Tech titans caution G7: frontier models are too dangerous to ignore

Sam Altman of OpenAI, Dario Amodei of Anthropic and Demis Hassabis of Google DeepMind were among the tech titans who joined G7 heads of state for a working lunch in Évian-les-Bains, France, on the final day of the summit. The official theme was “ensuring a safe, rapid and effective deployment of artificial intelligence.” What emerged was a collective warning from the people building the most capable AI systems that those systems are now too consequential and too dangerous to be governed by the companies building them alone.

“The technology’s future must be shaped by people, democratic institutions and society as a whole, not just by the companies building the most capable systems,” Altman told the gathering, attended by President Donald Trump and other G7 leaders. 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.”

It was not a reassuring message, even if it was intended to be. The CEO of the world’s most prominent AI company was, in effect, telling world leaders that the industry he leads cannot be trusted to regulate itself.

What Amodei and Hassabis pushed for behind closed doors

The public remarks were only part of the story. In a closed-door session, Amodei and Hassabis are reported to have gone further, calling for a US-led coalition to shape global AI rules and standards.

Amodei’s pitch centred on structured access to frontier models, restrictions on chip and component trade that would exclude China, and international cooperation on what he described as the most serious risk categories: cyber operations, bioterrorism and intelligence. Recent releases of powerful AI models with advanced cyber capabilities prompted fresh concerns from businesses and governments about digital security weaknesses.

The timing of Amodei’s call for US leadership was not lost on observers. It came five days after the Trump administration forced Anthropic’s latest models offline to comply with export controls, restrictions the company said it did not believe were warranted. Macron called it a “good thing” that US officials recognised frontier models could be dangerous, but criticised the move as a “strictly nationalist” reaction.

The infrastructure question nobody answered

The G7’s AI debate is increasingly inseparable from questions about where compute lands, which jurisdictions control it, and who can access it, and that is now a live question, not a future policy consideration.

Sovereign AI infrastructure has become one of the defining phrases of 2026. Governments are committing billions to build it, operators are being asked to host it, and enterprises are beginning to require it in procurement.

Capacity has covered this trend across multiple markets: from the BT and Nscale partnership targeting 14MW of sovereign AI data centre capacity across UK sites, to SoftBank’s October 2026 GPU cloud launch positioning sovereign compute as a commercial differentiator in Japan, to the fragmentation of cloud strategy across Asia as markets from India to Singapore tighten data localisation rules.

Training and running frontier models swallow enormous amounts of compute and electricity, and the companies are signing power and data centre deals that rival the budgets of small countries. Governments increasingly want a say in where that hardware lands. The G7 summit made that pressure visible at the highest level of global politics, but produced no framework for how it should be resolved.

What the summit did clarify is the direction of travel. The US has made clear its opposition to multilateral agreements that could undermine its ongoing industrial advantage, while the EU and Canada are moving to reduce, not strengthen, their ties to US tech companies. If those positions harden into policy, data centre operators with cross-border footprints will face a more fragmented procurement environment, not a simpler one. One where the model you can deploy, the chips you can buy, and the jurisdiction you can serve are all subject to political conditions that did not exist two years ago.

OpenAI’s Sam Altman: ‘AI could surpass human intelligence by 2030’

Microsoft takes over Stargate Norway data centre as OpenAI stalls investments