On Wednesday June 17, 2026, twelve AI executives walked into a closed-door working lunch with the G7 heads of state in Évian-les-Bains, France. Dario Amodei was at the table. Demis Hassabis was at the table. Sam Altman was at the table. Arthur Mensch was at the table. They had been invited as guests. They left as the room's centre of gravity. Amodei and Hassabis used the slot to ask the allies to sign up for what CNBC and TheNextWeb both framed in the same words: a US-led AI coalition — structured access to frontier models, a chip-trade ring that excludes China, a joint posture on AI risk in cyber and biosecurity. Canadian Prime Minister Mark Carney said yes. The French and German chairs reserved the right to bring a list of conditions, and the list is the story.

Five days

The lunch was on Wednesday. The previous Friday, June 12, the US Commerce Department had ordered Anthropic to block foreign nationals from Fable 5 and Mythos 5; within hours Anthropic had pulled both models offline globally. The Loop covered the export directive itself and the Andy Jassy phone call that triggered it. Read those two pieces and the Évian lunch becomes a different event. Amodei is not at the G7 because he is a tech personality. He is at the G7 because the US government has just demonstrated that it can switch off his company's flagship product overnight, and a large share of his addressable revenue is European, and the European response will determine whether Anthropic gets to be a global firm or a domestic one.

European leaders read the same five days the same way. French President Emmanuel Macron, in floor remarks that Bloomberg, Euronews, and The National each carried in different versions, said the Anthropic episode had "clarified the stakes" and warned that if Washington "from one day to the next can turn off the switch," the multi-trillion-dollar US AI franchise would damage itself. The phrase the European press has been quoting since Wednesday afternoon: "we will fracture the world, and we will have no effective solution."

The Hassabis dovetail

Hassabis pitched the safety frame: an international standard-setting body that would write the rules for safe frontier development without strangling the deployment cycle. The press summary is bland; the strategic move is sharp. DeepMind ships out of Mountain View and London, holds dual exposure to US and UK jurisdiction, and has the most to lose from a fragmented compliance map. He is asking for one centre of gravity for the rule-writing — implicitly, on his side of the Atlantic.

Altman ran a different line. His proposal, in the only direct quote any of the three CEOs gave the press, was for "an international forum for discussion that establishes globally accepted standards for testing, provides expert and impartial analysis of capabilities and risks, and serves as a venue for cooperation among nations." OpenAI's IPO roadshow needs a regulatory narrative that does not start with the New York Attorney General's office, and a G7 testing forum is exactly the venue that lets it.

Three labs, three different framings, one position: please write the global rules with us at the table, before the next national export-control letter does it again.

What Europe brought

The European pushback was not a refusal. It was a price.

European Commission President Ursula von der Leyen described, per The National, a "mutual interest" between US and European firms in sharing technology. German Chancellor Friedrich Merz said, in the same writeup, that "the potential of these new technologies should be available for all countries." The EU Commission's tech-sovereignty spokesperson Thomas Regnier delivered the Brussels line on the record to Euronews: "We are a trusted partner. I would challenge you to find a more trusted partner than Europe." That last sentence is the corporate-affairs-team way of saying we know you have other options and we'd prefer you didn't price them. The Italian MEP Brando Benifei, to the same outlet, was less polished: "the Anthropic kill switch shows that tech sovereignty was never abstract."

Translated into procurement language: Europe will sign the trusted-partners scheme, but the access guarantees have to be written down. The version of the coalition where Washington keeps the kill switch as a sovereign prerogative is the version Europe will fund a parallel stack against. France has already lined that parallel stack up — Mistral, civil-service deployments, the €655 million national AI commitment Arthur Mensch keeps citing. Berlin has the same option through the Aleph Alpha programme. Brussels has the AI Act. None of these are bluffs.

