In June, a foreign government removed access to the most capable AI models on the market from every user in Europe, in under four hours, with no notice and no appeal. The question for boards is not whether it could happen. It is what happens the next time it lasts longer. Part 1 of two.
On the afternoon of Friday 12 June 2026, at 5:21pm Eastern time, the United States Department of Commerce sent Anthropic a directive. It came from the Bureau of Industry and Security under the signature of Commerce Secretary Howard Lutnick, it invoked national security authorities under the Export Administration Regulations, and it ordered the company to suspend access to its two most capable models, Claude Fable 5 and Claude Mythos 5, for any foreign national, whether located outside the United States or inside it, explicitly including Anthropic’s own non-citizen employees.
A consumer and enterprise platform cannot sort users by passport in real time. So Anthropic disabled both models globally rather than attempt selective enforcement. The models had launched three days earlier. Within hours of the directive they were gone, for everyone, everywhere.
European enterprise customers in finance, healthcare, software and critical infrastructure found workflows they had built around those models simply stopped working. There was no notice, no exception process, and no effective recourse. The blackout lasted 19 days. Commerce lifted the controls on 30 June and access was restored on 1 July.
Most European coverage treated this as a technology story: a dramatic launch, an abrupt suspension, a resolution within a month. That framing misses what actually occurred. For nineteen days, the world’s largest single market was denied access to a critical technology by the administrative act of a government it does not elect, for reasons it was not told, with no legal remedy available to any European company affected.
It resolved quickly. There is no structural reason the next one has to.
What Made This Different From an Outage
Boards are used to vendor risk. Suppliers fail, services degrade, contracts get renegotiated on unfavourable terms. Standard enterprise risk management handles this through redundancy, contractual remedy and exit planning. The June episode does not fit that framework, and understanding why is the whole point.
Three features set it apart.
It was based on nationality, not on conduct. No European company did anything wrong. No European government was in dispute with Washington. The trigger was a reported jailbreak of the models’ safety guardrails, a technical dispute between American parties. Anthropic characterised the underlying finding as narrow and not universal. European users were nonetheless cut off, because the directive was written around who they are rather than what they had done. There is no compliance posture, no contract clause and no amount of good behaviour that protects against a criterion you cannot change.
It was administrative, not commercial. A vendor dispute proceeds through negotiation, notice periods and courts. This proceeded through a letter. The time from directive to global unavailability was measured in hours. No European regulator was consulted, and none had standing to object.
There was an asymmetry in the restoration. On 26 June, access to Mythos 5 was partially restored for roughly 100 approved US companies and agencies while Fable 5 remained blocked. For a period, American critical infrastructure operators had frontier capability that European ones did not. Whatever the intent, the practical effect was a capability gap opened and closed by administrative decision in another jurisdiction.
The Cloud Security Alliance drew the conclusion most directly, advising that security and governance teams must now treat frontier AI model access as a contingent operational dependency, subject to abrupt administrative revocation. That phrase, contingent operational dependency, is the one worth carrying into a board meeting. It is not how most European companies currently classify their AI providers.
How Exposed Is Europe, Actually
The honest answer is: more than the conversation suggests, and in a way that compounds.
Europe controls less than 5 percent of global frontier-scale AI compute. US hyperscalers hold 70 to 72 percent of the European cloud market, and roughly 80 percent of EU digital infrastructure sits with non-EU providers. Private AI investment in the United States runs approximately 24 times European levels. The nineteen EuroHPC AI Factories, Europe’s flagship public compute programme, aggregate around 600 petaflops, which is meaningful and still materially smaller than a single US Stargate site.
The compounding is the part that matters for risk assessment. A European company running an AI workload is typically dependent at three separate layers: the model itself, usually from OpenAI, Anthropic or Google; the cloud on which it is served, usually AWS, Azure or Google Cloud; and the silicon underneath, almost universally NVIDIA. Each layer is separately subject to US jurisdiction, and a disruption at any one of them is sufficient. Redundancy at the model layer does not help if the constraint arrives at the compute layer.
There is a European alternative, and it is stronger than it was a year ago. Mistral reports roughly $1 billion in annual recurring revenue across more than 125 large enterprises including Airbus, ASML and HSBC, has just raised €3 billion for sovereign infrastructure, and ASML has planned a €1.3 billion stake that would make it the largest shareholder. France chose Mistral over OpenAI for government cyber work following a tax agency breach affecting roughly 700,000 taxpayers. Germany has Aleph Alpha. Switzerland has Apertus.
Two qualifications keep this from being a solution. On frontier capability, Mistral Large 3 is not in the same class as Fable 5 or GPT-5.6 on every enterprise evaluation, which is precisely why European buyers reach for US APIs when the task is hard. And Mistral itself distributes through Azure, AWS Bedrock and Google Cloud Vertex, depending commercially on the incumbents it proposes to displace. That is not hypocrisy, it is the operating reality of bootstrapping against an entrenched stack, but it means the European option currently inherits part of the exposure it is meant to remove.
What Brussels Did
This is the part that should concern boards most, because it determines whether the exposure gets addressed at the policy level or stays entirely with individual companies.
