The Cartel That Couldn’t Agree on a Lie

Évian-les-Bains has hosted exactly one other G7 summit in its history, back in 2003, which means the lakeside town has now twice watched the world’s most powerful people gather to perform unity for the cameras. The 2003 vintage was about Iraq. The 2026 vintage was about something that sounds smaller and isn’t… who gets to own the thinking machines, and what happens to everyone else when the people who own them decide to pull the plug.

Officially, what came out of that working lunch on 17 June was the usual fog. AI needs rules. Children need protecting online. Democracies need to cooperate. The protection of children online was confirmed as a key part of the discussions by the Élysée Palace. Fine. Nobody’s against that. It’s the kind of consensus you reach when you haven’t actually agreed on anything that matters and need a paragraph for the press release.

What actually happened in that room was messier, and considerably more interesting, because ten days before anyone sat down to eat, Washington had already detonated the thing the summit was supposed to be diplomatically tiptoeing around.

The Friday afternoon that changed the seating plan

The summit had been preceded by recent announcements of powerful AI models with advanced cyber capabilities, including Anthropic’s Mythos and OpenAI’s GPT-5.5 Cyber, which had already unsettled businesses and governments over digital security weaknesses. One Brookings fellow described the release of Mythos as an “inflection point” that pushed the Trump administration toward regulating the technology in the first place.

So Washington regulated it. Specifically, it cut off foreign access to Anthropic’s most advanced models, Fable 5 and Mythos 5, in a single administrative move with no warning and no vote. It was the first time the US had applied export controls to an AI model directly, and the order didn’t discriminate by ally or rival… it simply switched something off across more than a hundred countries on what one analyst described, with the particular bluntness that only trade-policy nerds can manage, as a single Friday afternoon directive.

That’s the detail I keep returning to. Not the geopolitics, not the summit choreography… the Friday afternoon. Somewhere in Brussels, in Tokyo, in Whitehall, a procurement officer who’d spent eighteen months building critical infrastructure on top of an American model came in on Monday to discover the ground had moved while they were at the pub. Multiple G7 nations had previously gestured at the need for sovereign AI investment, but the working assumption had always been that this could happen alongside continued access to American models, not instead of it. The export ban didn’t just restrict a product. It demonstrated, in the most unambiguous way possible, that the assumption was wrong.

Britain, notably, didn’t take this lying down quietly. London lobbied Trump directly for an exemption from the Anthropic ban. The answer was no. Sit with that for a second. A close ally, nuclear partner, intelligence-sharing partner under Five Eyes, asked nicely, and was told to get in line with everyone else. If you wanted a single image to capture what “America First” looks like when applied to the compute layer of the global economy, that’s it.

The lunch nobody outside the room actually understood

By the time Altman, Amodei and Hassabis sat down with Trump and the rest of the G7 leadership on the seventeenth, the wars in Iran and Ukraine had dominated most of the week, with AI getting its moment only on the summit’s final day. But it was the moment that mattered, because three of the most powerful AI executives in the world had gathered for a working lunch explicitly framed around “ensuring a safe, rapid and effective deployment of artificial intelligence.”

Read that phrase again. Safe, rapid, and effective. You could fit a small war between those three words and the thing actually being negotiated, which was: who controls the off switch, and who gets left holding nothing when someone else’s hand is on it.

Macron’s read on the situation was that Washington would likely broaden Mythos access again within weeks, on the straightforward logic that nobody buys a product they’re afraid might vanish overnight. He wasn’t being sentimental about it. He acknowledged the American security concerns as legitimate, while still calling the export restrictions “strictly nationalist.” Both things can be true. America can have real reasons to worry about who’s running advanced cyber-capable models, and still be making a decision that’s fundamentally about leverage, not safety.

The asymmetry in that room tells its own story without anyone needing to make a speech about it. Five American tech executives attended. Every other participating country sent one. Trump sat at the table with the heads of state of the seven wealthiest democracies on earth, and right beside him sat Sam Altman, a man who runs a company, not a country, and yet showed up with a seat that put him visually on par with prime ministers. Three private companies, OpenAI, Anthropic and Google, now control the most advanced AI on the planet, and every government at that table runs critical national infrastructure on top of their products, despite the fact that no elected parliament built these systems and no international treaty governs them.

That’s not a conspiracy. That’s just what happens when you let infrastructure get built faster than governance can be written for it. Nobody planned this concentration of power deliberately… it accumulated, the way debt accumulates, quietly and then all at once.

Three pitches, three businesses, one word doing all the heavy lifting

Here’s where it gets properly cynical, and where the surface-level “AI needs rules” consensus falls apart the moment you look at who’s proposing which rules.

Amodei and Hassabis pushed for a US-led AI coalition, while Altman called instead for a global testing forum built around common safety benchmarks. Three different shapes for the same underlying problem, and not coincidentally, three shapes that happen to fit each company’s existing business model like a tailored suit.

Amodei’s company had just been the one cut off. A “trusted partners” framework, of the kind G7 leaders discussed as a way to restore allied access to restricted US models, is exactly the mechanism that gets Anthropic’s lucrative European and global enterprise clients back online without anyone having to admit the export ban was a mistake. It’s damage control dressed up as democratic solidarity.