Carney said yes

The cleanest moment of the lunch belonged to Canadian Prime Minister Mark Carney, who agreed that the US could lead the coalition. CNBC and TheNextWeb both have it. The same Carney also told the room, per Yahoo's wire writeup, to "build out and diversify" rather than commit to a single model. Both statements are consistent. He is endorsing the coalition because Canada is too small to opt out, and he is hedging by telling the other allies to keep their second-source vendors warm. It is the most honest line of the day.

The trusted-partners scheme

The actual G7 deliverable, per The National, is a trusted-partnership platform to be stood up within one month, with the leaders set to reconvene in September. A diplomatic source paraphrased the working rationale: "if you don't want China to catch up on US and European tech, then let's set up trusted partners to exchange innovations." That is the formal name for what the AI bloc is becoming. The platform is the wiring; the September meeting is the test.

Two things to watch in the next four weeks. First, the membership list. The G7 plus India was the room at Évian. If Singapore, South Korea, Australia, Japan-with-deeper-access, and one or two named European technical bodies appear on the founding roster, the coalition is being designed wide. If it stays at seven, it is being designed as a US-and-friends fence. Second, the access mechanism. Either the scheme is a real exemption framework where Anthropic's foreign Brussels-based engineers can use Mythos 5 next month, or it is a press release. There is no middle layer.

The room

The lab side of the table is worth listing once, because it is the new shape of the AI industry. Per Fortune's and FourWeekMBA's attendee lists: Dario Amodei (Anthropic), Demis Hassabis (DeepMind), Sam Altman (OpenAI), Arthur Mensch (Mistral), Aidan Gomez (Cohere), Robin Rombach (Black Forest Labs), Victor Riparbelli (Synthesia), Marc Benioff (Salesforce), Alex Wang, Pratyush Kumar (Sarvam AI), Ren Ito (Sakana AI), and a representative from Domyn. Twelve seats, four continents. The US officials present were Treasury Secretary Scott Bessent, Commerce Secretary Howard Lutnick, and Secretary of State Marco Rubio. The same Bessent who, per the Wall Street Journal reporting the Loop covered last week, took the Andy Jassy phone call that started the Friday export directive. The chain has now closed: investor calls Treasury, Treasury moves Commerce, Commerce switches off Anthropic, Commerce-Treasury-State sit across from Anthropic at the G7 five days later. Six weeks ago none of this was infrastructure. It is now.

What this is

This is the AI bloc formalising. Not as a metaphor, as a procurement framework with named members, a one-month deadline, and a September check-in. The labs are not lobbying the bloc from outside. They are sitting inside the room writing the rules with the heads of state, and the heads of state are taking the meeting because the labs now have the kind of capability that state-level actors used to monopolise. The clean diplomatic line of the day, from a senior G7 official quoted by The AI Chronicle: "we are no longer discussing whether AI will change the world, but who will be holding the steering wheel." The lab CEOs were sitting at the steering wheel; the heads of state were sitting in the passenger seats. That is the picture from Wednesday.

What to watch

1. The 30-day platform. If the trusted-partnership scheme has a published membership list and at least a draft exemption mechanism by mid-July, it is real. If the next press release is in September with no platform, the Évian commitments were theatre.

2. Anthropic's foreign-engineer carve-out. The cleanest single test of whether the coalition has teeth is whether non-US Anthropic employees can use Mythos 5 again before Q3 ends. The case is small enough to fix quickly and high-signal enough to set the precedent.

3. The Macron list. Paris will not sign a trusted-partners platform that does not include a written limit on the unilateral US kill switch. If Washington refuses, France will hard-fork the coalition through Mistral and Brussels, and the next G7 hosts a competing scheme.

4. Carney's diversification line. Ottawa endorsed the US lead in public and told the room to keep their second-source vendors warm in the same lunch. The next Canadian federal procurement that names a non-US model on its shortlist will tell you which sentence Carney meant.

The model wars are not the only wars that matter, and the diplomatic ones are not new. What is new is which side of the table the lab CEOs sit on. On Wednesday it was the same side as the heads of state.