The Jurisdiction Problem: Could US AI Companies Leave America?

US AI companies could theoretically relocate abroad to escape domestic regulatory pressure, but doing so would be extraordinarily difficult in practice. Federal export controls, ITAR restrictions, and existing contracts with US government agencies would follow the technology regardless of corporate headquarters. Talent, capital markets, and critical infrastructure remain deeply embedded in the United States, making genuine jurisdictional escape largely illusory.

Published June 13, 2026

1. What Happened Yesterday

At 5:21 PM Eastern Time on June 12, 2026, Anthropic received a letter from Commerce Secretary Howard Lutnick ordering the company to suspend all access to its two most capable models, Fable 5 and Mythos 5, for any foreign national anywhere in the world. That definition of “foreign national” explicitly included Anthropic’s own non-citizen employees working inside the United States.

Because Anthropic cannot reliably separate foreign nationals from the rest of its user base in real time, it had no practical option except to pull both models offline for every customer on Earth. The company called it a misunderstanding, said it was working to restore access, and apologised to its customers. The Commerce Department declined to comment.

The stated trigger was a claim from another company that it had jailbroken Mythos. Anthropic disputed the severity of the jailbreak, arguing it was narrow and that comparable capabilities could be elicited from other publicly available frontier models including those it was not being asked to restrict. The government disagreed. The directive stood.

This is not the beginning of Anthropic’s troubles with the current administration. In February 2026, the company refused to remove safety guardrails that limited its models from being used for mass domestic surveillance and fully autonomous weapons systems. The Pentagon responded by designating Anthropic a supply chain risk. The administration ordered federal agencies to stop using Anthropic technology. A federal judge later blocked that move on First Amendment grounds, but the damage to the commercial relationship was done. By June, the DoD had formalised contracts with OpenAI, Google, and Microsoft as replacements.

So the question being asked in every AI leadership team this weekend is not hypothetical. It is operational: what does this mean for us?

2. The Question Nobody Wanted to Ask Out Loud

For the past three years, the received wisdom in the AI industry was that the United States was the only serious jurisdiction for building frontier AI. The capital was there. The talent was there. The compute infrastructure was there. The regulatory environment was light. The military and intelligence spending was enormous and growing.

That calculus has not reversed overnight. But yesterday’s order introduced a variable that no frontier AI company can now ignore: the US government has demonstrated that it is willing to reach inside a private company’s customer relationships and shut off access to its products, globally, without notice, without specific justification, and without a clear path to restoration.

The mechanism used was export controls, a regulatory instrument typically applied to physical goods, semiconductors, and weapons systems. Applying it to a software inference endpoint is genuinely novel. It establishes that the most advanced AI models are now legally equivalent to dual-use defence hardware. Every frontier AI company operating under US jurisdiction just had that precedent set.

The question is no longer whether governments will treat AI as a strategic national asset. That question was settled in 2024. The question now is whether the country that hosts the world’s most capable AI labs will use that position to restrict, direct, or weaponise their output, and what happens to those labs if they resist.

3. What Leaving Would Actually Mean

Let us be precise about what “leaving the US” would mean for a company like Anthropic or OpenAI, because the phrase covers several distinct scenarios with very different implications.

The first and most limited scenario is legal reincorporation. A company could establish a parent entity in a jurisdiction with more favourable legal treatment, while keeping its operational footprint substantially in the United States. This is a tax and governance play, not an operational one. It does not help when a Commerce Department letter arrives because the engineers, the servers, and the models remain subject to US law regardless of where the holding company is registered.

The second scenario is compute decentralisation. A company could establish data centre capacity and inference infrastructure in multiple sovereign jurisdictions so that no single government directive could take the full product offline. This is technically achievable and commercially sensible as a resilience strategy regardless of political risk. The challenge is that the most capable training runs require the kind of contiguous, high-bandwidth GPU cluster density that currently only exists in a handful of locations, most of them in the United States.

The third and most consequential scenario is genuine headquarters relocation, taking the research organisation, the training infrastructure, and the leadership out of American jurisdiction. This is what people mean when they talk about “leaving the US” in any meaningful sense.

That third scenario faces enormous practical obstacles. The US controls export of the most advanced AI chips. NVIDIA H100 and Blackwell hardware is subject to export licensing. A company that relocates its training infrastructure outside the United States would need to source compute from a jurisdiction that can access that hardware, which largely means it would still be dependent on the goodwill of the US government for its most critical input. You cannot escape American jurisdiction by moving your offices if your GPUs still require an export licence.

