No AI sovereignty without infrastructure sovereignty

To understand where AI is headed, it helps to go back to the "Atoms for Peace" speech Eisenhower gave at the UN in 1953, and to the logic that later took shape in the International Atomic Energy Agency (IAEA).

Eisenhower wasn't speaking from a world where atomic power was still held by a single power. By 1953 the American monopoly had already broken. The Soviet Union had detonated its own bomb in 1949 and the nuclear race was underway. The problem wasn't keeping atomic knowledge from leaking out of the United States. That had already happened, largely through espionage. The challenge was how to channel a technology that was starting to move between powers, laboratories, alliances and national programs.

So control couldn't rest on secrecy alone. It had to lean on something more concrete: the critical material, the infrastructure able to produce it, and an institution that could manage, protect and inspect it. That logic was set in 1957 around a dual mission, promote the peaceful use of atomic energy and make sure that assistance wasn't turned to military ends.

When a technology becomes strategic, power starts deciding who gets access, under what conditions, within what limits, and on top of what infrastructure.

That's what's happening with frontier models, the most advanced of each generation. They aren't an incremental software upgrade. They're a new layer of general capability. Once they connect to repositories, data, credentials, APIs and infrastructure, they stop being just models and start acting as operational engines that amplify a person, a team, or a whole organization.

The critical point shows up in cybersecurity. These models speed up the search for vulnerabilities, automate analysis and lower the barrier to offensive use. That's where the debate changes scale. It's why governments are starting to treat frontier models as a national security matter.

The same pattern already runs with chips. The United States has spent years restricting access to the most advanced semiconductors. But a chip's capability no longer needs to cross a border. It can be rented in the cloud, used from another country, or diverted through intermediaries. In May 2025, US legislators introduced the Chip Security Act, a bill proposing location verification and diversion reporting for advanced AI chips.

If the critical input of the nuclear age was enriched uranium, today it's the chip. And whoever controls the chip controls the capability.

That same criterion started reaching civilian access. We got used to picking a provider, putting in a card, and using the best available model. That changed in mid-June 2026. Anthropic launched Fable 5 and three days later had to suspend it, along with Mythos 5, over a US government export-control directive that blocked access for anyone without US citizenship. To comply, it shut down both models for all of its customers.

On June 26 the government allowed a partial restoration of Mythos 5, Anthropic's cybersecurity model, only for previously vetted US organizations. The same day, OpenAI released its new GPT-5.6 model in preview to a small group of partners, also at Washington's request.

The underlying signal is clear. Access to frontier models is starting to depend on lists, jurisdictions, approved organizations and use cases. KYC stops being an identity check and starts working as an access layer: who can use what level of intelligence, from which country, and under whose responsibility.

You don't need to ban AI to restrict its use. Ordering access is enough.

Knowledge, on the other hand, circulates. Just like atomic technology once the monopoly stopped being a viable strategy.

China is already showing it. Z AI released GLM-5.2 as an open-weight model under an MIT license, with a one-million-token context window. It didn't generally beat the leading closed models, but it keeps closing the gap. And in cybersecurity, which is what worries governments most, a Semgrep evaluation placed it above Claude Opus 4.8 at vulnerability detection, at a fraction of the cost.

Consuming a model by API from a Chinese provider can cut costs. But it carries a less visible cost, handing operational context, sensitive data and usage patterns to a company that, under China's intelligence law, can be compelled to share them with its government.

The model can travel. The real capacity to run it at scale can't.

Model access isn't sovereignty. At scale, what matters is where it runs, on which chips, on which cloud, in which datacenters, with how much energy, and with what operating team. If that layer is in third-party hands, sovereignty stays borrowed.

For a company, this reading helps order roadmap, vendors, architecture and risk. For a country or an industry it marks something deeper. It defines who has real access to capability, who depends on others, and who can sustain it over time.

In the encyclical "Magnifica humanitas", from May 2026, Pope Leo XIV framed it from another angle. He warned that technological power is concentrating in a handful of private companies that already hold more power than many governments and sit outside their control. And that a more ethical AI isn't enough if a few decide that ethics.

The Pope arrives through ethics where security arrives through risk.

The problem isn't only the model. It's who concentrates the capability.

The visible debate is access to models. The real debate is infrastructure, capability and dependence.

This is where much of the economy, security and politics of the coming decades starts getting decided.