Europe is regulating artificial intelligence as if it already has world-class frontier AI companies, but it does not. A recent U.S. government directive suspending access to Anthropic's advanced models has exposed a critical strategic vulnerability: the European Union's approach to AI governance may be actively preventing the continent from developing its own competitive AI champions.

What Is the EU AI Act's Compute Threshold, and Why Does It Matter?

The EU AI Act classifies general-purpose AI models as presenting "systemic risk" when they exceed a specific computational threshold. Specifically, models trained using more than 10 to the 25th power floating-point operations, or FLOPs, face additional regulatory obligations including safety testing, incident reporting, and cybersecurity requirements. To put this in perspective, this is an enormous amount of computing power; by June 2025, more than 30 publicly announced AI models from 12 different developers had already crossed this threshold, meaning the rule already applies to today's most advanced systems, not hypothetical future ones.

The problem, according to technology analysts, is that this threshold creates a perverse incentive. Training a frontier AI model requires massive compute resources, capital, specialized talent, and infrastructure. The regulatory framework essentially penalizes companies for reaching the scale necessary to compete globally. It is as if Europe told its aircraft industry that small planes could be built freely, but once a company approached Boeing or Airbus scale, it would face a special regulatory burden.

How Does This Regulation Change Company Behavior?

When a regulatory threshold is attached to scale, companies respond rationally. Rather than risk entering a heavier compliance regime, frontier AI developers may choose to avoid training above the threshold, split development work across jurisdictions outside Europe, delay product releases, reduce capabilities available to European users, or structure their operations so that cutting-edge work happens elsewhere. These are not theoretical concerns; they are predictable business responses to rules that make scale more expensive and complicated.

The timing is particularly problematic. The United States can restrict access to frontier models by government directive, as demonstrated by the Anthropic suspension. China is building national AI capacity as a strategic priority. Gulf states are investing heavily in compute infrastructure. Meanwhile, Europe is defining compliance paperwork for companies that have not yet been built.

What Are the Key Regulatory Burdens on Advanced AI Developers?

Beyond the compute threshold, the EU AI Act imposes several obligations on providers of general-purpose AI models, particularly those classified as high-risk or systemically important:

• Technical Documentation: Companies must maintain detailed records of model architecture, training processes, testing methodologies, and evaluation results, creating substantial compliance overhead.
• Downstream Transparency: Providers must share information with downstream users and suppliers, limiting proprietary control and competitive advantage.
• Copyright Compliance: Companies must establish and publish policies governing the use of copyrighted training data, a requirement that may conflict with the data needs of frontier model development.
• Training Content Summaries: Developers must publish detailed summaries of training data according to templates provided by the EU AI Office, further constraining operational flexibility.
• Adversarial Testing: For systemically risky models, companies must conduct and document adversarial testing to identify vulnerabilities, adding time and cost to development cycles.
• Incident Reporting: Serious incidents must be reported to regulators, creating ongoing surveillance and potential liability.
• Cybersecurity Requirements: Enhanced security protections for model infrastructure are mandatory, increasing operational complexity.

Each obligation may sound reasonable in isolation, but together they create a pre-market and lifecycle burden precisely for the companies Europe most needs to create.

Why Is the Anthropic Suspension a Wake-Up Call for Europe?

The U.S. government directive that forced Anthropic to suspend access to its Fable 5 and Mythos 5 models to foreign nationals, including foreign employees, illustrates a hard truth: frontier AI is no longer just software. It is strategic infrastructure, industrial capacity, scientific acceleration, medical capability, military relevance, and economic sovereignty. If access to advanced AI systems can be withdrawn by government order, Europe cannot afford to remain merely a customer of American AI systems.

European researchers, engineers, physicians, companies, and institutions that depend on systems trained, hosted, controlled, and politically governed elsewhere do not control their own technological future. They rent it until someone else changes the terms.

Does Europe Have the Ingredients to Build Competitive AI?

Europe does not lack the raw materials for AI leadership. The continent has world-class scientists, engineers, universities, hospitals, research institutes, industrial companies, and public datasets of significant value. Many intellectual foundations of modern artificial intelligence, including optimization, medical imaging, robotics, speech processing, and interpretable machine learning, have deep European roots. Europe also has a strong ethical tradition insisting that technology serve human dignity.

The strategic error, analysts argue, is that Europe has chosen to regulate the race car before building the engine. By attaching heavy compliance burdens to the exact scale of operation necessary for global competitiveness, the EU AI Act may inadvertently ensure that frontier AI development happens elsewhere, leaving Europe dependent on foreign systems for decades to come.

What Should Policymakers Consider Moving Forward?

The core tension is between two legitimate goals: ensuring AI safety and enabling European innovation. The current regulatory approach prioritizes safety documentation at the expense of industrial capacity. A different strategy might focus on building European AI champions first, then refining governance as the industry matures. This would require rethinking the compute threshold, streamlining compliance for early-stage frontier developers, and creating incentives for scale rather than penalties.

Without such a shift, Europe risks becoming a museum of AI regulation, with detailed rules for an industry that has not yet been built on European soil.