The European Commission’s formal response was to take note of the US decision and begin assessing its implications and practical consequences for European users. A Commission spokesperson, asked about the episode, said that where Anthropic chooses to locate itself would be Anthropic’s decision, added that the Commission had not been left out and had done a good job discussing matters bilaterally with Anthropic and OpenAI, and offered that the EU believes itself to be a trusted partner rather than a security risk.
The one forceful institutional response came from Vienna. Austrian State Secretary for Digitalization Alexander Pröll wrote to Commission Executive Vice President Henna Virkkunen, and his framing was notably sharper than anything from Brussels. He observed that overnight, a single market of 450 million people had been cut off from a cutting-edge innovation, not by its own decision but by that of a foreign government, described it as a show of power and a reminder of how vulnerable Europe is when access to the key technology of the age can be withdrawn at the stroke of a pen, and asked whether Europeans are prepared to be the architects of their technological future or wish to remain mere administrators.
The diagnosis was correct. The proposed remedy was to invite Anthropic to establish itself within the European Union, offering legal certainty, market access, capital and aligned values. The letter contained no funding figure, no timeline and no operational plan, and Anthropic, which has committed roughly $50 billion to US infrastructure, did not publicly respond.
Set aside whether that specific proposal was realistic. Notice what it reveals about the reflex. Confronted with evidence that European access to critical technology depends on decisions made in Washington, the most forceful European response available was to ask an American company to move. Not to accelerate European capability, not to mandate continuity planning, not to change procurement rules. To extend an invitation.
Meanwhile the substantive conversation happened privately. Germany’s most senior cybersecurity official, Claudia Plattner, warned national lawmakers in a closed-door meeting that Chinese companies such as Alibaba could soon match the capabilities Europe had just been denied, leaving Europeans exposed. That is a serious strategic observation. It was made behind closed doors while the public response consisted of taking note.
And then the timing, which is almost too neat. The cut-off arrived days after the EU had announced a Tech Sovereignty Package. In the same week that a foreign government demonstrated it could switch off Europe’s access to frontier AI, the European Parliament was scheduled to vote on simplifying the bloc’s rules on artificial intelligence.
That juxtaposition captures the structural problem better than any argument could. Europe has built the world’s most developed regulatory apparatus for governing how AI systems behave: risk tiers, transparency obligations, prohibited practices, conformity assessments, fines reaching 7 percent of global turnover. It has built almost nothing governing whether Europe can continue to obtain those systems at all. The AI Act has a great deal to say about what a model may do to a European citizen. It has essentially nothing to say about what happens when the model is withdrawn.
Regulating conduct while ignoring continuity is a coherent policy only if supply is assumed. June demonstrated that it should not be.
The Case That This Is Overstated
A board paper that only argues one side is advocacy, so here is the strongest version of the opposite case, and parts of it are genuinely persuasive.
Commercial incentives cut hard against disruption. US providers earn substantial revenue in Europe and have no interest in demonstrating that their services are politically contingent. The June action was reversed in 19 days, partly because it was commercially and diplomatically costly, and a bipartisan group in the US House had demanded the legal basis for the controls before the reversal. The system corrected.
The action was also narrow. It targeted two specific models with specific claimed cyber capabilities, not a country and not the general availability of AI. Other Anthropic models remained online throughout. The precedent established is that a US administration can restrict a particular frontier model on national security grounds, which is a meaningfully smaller claim than the ability to cut a region off from AI generally.
And Europe’s more pressing problem may be the mirror image of this one. Only 20 percent of EU enterprises use AI, against roughly 50 percent in the United States. A continent that hesitates to adopt out of sovereignty anxiety will lose more economic ground than one that adopts widely and hedges intelligently. Sovereignty concerns can become a rationalisation for the caution that was already there, and Europe cannot afford that trade.
All of this is fair. None of it changes the underlying structure. The correct reading is not that Europe faces an imminent cut-off, because it probably does not. It is that European companies have been operating on an implicit assumption of guaranteed access, that assumption was tested in June and failed, and the failure was survivable only because it was short. The risk to price is not the 19-day version. It is the same mechanism applied for a quarter, during a trade dispute, or to a category of models rather than two of them.
What This Means for a Board
Three questions are worth putting on an agenda before the end of this quarter.
Which internal processes would stop working if a specific frontier model became unavailable within four hours, and how would anyone find out? Most organisations cannot answer this, because AI adoption has happened process by process rather than through a central register.
What is the actual fallback, and has it ever been tested? A fallback that exists as a paragraph in a vendor policy is not a fallback. The June episode gave every European company a free live-fire exercise. Very few appear to have run the retrospective.
And how much of the exposure is concentrated? A company using three US model providers has less redundancy than it thinks if all three are served on US cloud infrastructure running US silicon under the same jurisdiction.
Part 2 takes those questions seriously and works through what a credible response looks like, at the company level and at the policy level, including what European governments would actually need to do rather than what they have so far offered to do.
Key sources:
Tech Policy Press. “Did the US Government Just Set An AI Export Precedent by Blocking Mythos?”
Raconteur. “Mistral bets big on European sovereign AI,” on the hyperscaler distribution paradox.
Hosting Journalist. “Europe’s Mistral AI Secures €3B for Sovereign AI Infrastructure.”