According to Anthropic’s statement on the Fable 5 and Mythos 5 suspension, the US government ordered the company to suspend foreign-national access to both models, including access by foreign-national Anthropic employees. Reuters framed the order as an escalation from chip and tool export controls toward direct controls on access to frontier AI systems.
The suspension proves the sovereignty vulnerability while leaving the exceptionalism claim weak. Anthropic says the government gave no specific technical details, that the relevant jailbreak was narrow and non-universal, and that similar vulnerability-finding behaviour exists in rival public models. The stronger conclusion is not that Anthropic alone crossed a forbidden line, but that frontier agentic systems are now close enough in operational capability for access control to become a discretionary state instrument.
The benchmark picture supports that interpretation. In Artificial Analysis’ direct Claude Code versus Codex comparison, Claude Code with Fable 5 narrowly leads Codex with GPT-5.5 on DeepSWE, 66% to 64%, but Codex with GPT-5.5 leads on Terminal-Bench v2, 84% to 82%, runs at $5.07 per task versus $11.75, takes 10.1 minutes versus 23.5, and uses 12.3M tokens versus 14.1M. On Agents’ Last Exam, GPT-5.5 wins more clearly: Codex with GPT-5.5 reaches 24.0% overall pass rate, ALE-Claw with GPT-5.5 reaches 23.0%, and Claude Code with Fable 5 reaches 22.0%. These are not the numbers of a unique Anthropic security category. They are the numbers of a frontier capability cluster where the same state logic could just as easily be applied to GPT-5.5.
Sovereignty begins at parity
Europe’s sovereignty argument is now correct at the level of risk, but most European answers are still wrong at the level of execution. A good European model is not frontier sovereignty. A useful public-sector model is not frontier sovereignty. A model that is "good enough" for ordinary administrative workflows may be valuable for resilience, compliance, or industrial policy, but it does not solve dependency at the frontier.
The dependency exposed by the Anthropic order exists at the top capability tier, so the answer must exist at the top capability tier. If Europe loses access to the leading US systems, the European substitute must operate at the same level on software engineering, scientific reasoning, cyber-defence, autonomous tool use, multimodal professional work, long-horizon task completion, and private adversarial evaluations. Anything below that is a regional fallback with sovereignty branding.
This is the parity veto. Europe can build useful AI below the frontier, and it should, but it should not call that frontier sovereignty. If the best European system is worse than the best American system, the highest-leverage users will still route serious work through American models whenever policy, procurement, or access allows it.
The deadline is 2029/2030
The serious deadline is not 2035 or 2040. Those dates belong to industrial-policy theatre because the operating layer will already have been installed. By then, workflows, tools, data surfaces, security patterns, procurement contracts, developer habits, and organisational assumptions will have hardened around whichever systems dominated during the transition.
The defensible target is frontier model parity in 2029 and operational parity by 2030. The model must arrive first because enterprises and governments do not adopt base models directly. They adopt coding agents, enterprise agents, RAG systems, connectors, deployment environments, audit trails, procurement vehicles, support structures, and trained internal operators.
The 2030 line is anchored in adoption, not vibes. The EU’s Digital Decade target for enterprise adoption of cloud, data analytics, or AI is 75% by 2030. Eurostat’s 2025 enterprise AI adoption data already puts Sweden at 35%, while Statistics Sweden’s 2025 AI survey reports AI use by 72% of large Swedish enterprises. The Swedish government’s plan for a national AI workshop for public administration begins in 2026 and is supposed to be fully operational by 2030.
The labour-market clock points to the same window. The World Economic Forum’s Future of Jobs Report 2025 frames 2025-2030 as the transition period in which 22% of jobs are structurally disrupted and nearly 40% of worker skills change. My coordination-shift timeline reaches the same endpoint from the software-work side: practitioner ignition in 2025, organisational recognition in late 2025, public-sector forcing around 2029-2030, and post-shift normalisation around 2030-2031.
After 2030, catch-up becomes migration. Europe would not be choosing which AI substrate to adopt; it would be trying to replace a substrate already embedded in knowledge work, software production, procurement, public administration, security controls, and enterprise architecture. That is a radically harder problem than reaching parity before the operating layer settles.
The compute plan is too small
The EU’s current plan does not meet the parity test. The Commission’s AI Continent plan speaks of €200 billion for AI, €20 billion for up to five AI gigafactories, 19 AI factories, and a goal to triple EU data-centre capacity within five to seven years. The Commission’s AI Gigafactories timetable still has the formal call expected in summer 2026 and construction of the first gigafactory in 2027.
That is useful infrastructure, not a 2030 parity programme. Epoch AI’s analysis of frontier AI supercomputers estimates that the leading AI supercomputer in March 2025 already used 200,000 chips, cost about $7 billion in hardware, and required 300 MW. If the observed trend continues, Epoch projects that the leading system in June 2030 would require around 2 million chips, $200 billion in hardware, and 9 GW of power.
