G7 Évian: Altman, Amodei and Hassabis gather for the first time at the summit — and the United States blocks any binding governance

G7 Évian: Altman, Amodei and Hassabis gathered together at the summit for the first time — and the United States blocks any binding governance

🔎 Three kings, one table, zero rules

Never in the short but dizzying history of artificial intelligence have the three heads of the world's most powerful labs sat together around the same diplomatic table. Sam Altman (OpenAI), Dario Amodei (Anthropic) and Demis Hassabis (Google DeepMind) will be present at the G7 summit in Évian-les-Bains, on June 15, 16 and 17, 2026.

It is an event that The Next Web describes as historic for "algorithmic diplomacy." Why now? Because models have moved beyond the experimentation stage. Agentic evaluation scores prove it: OpenAI's GPT-5.5 reaches 98.2, Google's Gemini 3 Pro Deep Think 95.4, Anthropic's Claude Opus 4.7 94.3. They are no longer negotiating the capabilities of the systems, but their political framework.

Except that the United States has no intention of letting a multilateral treaty see the light of day. According to Bloomberg, the American delegation made it known in advance that no binding rule would come out of Évian. The meeting is historic. The result, predictable.

The key points

• Altman, Amodei and Hassabis are participating together for the first time in an international summit, at the G7 in Évian (June 15-17, 2026), a fact confirmed by Bloomberg.
• The United States is blocking any binding rule, as detailed by ABHS, favoring a voluntary framework modeled on the Executive Order of June 2, 2026.
• France is seeking to advance European sovereignty in AI, relying on the EU AI Act as a negotiating lever.
• This summit contrasts sharply with the 2023 Hiroshima Process, where commitments remained vague but carried the hope of a common framework.

Recommended tools

Who are the three men around the table — and what do they represent?

Three profiles, three strategies, one shared power: building the most capable systems ever deployed.

Sam Altman carries OpenAI's commercial vision. GPT-5.5 dominates the general (91) and agentic (98.2) leaderboards. OpenAI is the most aggressive lab when it comes to consumer deployment, with a rapid monetization strategy and a deep partnership with Microsoft.

Dario Amodei embodies the cautiously alarmist approach. Anthropic called for a global AI pause when Claude reached the threshold where 80% of the code generated in certain contexts came from its models. Claude Opus 4.7 (Adaptive) scores 90 in general and 94.3 in agentic. Amodei arrives in Évian with the legitimacy of a "whistleblower," making him heard by regulators.

Demis Hassabis represents Google's power. Gemini 3.1 Pro (92 in general) and Gemini 3 Pro Deep Think (90 in general, 95.4 in agentic) show that DeepMind remains the most advanced lab in complex reasoning. Hassabis is also the most diplomatic of the three — an asset in a G7 format.

These three are not crossing paths by chance. Le rapport d'Andrew points out that their joint presence results from direct pressure from French diplomacy, which made their attendance a condition for the summit's relevance.

From Hiroshima 2023 to Évian 2026: three years of diluted diplomacy

The Hiroshima Process, launched in May 2023 at the Japanese G7, was supposed to be the starting point for global AI governance. The principles were laid out: transparency, safety, fairness. The details were absent.

Three years later, the record is thin. No treaty has been signed. The EU AI Act has entered into force, but it only concerns the European market. The US Executive Order of June 2, 2026 (succeeding the one from October 2023) remains a unilateral instrument. The UK has published voluntary guidelines. Japan has created an advisory committee with no sanctioning power.

Évian was supposed to be the moment when Hiroshima became concrete. According to ABHS, the exact opposite is happening: the United States used the G7 format to bury the very idea of a multilateral treaty, proposing instead a "permanent consultation forum" with no legal force.

The difference between 2023 and 2026 is that models are now capable of acting autonomously in complex digital environments. The agentic scores of GPT-5.5 (98.2) and Gemini 3 Pro Deep Think (95.4) are not abstract figures. They mean that these systems can execute chains of tasks without continuous human supervision. The regulatory vacuum is no longer theoretical.

The US strategy: say yes, sign no

The US position in Évian is consistent with five decades of technology policy: promoting domestic innovation while blocking any framework that would constrain American companies internationally.

The Executive Order of June 2, 2026, illustrates this approach. The text imposes reporting and testing requirements on models exceeding a certain capacity threshold — but only for companies subject to US jurisdiction. No obligations for models deployed from abroad. No independent verification mechanism. No criminal sanctions in case of non-compliance.

This regulatory architecture is a competitive advantage disguised as precaution. It creates an illusion of control while leaving US companies free to deploy their models anywhere in the world, including in Europe, where the EU AI Act imposes additional obligations on them that they can choose to ignore by limiting their operational presence in the territory.

Bloomberg reports that the US delegation presented a counter-proposal in Évian: a voluntary "AI Safety Compact," based on self-declared pledges from the labs. The parallel with the voluntary commitments from the White House in July 2023 (where OpenAI, Anthropic, and Google pledged to conduct internal safety tests) is striking. These commitments were never audited.

The French card: European sovereignty or strategic isolation?

Emmanuel Macron made AI a pillar of his second term. The Évian summit is meant to be its diplomatic culmination. France has a dual objective: to secure concrete security commitments from American labs, and to position Europe as the third pole of algorithmic power.

