024, researchers at the University of California, Berkeley showed that routing between a powerful AI model and a cheaper one could cut the cost of answering a query by more than 85% on general tasks and 35–45% on hard reasoning problems, with only a 5% performance gap on the former.
The idea is catching on, and even inspiring new businesses. OpenRouter, a US-based commercial routing platform, offers access to hundreds of models from dozens of providers. It says it now processes about 100 trillion tokens (the units of text that models read and generate) per month, a five-fold increase over six months ago. This capability doesn’t necessarily need to depend on a foreign service provider. Open-source tools such as LiteLLM let organizations run routing on their own machines. This accelerating shift in the way AI systems work may have an unintended benefit for Europe. Multi-model systems are built to be model agnostic. They can switch from one model to another as prices and capabilities change. It’s not a huge stretch to think that the same design idea could be used to build systems capable of swapping one frontier model for another, or to degrade gracefully when a model becomes unavailable, instead of failing outright. European companies are already moving in this direction. At a startup conference in Paris last month, as the Fable situation was unfolding, Siemens, Renault, and Orange talked about how using US, Chinese, and European models could help them avoid depending on a single model provider.
Admittedly, the multi-model approach does not represent a complete solution to Europe’s sovereignty conundrum. Although it would expand options and reduce single points of failure at the model and application layers, it would not answer questions about how to best address data access issues and other sovereignty concerns related to cloud computing and AI infrastructure.
Multi-model systems would require extra upfront investment and testing to understand how systems behave when different models are swapped in and out. If routing simply swapped one US frontier model for another, it would still leave Europe exposed to Washington’s extraterritorial policy toolkit. And while Chinese models account for a large and growing share of routed traffic worldwide, they are unlikely to be an acceptable option for sensitive European workloads.
This is where public European money could make a difference. The EU already supports open-source projects such as the OpenEuroLLM consortium, which trains open-source models on European AI infrastructure.
Public funding for multi-model approaches and greater investment in open-source projects would be far cheaper than pursuing a European frontier model. This kind of engineered failover strategy is already used in the banking and energy sectors. Extending the idea to sensitive AI systems would preserve access to cutting edge AI. A shift towards multi-model systems will not make Europe self-sufficient. But it represents a smart, if partial, solution to Europe’s sovereignty conundrum.
By Kevin Allison and Venesa Rugova. Kevin Allison is a non-resident Senior Fellow with the Technology Policy Program at the Center for European Policy Analysis (CEPA). He is the Founder and President of Minerva Technology Futures, a geopolitical intelligence and policy advisory firm specializing in artificial intelligence and its associated technology stack. Venesa Rugova is a Senior Analyst at Minerva Technology Futures, where she specializes in artificial intelligence and cybersecurity policy. Article first time published on CEPA web page. Prepared for publication by volunteers from the Res Publica - The Center for Civil Resistance.
Comments