Position as of the close of 21 June 2026. Anthropic was still listing Fable 5 and Mythos 5 as unavailable. Bloomberg News reported that Dragos and Cisco retained access to Mythos Preview; Reuters carried that account. ENISA, invited to join Glasswing before the block, was told it would not receive access, though it remained in talks with Anthropic (Anthropic, 9 June 2026; Reuters, 19 June 2026).

On 17 June, with Fable 5 and Mythos 5 still suspended, the dispute over those models reached the G7 table at Évian-les-Bains. Leaders of the main Western economies shared a working lunch with executives from Anthropic, OpenAI and Google DeepMind. In parallel, they were negotiating a possible trusted partners scheme that would allow certain non-US countries or companies to access advanced models developed in the United States (Reuters, 17 June 2026).

Emmanuel Macron framed the problem in economic terms. No one will want to buy American AI if it can be switched off from one day to the next. He was not speaking only of a service outage. He was pointing to a dependency: whoever buys an API, a hosted model or compute capacity may not be buying technological autonomy at all, but a licence revocable from another jurisdiction.

The scene helps place Anthropic’s suspension within a wider frame. Frontier AI is no longer merely a digital commodity or a research tool. It is strategic infrastructure, a corporate asset, a legal object and a diplomatic resource. The frontier does not lie solely inside the model. It is administered through export rules, applied through private platforms, and renegotiated between governments, laboratories, security alliances and financial markets.

A hierarchy of access before the cut-off

On 12 June, a US directive ordered the suspension of access to Fable 5 and Mythos 5 for any foreign national, inside or outside the United States. Anthropic responded by disabling both models for its entire customer base. The rest of its catalogue, including Opus 4.8, was unaffected (Anthropic, 12 June 2026).

The news could be read as the abrupt shutdown of a global AI. The chronology tells a different story. On 7 April, Anthropic unveiled Project Glasswing and reserved Mythos Preview for a first network of organisations dedicated to defending critical software and infrastructure (Anthropic, 7 April 2026). On 2 June, the programme expanded to roughly 150 new organisations from more than fifteen countries. Together with the initial cohort, the entities potentially brought in numbered around two hundred, though each had to clear specific security requirements (Anthropic, 2 June 2026).

Fable 5 and Mythos 5 arrived on 9 June. Fable was offered for general use, but with safeguards that diverted certain sensitive queries towards Opus 4.8. Mythos 5 shared the same underlying model, but lifted part of those safeguards for organisations already enrolled in Glasswing, as an upgrade to Mythos Preview (Anthropic, 9 June 2026).

There was, then, no single AI circulating through a global market. There was a scale of access: Preview for selected organisations; a public version limited by safeguards; another with expanded capabilities for trusted networks; and finally a state-imposed restriction defined by nationality. The situation after 12 June confirmed that this taxonomy was not merely technical (Reuters, 19 June 2026).

The G7 did not abolish that hierarchy. It moved it into diplomacy. The trusted partners regime did not promise to universalise access, but to organise exceptions for those Washington was willing to certify as reliable allies (Reuters, 17 June 2026).

Sovereignty is also an economic relationship

Karl Polanyi does not allow us to call an AI model a “fictitious commodity” without qualification: he reserved that category for land, labour and money. His relevant contribution lies elsewhere. He questions the idea of an autonomous economy, detached from the institutions that make the market possible. An AI model does not reach its users through the price logic of an API alone. It depends on chips, energy, data centres, intellectual property, investment, export rules, regulatory bodies and diplomatic alliances.

The formulation of Iain Hardie and Donald MacKenzie helps to sharpen the point further. An economic actor is neither an isolated company nor an isolated state, but an assemblage of people, technical systems, devices and algorithms. Seen this way, the actor governing access to Mythos is neither Anthropic alone nor the Department of Commerce alone. It is the assemblage of laboratories, clouds, models, identity documents, chips, rules, compliance teams, security bodies and diplomacies that turns a computing capability into an unevenly distributed resource.

Technological sovereignty is therefore not the same as manufacturing everything within one’s borders. It means retaining the capacity to choose, repair, modify and negotiate when a supplier or a government changes the rules.

Digital sovereignty is jurisdiction over data

The suspension of Fable and Mythos reveals an access frontier defined by nationality and export. But the frontiers of AI also run through data: where it is stored, who can demand it, which transfers are authorised, and which body decides whether a piece of infrastructure is critical.

Anthropic’s policy already disclosed that dimension before the order. Published on 8 June and due to take effect on 8 July, it allows the company to request, in certain circumstances, age or identity verification. Should a person choose to comply, the policy provides for the collection of an official document, images or video, and facial-geometry templates (Anthropic, Privacy Policy). Its policy for Mythos-class models also introduced thirty days of traffic retention to detect sophisticated attacks and false positives (Anthropic, 9 June 2026).

China organises this question through a different legal architecture. Its Personal Information Protection Law also applies, in certain cases, to companies located outside the country that offer services to, or analyse the activities of, people within China. To transfer personal information beyond its borders it requires conditions such as a security assessment, certification or a standard contract; for certain critical operators or large-scale handlers it requires domestic storage and an assessment before any export (DigiChina, Stanford).