There are scenarios where this constraint loosens. If a European or Asian semiconductor ecosystem matures to produce competitive training accelerators, the chip dependency weakens. If open-weight models at frontier capability become the norm, the need for centralised training infrastructure diminishes. Neither of those is the world we are in today.

4. Where Would They Even Go?

The honest answer is that no jurisdiction currently offers everything a frontier AI lab needs.

The United Kingdom is making the most aggressive play. The government launched a £500 million Sovereign AI Fund in April 2026, operating at venture capital speed rather than bureaucratic speed. It offers GPU hours from the national AI Research Resource supercomputer network, fast-tracked visas for technical talent, and a regulatory posture deliberately calibrated to attract companies that find the European AI Act too constraining. Prime Minister Starmer has used the phrase “AI maker, not AI taker” repeatedly, and the policy architecture behind it is more coherent than the slogan suggests. UK AI startups raised over £3 billion in the first quarter of 2026 alone, on pace to shatter the 2025 record of £6 billion for the full year.

But the UK has a real problem: energy costs. Research published in April 2026 showed that UK electricity prices are running at roughly four times the US equivalent for industrial loads. The most compute-intensive AI workloads are the most sensitive to that differential. The government’s AI sovereignty ambitions and its power infrastructure are in direct contradiction, and the grid investment needed to close that gap is measured in years, not months.

France is competing aggressively on capital, with €109 billion in AI commitments announced. Canada has talent depth and a more welcoming immigration posture than the US. Singapore has positioned itself as the neutral ground for AI companies navigating US-China tensions, with strong infrastructure and pragmatic regulation. The UAE has sovereign capital, proximity to emerging markets, and minimal regulatory friction.

The European Union as a bloc is a more complicated case. The AI Act is being progressively softened under competitive pressure, and France, Germany, and the UK have all pivoted from regulatory ambition to industrial policy. But the EU’s fundamental regulatory culture, its instinct toward precaution and fragmentation, makes it a difficult base for a frontier lab that needs to move fast and iterate aggressively.

5. The Anthropic Paradox

Anthropic’s situation has a specific irony that is worth dwelling on.

The company was founded explicitly on a mission of AI safety. Its refusal to remove guardrails for domestic surveillance and autonomous weapons was not commercial positioning. It was the mission. The administration’s response to that refusal was to treat the company’s safety commitments as a liability.

Anthropic now finds itself simultaneously on a Pentagon blacklist for being too safe to use in government weapons applications, and under a Commerce Department export restriction for producing models whose capabilities are deemed too dangerous for foreign nationals to access. The government considers Anthropic’s most powerful models too dangerous to let the world see and too cautious to give the Pentagon what it wants.

That is not a comfortable position for a company that is also preparing an IPO, that recently raised $65 billion at a valuation of $965 billion, and that has just had its two flagship frontier models taken offline at five hours notice.

The question for Anthropic’s board is not whether to leave the United States. It is whether the current regulatory trajectory makes the United States a structurally hostile environment for the kind of company Anthropic is trying to be.

6. The Structural Argument for Leaving

Set aside Anthropic’s specific situation and consider the structural question for any frontier AI lab.

The US government has now established four things in sequence. First, it will spend enormous sums to support AI development it controls. Second, it will shut out companies that refuse to make their models available for uses those companies consider unethical. Third, it will restrict access to frontier models if it believes those models create national security risk. Fourth, it will do all of this without specific justification, without notice, and without a clear appeals process.

A company operating under those conditions faces a fundamental principal-agent problem. Its principals are its customers and investors. Its agent relationship with the US government is involuntary and unilateral. Every product decision it makes is subject to override by a regulator that has no accountability to its commercial interests and limited accountability to its stated rationale.

For most industries, this kind of regulatory risk is manageable because the regulatory interventions are predictable, bounded by statutory frameworks, and subject to judicial review. Export controls applied to software inference endpoints are none of those things. The precedent set yesterday is that the government can reach into a live commercial product and turn it off globally, today, without meaningful notice or specific justification.

If you are running a frontier AI lab and you are thinking clearly about risk, you are now asking whether you can build a durable business in a jurisdiction where that is the regulatory baseline.

7. The Structural Argument for Staying

The counterarguments are substantial and should not be dismissed.

The US compute ecosystem remains unmatched. The concentration of GPU capacity, high bandwidth networking, and the engineering talent to run it at scale is extraordinary and not replicable quickly elsewhere. A frontier lab that left the US tomorrow would spend years rebuilding infrastructure that it currently takes for granted.