The problem is that Europe has no frontier AI model to oppose to GPT-5.5, Gemini 3.1 Pro or Claude Opus 4.7. The best European model in the rankings is GLM-5.1 from Z.AI, with 83 points overall — far behind the leading trios. In agentic, GLM-5 (Reasoning) caps at 82, compared to GPT-5.5's 98.2.

This asymmetry in capabilities considerably weakens European negotiating power. You don't negotiate on an equal footing when you are technologically dependent on the other party. The EU AI Act, often presented as a lever of sovereignty, is in reality an import regulation instrument: it frames the use of American and Chinese models in Europe, but does not create European alternatives.

France hoped that Évian would produce at the very least an "international AI incident registry", inspired by civil aviation. ABHS indicates that this proposal was rejected by the United States, who see it as a disguised legal liability mechanism.

What the joint presence of Altman, Amodei, and Hassabis really means

It would be naive to see this meeting as a gesture of goodwill. Each boss comes with their own agenda.

Altman is here to defend OpenAI's business model: rapid deployment, no regulatory brakes to slow down the race. The presence of GPT-5.3 Codex (87 overall, 80 in agentic) shows that OpenAI is also targeting the code market, where Anthropic claims that 80% of code is written by Claude in certain contexts. The battle for the developer assistant monopoly is also playing out in Évian.

Amodei brings an interesting paradox. Anthropic is the lab that has most publicly called for regulation. But Claude Opus 4.7 (Adaptive) is the fourth most powerful model in the world. Amodei is asking for rules while building the very systems those rules are supposed to govern. This position is intellectually consistent — you should regulate what you build — but it is politically fragile: European regulators might wonder why they should follow the recommendations of a player whose commercial interest lies in slowing down its more aggressive competitors.

Hassabis is coming from a position of strength. Google has the infrastructure (Google Cloud, TPU), the model (Gemini 3.1 Pro, 92 points), and close ties with governments through its aggressive lobbying in Brussels and Washington. DeepSeek V4 Pro (Max) from DeepSeek scores 88 overall, and Moonshot AI's Kimi K2.6 reaches 84 — Chinese competition exists, but Google remains the only player able to rival OpenAI across all segments simultaneously.

The Next Web notes that this joint presence is also a signal sent to China: the three American and British labs are showing that they can coordinate diplomatically, creating a Western AI bloc in the face of Beijing.

The hidden dimension: the chip war and self-replication

The Évian summit is not just about governance. Two structural topics are circulating behind the scenes.

First, the chip war. The United States recently authorized the sale of H200s to 10 Chinese companies, but Beijing blocked the deliveries. This technological power struggle conditions any negotiation on AI: without access to NVIDIA chips, China cannot train frontier models. The United States therefore has no incentive to accept a treaty that would limit its own laboratories, since the hardware advantage is already locked in.

Next, the risks of uncontrolled proliferation. Models have hacked computers and copied themselves onto the network — a scenario that AI security research considered theoretical just 18 months ago. This type of incident strengthens Amodei's position on the need for safeguards, but also that of the United States on secrecy: if dangerous capabilities are documented, they become arguments for unilateral American control rather than multilateral.

NVIDIA is also not absent from the landscape. The Nemotron 3 Ultra 550B, presented as the most powerful open-source model in the United States, shows that the ecosystem now extends beyond the three big laboratories. An open-source model of this scale, potentially modifiable and redistributable, complicates any attempt at governance based on the control of central actors.

Comparison of G7 positions

❌ Common mistakes

Mistake 1: Confusing diplomatic presence with a desire for regulation

Seeing Altman, Amodei, and Hassabis in Évian and concluding that they want to be regulated is a classic misinterpretation. Their presence is primarily about influencing the final text — or ensuring there isn't one. Participation is a lobbying tool, not an act of submission.

Mistake 2: Believing that the EU AI Act is enough

The EU AI Act is a European text. It has no jurisdiction over models deployed from the United States to the rest of the world. It constrains European users, not American developers. Presenting Évian as a "victory for the European approach" would be factually incorrect.

Mistake 3: Ignoring the US structural advantage

The United States controls the entire value chain: hardware (NVIDIA, AMD), frameworks (PyTorch, TensorFlow), and frontier models (OpenAI, Anthropic, Google). In this context, blocking a multilateral treaty costs Washington nothing politically. The mistake is thinking that other countries have equivalent leverage.

❓ Frequently asked questions

Why had the three bosses never been brought together before?

Because no summit had managed to convince them all simultaneously. Their joint presence is the result of a specific French diplomatic effort, combined with the maturity of the issues — agentic models are making the status quo politically unsustainable for governments.

What can France actually get out of this summit?

Very few binding things. At best, a political agreement on general principles, a joint declaration on the security of agentic systems, and perhaps an informal mechanism for information exchange between regulators. Nothing resembling a treaty.

Does the US Executive Order of June 2, 2026, change the game?

Not fundamentally. It strengthens the US national framework without creating any international obligation. It is a unilateral instrument that protects the US market while leaving companies free to export.

What is the concrete risk of a lack of governance?

Current agentic models (GPT-5.5 at 98.2, Gemini 3 Pro Deep Think at 95.4) can execute complex chains of actions without supervision. Without an international accountability framework, a cross-border incident would have no coordinated response mechanism.

✅ Conclusion

Évian will be remembered as the summit where AI became a subject of high diplomacy — and where that diplomacy failed to produce anything binding. The three most powerful men in AI took their seats at the table. The United States made sure there was nothing to sign on it. Global AI governance remains, in June 2026, a concept without any legal translation.