OpenAI’s Stargate capacity and investment plan points in the same direction: nearly 7 GW of planned capacity and more than $400 billion of investment over three years, on the path to a $500 billion, 10 GW commitment. Against that trajectory, a 100,000-processor European gigafactory is not a frontier-sovereignty asset. It is one cell in a much larger machine.
A credible European parity programme is closer to a €300-500 billion front-loaded effort than a €20 billion facility programme. The exact number is less important than the structure: chips, power, grid connection, data centres, cooling, networking, storage, security, failed runs, inference capacity, post-training, evals, productisation, and refresh. Most of the first-order capital goes into infrastructure, which is politically toxic because it means concentration rather than balanced distribution.
The grid is part of the model
Europe cannot treat energy as a separate policy file. Frontier AI is power, chips, networking, cooling, storage, and operational discipline converted into model capability. The strategic question is not whether Europe likes AI sovereignty, but whether it can connect gigawatt-scale power to a small number of training sites before the operating layer is already American.
The constraint is already visible. The IEA’s analysis of European data-centre energy constraints says EU grid-connection wait times can range from two to ten years depending on country and region, with major data-centre hubs such as Frankfurt, London, Amsterdam, Paris, and Dublin averaging seven to ten years. Reuters has also reported that European grid delays are already shaping data-centre geography, with some transmission connections taking up to seven years while the data centre itself can be built far faster.
This is where the Green Deal analogy becomes useful and dangerous. The useful overlap is infrastructure: power, grid, permitting, cooling, land, transmission, nuclear, hydro, storage, and industrial electricity contracts. The dangerous overlap is the European habit of translating hard technological constraints into balanced programmes, stakeholder alignment, regional benefit allocation, and long timelines.
The model is not the operating system
Even if Europe trains a frontier model by 2029, sovereignty does not exist until that model becomes an operational system. The US labs are not only selling models; they are building operator layers for enterprise work. OpenAI’s Deployment Company is designed to embed engineers with enterprises, redesign workflows, and connect AI systems to customer data, tools, controls, and business processes. OpenAI’s Frontier enterprise platform is built around secure, production-ready AI agents integrated with systems of record.
Anthropic is moving in the same direction. Claude Enterprise bundles governance, data controls, and admin infrastructure for organisation-wide deployment. Anthropic’s Managed Agents architecture treats agents as long-running execution systems with sessions, harnesses, sandboxes, tools, and enterprise environments. The serious layer is not chatbot access; it is the model, the harness, the deployment pattern, the control plane, and the transformation labour around it.
Europe therefore needs more than AI researchers and data-centre engineers. It needs developer-tool builders, enterprise platform engineers, security engineers, product designers, eval engineers, compiler specialists, RAG and search specialists, identity experts, forward-deployed engineers, procurement operators, and support organisations that can make frontier capability usable inside real firms. The lab is necessary, but the operator stack decides adoption.
Inferior systems create inferior feedback
The buyer problem is decisive. Who buys the European system if it is worse? Governments and regulated firms may buy it for compliance, resilience, or political reasons, while the highest-leverage users keep using better US systems whenever they can.
That creates a weak feedback loop. Captive customers tolerate inferior performance because policy instructs them to tolerate it. Vendors learn to satisfy procurement language. The product becomes sovereign in documentation while losing the frontier in practice.
A serious sovereignty programme needs buyers who can reject the product. European software firms, banks, defence suppliers, energy companies, manufacturers, telecoms, public agencies, hospitals, and universities should be anchor users, but they must bring hard tasks, hard datasets, hard acceptance criteria, and hard deployment constraints. A customer who cannot say no is not a customer; it is a subsidy channel.
The real choice
Europe’s choice is not between dependency and a comfortable sovereign alternative. It is between dependency and an uncomfortable level of concentration. Europe would need to commit money before the politics feels ready, pick sites before every region is satisfied, build grid and compute before the committee cycle completes, fund the operator layer instead of only research, and judge the result against US frontier systems rather than internal milestones.
The European Court of Auditors’ special report on EU AI ambition already called for stronger governance and more focused investment. Frontier AI turns that bureaucratic warning into a strategic deadline. If Europe does not have frontier model parity in 2029 and operational parity by 2030, the dependency gap will widen through integrations, data gravity, developer habits, procurement defaults, enterprise agents, and operating-model redesign.
The Anthropic order proves the access risk. The benchmark data shows that the risk class is not uniquely Anthropic’s. The infrastructure data shows that Europe’s current plan is too small and too slow. The enterprise product layer shows that model parity is only the first gate. Everything below parity is dependency with better stationery.