Dario Amodei published an 8,000-word policy essay on June 10 calling for the U.S. government to have legal authority to block frontier AI models that fail safety testing. Two days later, Bloomberg confirmed that Amodei, Sam Altman, and Demis Hassabis will all attend the G7 Leaders’ Summit in Évian-les-Bains, France, from June 15 to 17. For the first time, the three CEOs building the most powerful AI systems on Earth will sit across the table from the seven heads of state who regulate them.
The timing is not coincidental. Each company arrived at Évian with a distinct policy framework, and the distance between those frameworks reveals how fractured the AI industry’s relationship with government has become.
What’s Actually on the Agenda at Évian
French President Emmanuel Macron released the guest list through the Élysée Palace. Alongside the G7 leaders (Macron, Trump, Starmer, Merz, Meloni, Takaichi, and Carney), the summit includes invited heads of state from Brazil, India, Kenya, South Korea, and Syria. The AI executives join a working lunch on Wednesday, June 17, covering AI infrastructure, network security, and regulatory coordination.
Two specific agenda items stand out. First, youth safety: G7 digital ministers agreed on a joint framework for protecting children online in late May, and the Évian session is expected to formalize voluntary commitments from the labs. Second, frontier AI risks, with particular focus on cybersecurity and biological domains. Both topics gained urgency after Anthropic’s Mythos Preview identified thousands of zero-day vulnerabilities in internal testing through Project Glasswing, and after GPT-5.5 shipped with enhanced autonomous capabilities that raised new questions about agent safety.
This marks a structural shift. The Hiroshima AI Process from the 2023 G7 produced voluntary guiding principles and a code of conduct. The 2024 Apulia summit extended those with a reporting framework. Both were negotiated among diplomats and policy staff. Évian puts the CEOs themselves in the room with heads of state, turning what was a bureaucratic exercise into a direct negotiation.
Amodei’s Blueprint: FAA-Style Testing and $350 Million
Anthropic’s CEO did not arrive empty-handed. His “Policy on the AI Exponential” essay, published June 10, lays out the most detailed regulatory framework any frontier lab CEO has proposed. The core argument: voluntary governance is no longer sufficient.
The specific mechanism targets models trained above 10²⁵ floating-point operations, built by companies earning over $500 million in AI revenue or spending more than $1 billion on research. Those models would require mandatory third-party testing across four risk categories: cybersecurity, biological weapons, loss of control, and automated AI research that could accelerate the other three. The government would hold authority to block or reverse deployment if a model fails.
Amodei modeled the approach on the Federal Aviation Administration. “Frontier AI models, like airplanes, should be required to go through technical testing and auditing,” he wrote, “and their release should be blocked or reversed as a threat to public safety if they do not meet high standards.”
Alongside the essay, Anthropic pledged $350 million in two initiatives: a $200 million Economic Futures Research Fund financing research into wage insurance, retraining grants, and capital accounts for AI-displaced workers, plus a $150 million fellowship program for early-career Americans. The essay also outlines three escalating tiers of economic intervention tied to unemployment levels: pro-employment incentives at 5%, expanded insurance at 10%, and universal basic income at unprecedented displacement levels.
On geopolitics, Amodei called AI “the dominant future source of military and economic power” and advocated tightening U.S. export controls on frontier chips to China while building a democratic coalition that makes membership attractive through shared access to compute infrastructure.
Altman’s Counter: Voluntary Commitments and Youth Safety
OpenAI’s approach at Évian is structurally different. Rather than proposing binding regulation, the company is pushing for a package of voluntary commitments that AI labs would agree to at the summit, with youth safety as the centerpiece.
Chris Lehane, OpenAI’s chief global affairs officer, told CNBC that Altman expects tech firms to leave the summit having agreed to voluntary commitments, with youth safety “sitting at the top of Altman’s personal agenda.” The specific proposal calls for an international youth AI safety institute, modeled on Common Sense Media’s existing framework, that would coordinate governments, researchers, civil society, and industry on age-appropriate protections.
