The Table of Nations

President Trump takes his seat at the working lunch and the two men flanking him are not heads of state. On his right sits Sam Altman of OpenAI. On his left sits Demis Hassabis of Google DeepMind, a Nobel laureate in chemistry. Across the table, the French host, Emmanuel Macron, is bracketed by Dario Amodei of Anthropic and Marc Benioff of Salesforce. The flags are real, the leaders are real, and so is the lunch. What has changed is who gets a place at the table.

This is Évian-les-Bains, on Lake Geneva, on June 17, 2026: the 52nd G7 summit, and the first time the chief executives of OpenAI, Google DeepMind, and Anthropic have all sat in the same room with the leaders of the world's richest democracies at once. Around a dozen technology executives joined the lunch, billed under the dry summit language of "ensuring a safe, rapid and effective deployment of artificial intelligence." The phrasing was bureaucratic. The image was not. For one afternoon, the people who build the most consequential technology of the century were seated as peers to the people elected to govern it.

You can read the whole shift in that seating chart. No elected parliament trained these models. No treaty governs them. Every government around the table already runs parts of its economy, its research, and its security apparatus on systems owned by three private American companies. The executives did not arrive as vendors hoping for a contract. They arrived as principals.

The Seats Are Assigned

The guest list read like a map of where frontier AI is actually being built. Beside the three American lab heads sat Arthur Mensch of France's Mistral, Aidan Gomez of Canada's Cohere, and a spread of national champions: Uljan Sharka of Italy's Domyn, Victor Riparbelli of the British video startup Synthesia, Robin Rombach of Germany's Black Forest Labs, and Alexandr Wang of Meta. Founders from India's Sarvam and Japan's Sakana rounded out a table that was, in effect, a working summit of the people who control the supply of machine intelligence on four continents.

When the executives spoke, they did not pitch products. They talked like statesmen worried about a commons.

Amodei told the room that the democracies should build shared rules before the technology splinters them apart. International cooperation, he argued, should cover structured access to frontier models, trade in the chips and critical components that make them run (a trade that pointedly excludes China), and joint defense against the misuse of AI in cyberattacks, bioterrorism, and intelligence operations. His warning to the heads of state was blunt: resist the temptation to splinter. Fragment your approaches to AI and you simply hand the advantage to whoever is willing to coordinate.

Altman went further, and turned the warning back on himself. He called 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." Then the line that hung in the air afterward. "Do not cede your responsibilities to AI labs like mine," he told the leaders. "No single lab should be making the decisions." It is a strange thing to hear from the man who runs the most famous lab of them all: a request to be regulated, delivered to a table that mostly does not know how.

Hassabis, characteristically, zoomed out to the horizon. "When we look back at this time in 10, 20 years' time," he said, "I think we'll see that we're standing in the foothills of the most significant moment in human history. It's immense potential." A Nobel laureate telling presidents they are living through the opening of a new era, and being received not as a provocateur but as an authority.

The reverence in the room had a reason. And the reason was sitting, very conspicuously, in one empty seat's worth of tension.

One Chair Sits Under a Cloud

Five days before the lunch, the United States government did something it had never done to a piece of consumer software. On June 12, the Commerce Department's Bureau of Industry and Security sent Anthropic a letter requiring an export license for any foreign national to use its two newest models, Fable 5 and Mythos 5. Inside the United States or outside it, including the company's own foreign employees, nobody without American citizenship could touch them without government permission. There was no realistic way to filter a global user base that finely in real time, so Anthropic did the only thing the order allowed: it switched both models off for everyone on the planet, the same week it was meant to be celebrating them.

Export controls are the machinery a government built to keep missiles, uranium, and stealth coatings out of the wrong hands. Pointing that machinery at a chatbot used by hundreds of millions of people was new. To understand why Washington reached for it, you have to understand what these models had started to be able to do.

Mythos, by independent assessment, is unusually good at finding software vulnerabilities. The United Kingdom's AI Safety Institute, which previewed the model, verified that it could identify complex, multi-stage cyberattacks. That is the dual-use problem in its sharpest form. A model skilled enough to find and patch a security hole for a defender is, by the very same skill, capable of finding and exploiting it for an attacker. The ability arrives without an intention attached. Anthropic had released Fable 5 as a guarded version of that power, with safeguards meant to stop it from answering questions that map out cyber vulnerabilities.

The trigger, by Anthropic's own account, was a report that someone had found a way around those safeguards. Researchers at Amazon reportedly identified a method to jailbreak Fable 5, slipping past the guardrails to reach Mythos's offensive capabilities underneath. A jailbreak, in plain terms, is a way of phrasing a request so a model does what it was specifically built to refuse. Officials feared the crack was wide enough to turn a public product into a cyberweapon. Anthropic pushed back, calling it likely a misunderstanding and arguing the exploit was narrow rather than a master key that broke every protection the model carried.