The US capital markets are the deepest in the world for technology investment. Anthropic’s IPO preparations are proceeding. The venture and growth equity ecosystem that funds AI at scale is predominantly American. Moving headquarters to the UK or Canada would not stop American capital from investing, but it would complicate the regulatory and governance structures significantly.

The talent concentration in San Francisco, Seattle, and New York is still without parallel. The density of people who have trained frontier models, who understand RLHF and Constitutional AI at a research level, and who are available to hire is highest in the United States.

And there is the deeper point: the US government’s behaviour toward Anthropic is partly a function of this particular administration’s posture toward companies that publicly resist its priorities. Administrations change. The midterms are shaping up to be contested in part on AI-related spending. A different political environment in 2027 or 2029 might look very different.

8. What This Means for the Rest of the World

For countries outside the United States, yesterday’s order was a data point that validates a concern that has been building for two years: AI capability developed inside American jurisdiction is subject to American foreign policy priorities, regardless of what your commercial relationship with the provider says.

The European Central Bank had already praised Anthropic for limiting access to Mythos. European banks that were not given access to Mythos watched as Mistral began developing its own frontier model. That dynamic accelerates now. If Fable 5 and Mythos 5 remain restricted, every non-American enterprise that was planning to build on those models has to recalibrate.

Gartner data shows 47% of Western European enterprises were already actively reevaluating their dependencies on non-European cloud infrastructure before yesterday. That number will move.

For emerging markets, the implications are starker. Countries building national AI strategies on access to US-hosted frontier models now have evidence that access is a foreign policy instrument, not a commercial service. The conversation about sovereign AI is no longer theoretical. It is operational.

9. The Question That Matters Most

The question for AI companies is not really about geography. It is about governance.

What any serious AI lab needs is a stable, predictable, legitimate relationship with its regulatory environment. It needs to know what the rules are, that those rules will not change without process, and that if the government wants something from it, there is a negotiation rather than a directive.

The United States, for all its advantages, is currently failing to provide that. Not because of bad faith, but because the regulatory frameworks for AI simply do not exist at the sophistication needed. Export control law was written for semiconductor chips and missile guidance systems. Applying it to a software model at five hours notice is not a reflection of malice. It is a reflection of institutions trying to govern something they do not fully understand with tools designed for a different era.

That is fixable. Whether it will be fixed quickly enough to matter is a different question.

10. What Anthropic Should Do

Anthropic is not going to leave the United States. The IPO is too close, the capital relationships are too American, and the compute infrastructure is too deeply rooted. But it should do several things in the wake of yesterday.

It should publish its compliance posture clearly and publicly. Its customers need to understand what kinds of government directives it will comply with, how quickly, and with what transparency. The opacity of yesterday’s shutdown was almost as damaging as the shutdown itself.

It should accelerate its infrastructure decentralisation. Running all inference through a single jurisdiction is now clearly a concentration risk. Distributing inference capacity across multiple sovereign regions does not solve the problem of government directives but it reduces the blast radius of any single one.

It should work with peers to establish industry-level frameworks for how government directives should be handled. The individual company negotiating with a Commerce Department letter at 5:21 PM on a Friday is in a structurally weak position. An industry with established protocols, legal frameworks, and transparency mechanisms is in a much stronger one.

And it should continue to refuse to compromise its safety principles. That is the one thing it cannot negotiate away without becoming a different company.

11. The Longer Arc

The era of AI companies operating as pure commercial entities, accountable only to their investors and customers, is over. That was always going to be true once the capability crossed the threshold of genuine strategic significance. Mythos class AI has crossed that threshold. The governments of the world know it.

What is being negotiated right now, messily and without adequate frameworks, is the governance structure for the most powerful technology that has ever existed. That negotiation is happening in the US, in the EU, in the UK, in China, in Singapore, and in dozens of other jurisdictions simultaneously.

The Anthropic situation is a single data point in that negotiation. A significant one, because it establishes a precedent for export controls applied to AI inference. But a single data point.

The companies that navigate this era successfully will be the ones that build robust relationships with multiple sovereign jurisdictions, invest in regulatory engagement as a core capability, and maintain enough operational independence that no single government directive can take their entire product offline at five hours notice.

That is a very different operating model from the one that built the frontier AI industry. Whether the companies currently at the frontier can make that transition quickly enough is the real question.

Not whether they leave the United States. But whether they outgrow the assumption that being American was enough.

Andrew Baker is Group CIO at Capitec Bank. He writes about AI strategy, cloud infrastructure, and financial technology at andrewbaker.ninja and on Substack at @futureherman. Views expressed are his own.