The Trump administration's order blocking foreign access to Anthropic's Fable and Mythos models signals a structural shift in how AI capability is weaponised as geopolitical leverage.

Key Highlights

  • The US government demonstrated it can unilaterally restrict global access to frontier AI models, a power with no clear legal precedent.
  • Anthropic's Fable and Mythos represent a new tier of AI capability, built around cybersecurity exploitation at scale.
  • A tiered access hierarchy is forming: full capability for US defence, diluted versions for allies, safety-wrapped models for the world.
  • Non-American economies face a structural compute deficit, with the US holding roughly 15 times Europe's data centre capacity.
  • The geopolitical calculus now extends beyond trade, alliances, and currency to include who controls the most capable AI infrastructure.

The Lever Nobody Saw Coming

When the Trump administration ordered Anthropic to block foreign users from accessing its Fable and Mythos models on June 12th, the immediate reaction focused on the legal ambiguity of the order. That is the wrong frame. The more consequential signal was structural: a single executive decision in Washington can now determine who on earth accesses the most capable artificial intelligence ever built. That is not a policy footnote. It is a reconfiguration of global power.

The episode exposed something that had been building quietly since Anthropic's Mythos model leaked in March. Mythos, designed primarily to identify and exploit software vulnerabilities at a level no prior model had achieved, was never a product in the conventional sense. It was a capability threshold. When Anthropic launched Claude Fable 5 on June 9th as a safety-wrapped public version of Mythos-class intelligence, it was making a calculation: broad deployment with classifiers and hard blocks on high-risk domains, rather than concentrated access without them. The administration's response, three days later, suggested Washington had its own calculation to make.

A Hierarchy Is Already Forming

The analogy to military technology is uncomfortable but instructive. The United States kept the F-22 entirely for itself while offering allies the F-35, a capable but deliberately constrained alternative. Nuclear cooperation with Britain resumed only after Britain developed independent capability; with everyone else, the Non-Proliferation Treaty held the line. Cryptography export controls eventually collapsed under the weight of mathematical inevitability. Frontier AI sits at the intersection of all three precedents simultaneously.

What is forming now is a three-tier structure. The most capable models, specifically Mythos 5 and successors, remain restricted to a small consortium of vetted organisations under Project Glasswing, which itself operates under a degree of government visibility. The next tier is Fable-class: Mythos-equivalent reasoning and coding ability, but with safety classifiers that block responses in cybersecurity, biology, chemistry, and related high-risk domains. The third tier is everything else, the broader market of models with progressively more conservative safety architectures.

The practical implication is that America's offensive cyber capability, already formidable, is being augmented by an AI layer that allies cannot fully access and adversaries cannot currently replicate. If a model at version 5 can identify one more critical vulnerability than a model at version 4, and the attacker holds the better model, the asymmetry is not incremental. It is decisive.

The Compute Gap Compounds the Problem

Model access is only part of the equation. Running frontier AI at scale requires compute infrastructure, and the distribution of that infrastructure is not neutral. The United States holds approximately 15 times Europe's data centre capacity and continues to attract the majority of global AI infrastructure investment. Anthropic's own IPO trajectory, alongside OpenAI and SpaceX, is concentrating capital further into the American AI ecosystem.

Europe's position is particularly exposed. In the seven months to June 2025, demand for new grid connections in Britain rose from 41 gigawatts to 125 gigawatts, exceeding twice the country's peak power demand. Yet OpenAI paused a data centre project in Britain in April, citing regulatory friction and energy costs. The infrastructure bottleneck is a policy failure as much as a market outcome.

East Asia presents a marginally stronger position. Taiwan, Japan, and South Korea each hold critical nodes in the global semiconductor and fabrication supply chain. Dutch lithography machines from ASML remain indispensable to any advanced chip production, including American chip production. These dependencies are genuinely bilateral in ways that European AI dependence is not.

What the Rest of the World Does Now

The instinct in many capitals will be to treat this as a trade or alliance problem and route it through existing multilateral mechanisms. That instinct will not be sufficient. The WTO has no framework for AI model access. NATO has no protocol for frontier model sharing. The response has to be structural.

The most immediate lever available to non-American economies is infrastructure investment. Compute is scarce enough globally that data centres can be built almost anywhere and attract significant economic activity. But this requires energy availability, planning reform, and regulatory environments that do not treat AI infrastructure as a compliance liability. Few governments have treated these preconditions with the urgency they warrant.

Open-weight models offer a partial alternative. Models that anyone can download and run locally are structurally resistant to the kind of access restriction the Trump administration exercised on June 12th. Their capabilities continue to advance. But open-weight alternatives currently lag Mythos-class models in the specific domains, particularly offensive cybersecurity, where the capability gap matters most. That lag may not persist indefinitely, but it is real today.

A New Axis of Dependence

For decades, American power over allied economies operated through familiar channels: dollar-denominated trade, military guarantees, technology licensing. AI is adding a new axis to that dependence structure, one that is harder to hedge, faster-moving, and less governed by existing international frameworks.

The June 12th order may not last. Courts may constrain it. Anthropic has argued the underlying jailbreak concern was overstated. But the demonstration has been made. The question of who accesses the world's most capable AI is now, demonstrably, a question that can be answered in the Oval Office. That is the structural reality every capital outside Washington now has to plan around.

This article is for informational and analytical purposes only. It does not constitute financial or investment advice.