The sovereignty speech that arrived without a supply chain

A G7 head of state has named frontier AI access as a national sovereignty risk, on the record, forty-eight hours after the model in question was switched off for his country's researchers, and twenty-four hours before the lab's CEO walks into the room. That is the political event. The harder question is what Mark Carney intends to do about it, because the speech arrived without a supply chain attached.

The forcing function. On Friday 13 June, the United States issued what the Los Angeles Times described as a "sweeping" order, and Anthropic shut down non-US access to Mythos 5 and Fable 5, its flagship and near-flagship models. On Sunday, in Ireland, Carney told reporters that the restrictions "underscore risks of dependence on a limited number of American providers." On Monday, in Évian, Dario Amodei, Sam Altman and Demis Hassabis were scheduled to brief G7 leaders directly. The compressed sequence, shutdown, sovereignty speech, lab CEOs in the room, is the story. It is the first time a sitting G7 leader has reacted to a specific frontier-model shutdown by reaching for the language of national dependence.

The shift this represents is larger than the news cycle suggests. The G7's AI conversation has, since Hiroshima 2023, been organised around governance: how to align, how to red-team, how to write codes of conduct. That conversation assumed access. Carney's intervention reframes it. The question is no longer how do we govern this; it is how do we make sure we can still use it on Tuesday.

What the export order actually does

The architecture of the US order matters, because it is not the export control regime people are used to. Prior rounds — chip controls on advanced GPUs, the October 2022 and October 2023 BIS rules, the AI diffusion framework — restricted goods crossing borders. Hardware. Model weights, in some readings. Things you could put on a manifest.

The June order, as reported, extends the filter to non-US citizens employed inside US laboratories. That is a structural escalation. A Canadian researcher at Anthropic's San Francisco office, a British engineer at OpenAI, a French scientist at DeepMind — under the new framing, each of them is a potential vector of export risk by virtue of nationality. The control is no longer on the artefact. It is on the person.

This is worth dwelling on, because it changes the political economy of frontier research. The US lab ecosystem has, until now, been a magnet — drawing the strongest researchers from allied countries into a single concentrated cluster. The implicit deal was that talent flowed in, capability flowed out, and the allied countries got API access in return. The new order partially breaks both legs of that deal at once: it puts allied nationals under suspicion inside the labs, and it cuts the API on the way out.

The official rationale, preventing adversary access to frontier capability, is coherent. It is not nothing. There is a real, documented body of US national-security analysis behind it, and dismissing the controls as pure overreach misreads how seriously Washington takes the model-weights-in-Beijing scenario. But coherence at the policy level does not insulate the order from its second-order effects on allies. Those effects are what Carney is responding to.

Why the speech matters more than its specifics

Here is the uncomfortable part of Carney's intervention: it contains almost no policy. No funding commitment. No named domestic programme. No bilateral instrument. No carve-out he is demanding from Washington. Canada does not have a frontier AI laboratory. It has Cohere, which is not in the Mythos-class capability tier; it has world-class research talent at Vector and Mila, much of which has historically been recruited away by the US labs the speech is now criticising; it has compute capacity that is a rounding error on what Anthropic alone is provisioning this year.

The speech, in other words, is a demand signal without a supply response. Read uncharitably, it is pre-summit positioning — a G7 head of state turning a bad weekend for his researchers into leverage in Évian. Read more carefully, it is something subtler: a public acknowledgement that the implicit deal has changed, made early enough that other allies can repeat the line without inventing it. Carney does not need to have the answer. He needs the question to be in the room when Amodei sits down.

That is the diplomatic move. A speech of this kind is a coordination device. The PMs and presidents of Japan, the United Kingdom, Germany, France, Italy and the EU now all have a head-of-state quote they can point to before saying their own version of it. If you wanted to ensure the Évian communiqué contains language on "resilient and diverse AI supply", a Sunday speech in Ireland is exactly how you would do it.

The speech is a demand signal without a supply response. It is also, deliberately, a coordination device for everyone else's demand signals.

The supply side, honestly assessed

If Carney is serious, and the test will be the autumn budget cycle, not the Évian photo op, Canada and its peers have three options. None of them is clean.

Buy European. Mistral is the obvious recipient of Carney's quote. It has positioned itself, with some success, as the sovereign European alternative. Procurement offices in Ottawa, Canberra and London will read the Sunday remarks as cover to widen Mistral pilots and trim US-vendor concentration in next year's framework agreements. The problem is that swapping concentrated US dependence for concentrated French dependence is not sovereignty; it is a different counterparty. Mistral is a private company with its own investor base, its own regulatory exposure, and, at frontier capability, a meaningful gap behind Anthropic and OpenAI on the benchmarks that matter for the workloads governments actually run.

Build domestic. This is the honest sovereignty answer and the one nobody wants to cost. A serious domestic frontier programme, sovereign compute, sovereign training runs, sovereign post-training, is a multi-tens-of-billions commitment over a decade, against an incumbent set whose annual capex now runs in the hundreds of billions. Canada alone cannot do this. A coordinated EU-plus-Five-Eyes-minus-US programme could, in principle. There is no evidence such a programme exists beyond the speech-level.

Run open weights. This is the option the reporting on Carney's remarks almost entirely ignores, and it is the most technically immediate. The Llama lineage, DeepSeek, the Qwen family, open-weight models in the rough capability band of last year's frontier, run on domestically controlled compute, with no API dependency on any foreign lab. They are not Mythos 5. They are, in many enterprise and government workloads, close enough. The awkward part is that the strongest open-weight options now come out of Chinese labs, which means the sovereignty hedge against Washington routes, partially, through Hangzhou. That is the conversation Carney did not have on Sunday, and the one his procurement officials will be having on Tuesday.

The pattern across both poles

There is a second piece of context that should not be lost. In the same fortnight, Beijing forced the unwind of Manus's foreign-investor structure, demonstrating its own willingness to use AI infrastructure as a coercive instrument when geopolitics demanded it. Washington has now done the same in the other direction. The pattern is not a coincidence. Both great powers are treating frontier AI access as a lever. Both are pulling.

The middle, Canada, the EU, Japan, South Korea, Australia, sits between two state actors who have each, within weeks, used their domestic AI ecosystems as instruments of foreign policy. The Hiroshima Process was designed for a world in which AI was a technology to be governed in common. The June events describe a different world, in which AI is an instrument of state power to be hedged against. The governance problem and the alliance-management problem are not the same problem, and the second one is harder.

This is the framing Amodei will walk into on Monday. He will be asked, in a closed room, why his company switched off two models for the researchers of every country represented at the table. He will have a good answer, the order was sweeping, compliance was not optional, the alternative was worse, and the answer will not be the point. The point is that the G7 leaders now know the kill-switch is real, and they have watched it being used.

What Carney said in Ireland is that the next G7 will be organised around that fact. Whether Canada turns up to that meeting with a budget line, a domestic compute plan, or just the same speech in a different suit is the question the next twelve months will answer. I would not bet heavily on the budget line. I would bet, with some confidence, that the speech will be repeated.

Footnotes

¶