Artificial intelligence is often presented as a sequence of models, products and technical improvements. But the signals of recent days point to a different scene. A government can intervene in the conditions of access to a model. A company can discuss bringing the state itself into its ownership structure. A federal department can reorganise universities and professional pathways around an AI-powered scientific infrastructure. And a country can try to keep a supplier out of strategic contracts without yet having an equivalent alternative.

These are not disconnected issues. They describe the same transformation. AI is no longer governed only through laws or codes of conduct. It is also governed by defining who gets access, with what training, through which contracts and under what ownership relationship with the state.

A border with a price tag

On 1 July, Anthropic restored global access to Claude Fable 5 after the United States lifted the export controls imposed in mid-June. The model circulated again, but not under exactly the same conditions for every person and organisation.

The company itself announced that, for Pro, Max and Team plans as well as certain Enterprise accounts, Fable 5 would be included up to 50 per cent of the weekly usage limit until 7 July. After that, it would be available through usage credits. Restoring access does not, therefore, close the episode of the border. It shifts it to another layer. Difference no longer operates only through nationality or jurisdiction. It is also expressed through the commercial architecture of the service.

These should not be confused. Paying for use is not the same as an export control. But both mechanisms can shape who has sustained access to a technical capability and who is left out when cost, administrative identity or security rules change. An AI interface is not only a screen. It is a customs gate made of permissions, classifications, limits and prices.

AIthropology Lab key point — Technological sovereignty does not begin only when a state blocks a model. It also begins when access is shaped by rules that users did not choose and can barely contest.

Training those who will be able to enter the platform

Genesis Mission, the United States Department of Energy initiative led by Darío Gil, offers a view of the same movement from another angle. Its public aim is not limited to linking supercomputing, artificial intelligence, quantum systems and scientific facilities. It also involves creating the human and institutional conditions required to use that infrastructure.

On 1 July, the DOE published its University Summit Report and the summary of its workforce development RFI. The associated analysis sets out the need to train 100,000 scientists and engineers over the next decade. Universities, curricula, intellectual property, interdisciplinary education, data practices and assessments of scientific merit appear as components of the same architecture.

The next verifiable moment will come on 17 July, when the Office of Science Advisory Committee meets in Washington. The committee has specific mandates on Genesis Mission, scientific facilities and quantum computing, with July 2026 as the horizon for the roadmaps requested by the DOE. There is still no public roadmap that would allow its content to be anticipated, but the timetable matters. It shows that the debate over scientific AI is becoming a debate over who will be trained, accredited and funded to take part in it.

22 July also warrants close attention. The executive order that launched Genesis Mission sets that date as the deadline for reviewing the capacity of laboratories and other federal facilities that could take part in AI-directed experimentation and manufacturing. In addition, Carl Coe, the DOE chief of staff, said in June that the department expected to present the mission’s first awards that day. This second expectation comes from a public statement reported by specialist press, not from an official DOE agenda or a published project list.

The question is not secondary. When a public platform defines the profiles it needs, it also sets which forms of knowledge will be legible as scientific contribution. Software, curated data and interdisciplinary coordination may gain recognition. But they may also be subordinated to metrics, national priorities and productivity languages set from above.

AIthropology Lab key point — A scientific infrastructure does not only process data. It organises careers, allocates resources and decides which forms of work will have a future within it.

The contract can also become ownership

While Anthropic negotiates the conditions of access to one of its models, OpenAI has discussed another form of relationship with the state. On 2 July, Reuters, citing the Financial Times, reported initial talks over the US government receiving a 5 per cent stake in the company. There is no public agreement or official decision. But the possibility is significant. The link between an AI company and the state would no longer be limited to licences, controls or procurement contracts. It could extend to ownership and the distribution of future benefits.

The proposal has emerged in a context of political pressure on advanced AI companies and debate over who should share in the wealth generated by these infrastructures. It should not be read as an automatic socialisation of their profits. For now, it is a political negotiation at an early stage. But it anticipates a change of scale. Regulation ceases to be merely an external force limiting the company. It can become a relationship of participation, incentive and shared responsibility.

In Spain, the debate over Palantir shows the reverse side of this politics. Several press reports published in recent days state that Moncloa has told public companies and companies with SEPI shareholdings to avoid new contracts with the US firm where they could compromise strategic information or national sovereignty.

The report must be read precisely. No general order has been published in the BOE, Spain’s Official State Gazette, nor is there a formal prohibition of universal scope. According to the available reporting, the instruction is not in writing. Nevertheless, this orientation is said to have stalled agreements under discussion with the Guardia Civil and Navantia. At the same time, the contract signed in 2023 between the Ministry of Defence and Palantir for military intelligence software remains in force.

This reveals a contradiction that should not be oversimplified. Removing or limiting a foreign technology may be presented as a gesture of strategic autonomy. But autonomy is not achieved simply by closing a door. It requires technical alternatives, public procurement capacity, specialist staff, auditing and time to build infrastructure that does not depend on a supplier that has already become central.

The decision over Palantir is not only about one company. It is about what happens when an analytics platform becomes so deeply embedded in security, defence and public administration that stopping its use becomes as difficult as starting it.

AIthropology Lab key point — Technological dependence is not measured only by the number of contracts. It is measured by the real difficulty of recovering decision-making capacity once a platform has become part of an institution.

Regulating from outside is not enough

The three cases point to the same trajectory. Anthropic shows that the circulation of a model can be shaped by geopolitical decisions and commercial conditions. Genesis Mission reveals that scientific AI requires the reshaping of institutions and the training of those who will operate within them. OpenAI opens the possibility that governance may reach company ownership itself. The case of Palantir in Spain exposes that sovereignty does not consist in declaring a preference, but in being able to sustain it technically and administratively.

AI governance does not happen in one law or one laboratory. It takes place in pricing plans, identity requirements, training calls, evaluation criteria, shareholdings, tenders, security agreements and purchasing decisions. At each of these points, decisions are made about who counts as a user, a worker, a legitimate supplier or a risk.

Looking at AI through anthropology requires a shift in the question. It is not enough to ask which model is more capable. We also need to ask who organises its conditions of access, which institutions become tied to its operation and what forms of autonomy can be sustained once the infrastructure has entered public life.

Sources