The proposed standards include privacy-preserving age estimation, stronger default safeguards when a user’s age is uncertain, annual youth safety risk assessments, parental controls over memory and data usage, clear protocols for self-harm and exploitation scenarios, and a ban on privacy-invasive advertising targeting minors.
This is the voluntary approach that Amodei’s essay explicitly rejected as insufficient. The gap between the two frameworks is the real story at Évian: one CEO wants the government to have the power to block model releases, while the other wants the industry to police itself with periodic commitments.
The Regulatory Capture Question Nobody at Évian Will Raise
Critics have already identified the structural problem with Amodei’s proposal. Mandatory third-party testing at the 10²⁵ FLOP threshold, combined with government-accredited evaluators and compliance processes, creates a regime that only well-funded labs can navigate. As one analysis put it, the framework risks building “a moat for incumbents” where the companies writing the rules are the same ones with the policy teams, legal budgets, and government relationships to comply with them.
Open-source researchers and smaller labs face the sharpest disadvantage. If frontier development requires government-approved evaluators and filing processes, independent AI research becomes dependent on incumbent approval. The irony is hard to miss: Anthropic is filing for a $965 billion IPO while simultaneously proposing regulations that would make it harder for competitors to reach frontier capability.
This is worth watching from inside an enterprise that buys AI services. Binding regulation at the frontier level does not just shape who can build models; it shapes the competitive dynamics that determine pricing, availability, and vendor lock-in for every company buying inference. When the U.S. government already runs 40 secret evaluations on frontier models through CAISI, adding a formal blocking authority changes the procurement calculus for any organization planning a multi-year AI deployment.
France’s Play: $87 Billion and a Personal Phone Call
Macron’s decision to put AI CEOs at the G7 table is inseparable from France’s broader strategy to become Europe’s AI infrastructure hub. SoftBank committed €45 billion (roughly $49 billion) over five years for French AI data centers, powered by the country’s nuclear fleet. MGX and Bpifrance are building a €7.5 billion AI campus. Salesforce committed €2 billion.
Macron personally called SoftBank’s Masayoshi Son and OpenAI’s Altman to pitch France as a destination. The G7 invitation is the latest move in that courtship: by putting these executives next to Trump, Starmer, and Merz, Macron positions France as the convening power for AI governance, not just a market.
This matters because the regulatory landscape is splintering. The EU AI Act was designed before agentic AI existed; its risk categories assume systems that assist human decisions, not systems that make and execute decisions independently. Trump’s June 2 executive order established a voluntary framework for frontier model engagement. The gap between Europe’s binding rules and America’s voluntary approach is widening, and Évian is where Macron hopes to bridge it.
What Évian Signals for Anyone Building on These Models
Three competing governance visions will collide at Évian. Amodei wants binding regulation with government blocking authority. Altman wants voluntary commitments with industry-led safety institutes. Macron wants infrastructure investment and a convening role for France. Hassabis, whose Google DeepMind already partners with the UK AI Safety Institute, arrives as the quiet third option: safety testing through national institutes rather than new legislation.
For enterprise buyers, the practical question is which framework wins. Binding FAA-style regulation would slow frontier model releases, increase compliance costs, and potentially reduce the number of viable frontier providers. Voluntary commitments preserve the current pace of releases but leave safety enforcement to the labs themselves. National safety institutes split the difference but create jurisdiction-by-jurisdiction testing requirements that complicate global deployment.
The Hiroshima AI Process Friends Group already spans over 50 countries. Whatever emerges from Évian will not be a final answer, but it will set the direction. When three CEOs controlling the majority of frontier AI capability sit across from seven heads of state controlling the majority of regulatory authority, the conversation stops being theoretical.
The guest list alone tells you where AI governance is headed: out of the conference circuit and into the room where trade wars, nuclear proliferation, and climate finance are negotiated.