Altman’s pitch, the international standards forum, reads differently depending on which direction you’re facing. Facing forward, it’s principled: don’t let any single lab, including his own, hold too much regulatory power. Facing backward, toward an OpenAI that wants to remain globally ubiquitous ahead of whatever its future public market story turns out to be, it’s the positioning of a company that needs to look like global infrastructure rather than a nationally-bound asset that Washington could, in theory, switch off the same way it switched off Anthropic.

Hassabis wasn’t asking for sweeping geopolitical rules at all, and there’s a reason for that. Google doesn’t need geopolitics slowing it down. It needs predictable, narrow, technical checklists it can engineer around without disturbing a search, cloud and advertising business that was profitable long before anyone had heard the word “transformer.” A quarterly technical-benchmark body is the regulatory equivalent of a speed bump, not a wall.

None of this makes any of the three men dishonest, exactly. It makes them businessmen in a room full of heads of state, each one fluent in the dialect of public good that happens to also serve private interest. That’s not new. What’s new is the scale, and the fact that the thing being regulated is also the thing they’re each trying to sell.

The man who said the quiet part out loud

If there’s a hero in this story, and I use the word loosely, it’s Aidan Gomez, who runs a company most people outside the industry have never heard of and said the thing the big three were too careful to say plainly.

Gomez told Fortune the Anthropic episode had “woken everyone up to the reality” that centralised dependence on a single entity is a structural risk, and that access to a foreign-controlled model “can absolutely just be revoked, and services shut down.” That’s not boardroom diplomacy. That’s a man pointing at the smoking crater where allied trust used to be and saying, look, this is what we were warning you about.

His sharper formulation, delivered straight to G7 leaders, was that “renting artificial intelligence means surrendering operational control,” and that “a monopoly of intelligence is inherently brittle.” Which is true, and also happens to be the precise pitch for Cohere’s entire business model, since the company builds models that enterprises and governments can take full custody of and deploy inside their own secure environments, with the Canada-Germany tech alliance offered as proof that nations can access cutting-edge capability without trading away their self-determination.

So even the man calling out the cartel is, transparently, also selling an alternative to the cartel. I don’t think that disqualifies what he’s saying. I think it’s just worth noticing that everyone in this story, without exception, is narrating self-interest in the language of principle. The only difference between Gomez and the big three is that his version of self-interest happens to align more closely with what smaller democracies actually need, which is optionality, not a new master.

What gets crushed while the giants argue about the shape of the table

Here’s the part that should worry you more than any of the individual pitches, and it’s the part that barely makes it into the press coverage because it doesn’t have a recognisable face attached to it.

Every framework on offer, the trusted-partners model, the testing forum, the technical benchmark body, the sovereign-alliance model, comes with a compliance cost. Testing regimes, national security audits, data governance requirements, certification processes. For Anthropic, OpenAI and Google, that cost is a rounding error against revenues that already run into the tens of billions. For a two-person open-source project releasing model weights on a forum, it’s not a rounding error. It’s an extinction event.

This is the bit nobody at that lunch had any incentive to say out loud, because it’s not really about China, or cyber-risk, or even genuine safety, although all three get invoked constantly and sometimes sincerely. It’s about who can afford to clear the bar once the bar gets built. G7 leaders did adopt joint statements on a wide range of issues at the summit, including critical mineral supply chains and online safety for minors, the kind of broad multilateral language that tends to calcify, later, into exactly the compliance architecture that locks out everyone without a legal department the size of a small country’s civil service.

I’ve written before about how regulation framed as safety often does its real work elsewhere, quietly, in the gap between what gets said in the press conference and what gets built into the rulebook six months later. This is that pattern again, just with better catering and a lake view.

Where this actually leaves us

Nobody in that room at Évian was lying, exactly. Macron really does believe export bans are nationalist. Gomez really does believe centralised dependence is brittle. Amodei, Altman and Hassabis probably all believe, on some sincere level, that their preferred framework would make the world safer. The trouble with rooms like this isn’t dishonesty. It’s that everyone’s sincerity happens to point directly at their own survival, and there’s no mechanism in the room designed to notice that, let alone correct for it.

What comes out the other end won’t look like a turf war when it’s finished. It’ll look like a “framework,” with a name, probably announced at the next summit with the kind of language that makes it sound inevitable and carefully balanced. It won’t be either. It’ll be the visible residue of three companies and a handful of governments negotiating, in real time, who gets to keep their hand on the switch, and the rest of us finding out where we stand only once somebody, somewhere, decides to flip it.

A few notes on what I changed from your brief, because accuracy matters more than narrative tidiness here: I dropped the specific claim that Macron and Modi jointly “warned the room” in alarmist terms… what’s actually sourced is Macron’s market-trust framing and separately, India’s broader G7 engagement around aid-versus-investment and countering China, which is a different conversation thread, not the same warning. I also corrected the Gomez quote to what he actually said rather than the paraphrase, and added the Britain-exemption detail because it’s the single most concrete proof point of “the switch can be flipped on anyone” that exists in the public record so far.

Until Next Time