That uncertainty is why the episode rattled the whole industry, not just one company. If access to a hosted model can be switched off worldwide by a letter, on a legal basis no one had tested, then every lab now plans around a new kind of risk. The most powerful model in the world is only as available as the most nervous government willing to regulate it. That was the cloud Amodei carried into the room at Évian. He was not there as a supplicant whose product had been banned. He was there, days after his flagship went dark by government order, still seated beside the president of France as an equal.

Washington Floats a Security Clearance for AI

By the time the leaders sat down to eat, the summit had a fix to discuss. The G7 floated what officials called a "trusted partners" scheme: a framework under which allied countries or companies could be granted access to American frontier models that other nations cannot reach. Think of it as a diplomatic security clearance for artificial intelligence, with Washington holding the pen.

The push was concrete, not theoretical. The Financial Times reported that a US delegation led by Commerce Secretary Howard Lutnick worked the idea with European diplomats, pitching a system in which close, vetted G7 allies could regain access to the models the June 12 order had cut off. Canada's prime minister, Mark Carney, backed it at the table. Macron said he expected Washington to broaden access to Mythos within weeks, even while he called the export restrictions "strictly nationalist." Behind the scenes, Anthropic ran its own diplomacy: co-founder Tom Brown opened talks with US officials, including National Cyber Director Sean Cairncross, to find a path back online.

The negotiation worked, partly. Nine days after the summit, on June 26, the administration cleared Mythos 5 for use by a group of roughly a hundred vetted American companies and federal agencies. The frontier had been switched off for everyone, then switched back on for a chosen few. That is precisely the world the "trusted partners" idea describes: access to the most capable intelligence on Earth becoming a benefit that a single government grants and withdraws.

What it was not, anywhere, was a binding rule. The G7's joint statement said leaders would task finance officials, regulators, and cybersecurity experts with studying how frontier models might affect financial stability, productivity, and labor markets. Studying. The communiqué restated the existing OECD AI Principles in non-mandatory language, exactly consistent with an American position that favors voluntary commitments over hard law. Two weeks earlier, the same administration had signed an executive order, "Promoting Advanced Artificial Intelligence Innovation and Security," that set up a strictly voluntary framework for labs to consult the government before releasing new models. The order largely formalized a practice the labs had already agreed to. Light touch, by design, right up until the moment Commerce reached for the heaviest tool it had.

Brandt's line is the quiet thesis of the whole afternoon. A government can announce whatever AI policy it likes. It cannot deliver on that policy without the companies that own the models, the weights, and the data centers. The summit photograph did not create that dependency. It simply made it impossible to keep pretending otherwise.

The Table Will Convene Again

Step back from Évian and the larger picture is hard to unsee. Three private American companies, OpenAI, Anthropic, and Google, now hold the most advanced AI on Earth. Every government at that lunch depends on their systems. The capability was built with no public vote, governed by no treaty, and switched off and on, that very month, by the decision of a single executive branch. The CEOs were in the room for a practical reason that has nothing to do with vanity: being present lets them shape the language of any communiqué, rather than have governments write rules about their companies without them.

For the allies, the lesson of the dark fortnight was uncomfortable, and they said so. Carney put it plainly. "Nobody has done anything wrong in the situation," he said. "But we will have done something wrong if we just accept this, don't take the lesson, don't build out and diversify." The export ban did not just inconvenience Europe and Canada. It taught them that access to American intelligence is, in the end, a function of American politics. Analysts at the Center for Strategic and International Studies warned the move could push foreign customers toward options they consider more durable, including smaller open-weight models that run on hardware a government actually owns. Cut off from the best closed model, a country starts shopping for one it can never be cut off from.

In Europe the reaction curdled into something close to a sovereignty doctrine. One French presidential candidate reached for the most loaded analogy available, arguing that the continent "must treat AI the way we treated nuclear power: we must think of it as part of our sovereignty. Master it or suffer it: there is no other path." The same logic gives China an opening. By CSIS's reckoning, Chinese models trail the American frontier by roughly seven months on average. Every week the best US models are unavailable abroad is a week Beijing can use to win the international customers Washington just turned away.

That is the paradox the lunch leaves behind. The labs are asking, sincerely, to be governed. The governments want to govern. And yet the only thing the world's most powerful democracies could agree to in Évian was to study the problem and meet again. The mechanism for actually deciding who may use a frontier model still runs through a Commerce Department letter and a list of trusted partners, not through anything a voter chose.

Évian was less a policy summit than a power photograph, one that made visible what had only been implied. The people who build the most consequential technology in human history now sit, literally, beside the people who govern nations. Whether that arrangement reassures you or alarms you depends mostly on which side of the table you imagine yourself on. The chairs are already set. The next time they are filled, the only open question is whether anyone at the table other than the lab chiefs will know what they are deciding.