This does not automatically turn the Chinese model into a guarantee of individual liberty. Its logic combines data protection, national security and the state’s capacity for oversight. But it does force us to view digital sovereignty as a conflict over jurisdiction, not merely as a dispute over owning models. The question becomes who decides over data flows, what rights the people affected hold, and which authority can compel a foreign platform to submit to local rules.

Europe is not merely an ally asking for access

Europe appears at Évian as part of a group seeking guarantees against dependence on the United States. But it is not only an ally hoping to enter the circle of trust. It is also trying to exercise regulatory, industrial and market power.

The European Union’s AI Act places general-purpose models under obligations of governance, transparency and safety, supervised by the European AI Office and national authorities. The obligations for general-purpose models already apply; the Commission continues to roll out the timetable for the rest of the framework (European Commission, AI Act; European Commission, AI Act governance).

The EU likewise uses the size of its market to regulate the platforms that dominate its digital infrastructure. In April 2025, the Commission fined Apple and Meta for breaches of the Digital Markets Act: not only on competition grounds, but also over conditions of choice, distribution and the use of personal data (European Commission, DMA).

The European paradox is plain to see. It wants access to frontier capability, yet fears that such access will create a dependency that erodes its autonomy. That is why, in June 2026, the Commission proposed a tech sovereignty package covering chips, cloud, AI and open source; the CADA proposal includes a common sovereignty assessment for cloud and AI. Days later, it selected a consortium to develop an open-source European frontier model in the Union’s twenty-four official languages (European Commission, tech sovereignty package; European Commission, CADA; European Commission, EUROPA consortium).

Europe is not outside the game of dependencies. It is trying to turn regulation, infrastructure and open access into instruments for reducing them. Its position is not that of a fully autonomous third bloc, but that of a regulatory power seeking to become a material power as well.

China, open weights and material bottlenecks

The Chinese reading of the case cannot be reduced to a defensive reaction. In its political-economic vocabulary, AI innovation is part of a strategy of endogenous development and technological autonomy. On the international stage, Beijing has proposed a global AI cooperation organisation and argued that the technology should not become the exclusive preserve of a few powers or companies (The Guardian, 26 July 2025).

That discourse must be read with double attention. It may open spaces for cooperation, shared standards and access for countries absent from the Évian negotiations. It is also a narrative with its own interests: it casts China as a supplier of infrastructure, models, standards and alliances.

Open weights widen the room for manoeuvre, though they do not abolish dependency. The public Qwen3-Coder repository makes it possible to examine, deploy and adapt a model without seeking permission from a US API (Qwen Team).

Open source, however, is not the same as autonomous infrastructure. Running and adapting models still requires stable electricity, data centres, chips, networks, situated data, technical staff and the legal capacity to decide where the data resides. Releasing weights can shift the frontier; it does not make it disappear.

Nor should the dispute over critical minerals and the dispute over models be treated as a simple symmetry. They are different bottlenecks. The International Energy Agency stresses that critical minerals concentrate risks of supply, processing and market power, and that the responses call for diversification, investment and long timeframes (IEA).

The asymmetry matters. A hosted model can be lost through an immediate order; a materials chain can be reconfigured, but not overnight. In both cases the dependency is material. What changes is the tempo, the assets and the prospects for exit.

Geopolitics is also narrated

The frontier is not negotiated solely in contracts, data centres and summits. It is also contested in language.

The United States presents access restrictions as a national-security requirement. At Évian, the vocabulary of trusted partners transforms a relationship of dependence into a selective guarantee. China presents global cooperation, the opening of models and support for the Global South as a response to technological monopoly. Europe presents regulation, rights and strategic autonomy as alternatives to a competition without rules (Reuters, 17 June 2026; The Guardian, 26 July 2025; European Commission, AI Act).

None of these formulas is a neutral description of the world. Each organises who appears as a risk, who as an ally, which kind of dependency is deemed legitimate, and which investments seem necessary. The geopolitics of AI is also a geopolitics of narrative.

Abner Cohen helps us read this dimension. He analysed how ruling groups validate their position by presenting their qualities as rare, exclusive and necessary to society as a whole; in formally egalitarian societies, those qualities can be objectified through ambiguous symbols and dramatic performances.

The G7 lunch proves no conspiracy, nor does it turn corporate executives into heads of state. But it does reveal a dramaturgy of authority. The people running laboratories able to open or close particular models take part in the space where states discuss the rules of their circulation. The rules may be public; the capabilities, the infrastructure and much of the information needed to apply them remain concentrated in private companies.

The Global South does not fit in the margins

The answer cannot consist of simply choosing between a US API, a European provider or a Chinese ecosystem. For Latin America, Africa, South Asia and South-East Asia, the question is whether they can build local capacity to operate, audit, adapt and govern AI systems.

A broad group of African researchers has set out a specific agenda: data dependency, scarcity of compute, labour risks, environmental costs and the absence of effective participation in global governance. Their proposal is not to ask for a charitable exception, but to build situated institutional, linguistic, educational and security capabilities (Segun et al., 2025).

Évian displayed a concrete geography of concentration and permission. Shanghai proposes a different grammar of cooperation, open to scrutiny. Brussels is trying to turn law and infrastructure into the capacity to decide. None of these paths resolves the problem on its own.

The anthropological question is not who holds a more benevolent key. It is who can take part in making the keys, decide which doors they open, and retain the capacity to change them.

Sources and further reading