The Biggest AI Summit Yet—But Was It Enough?

Last week, the world gathered in New Delhi for the fourth major AI summit—and the biggest by almost every measure. Bletchley Park, in 2023, was intimate and deliberately invite-only: a small group of governments and frontier labs trying to establish a shared vocabulary for AI risk. Seoul, the following year, kept that focused spirit. Paris, last year, was larger but still bounded.

Delhi was something else entirely: five days, more than 250,000 registered attendees, over 20 heads of state, delegations from more than 100 countries, and 400 official sessions running simultaneously across a venue the size of a small city.

We were there. And the scale was genuinely striking.

So was the timing. Less than two weeks before the summit opened, Anthropic released Claude Opus 4.6, which posted the largest single-generation leap on ARC-AGI-2 ever recorded and led all frontier models on Humanity’s Last Exam. But the mirror of this surge in capabilities is the rise of uncertainty around risks: our ways of understanding and testing for danger—from benchmark saturation to the limits of internal red-teaming—are becoming maxed out. And those capabilities and risks are just the floor.

This was the backdrop against which the AI Impact Summit convened. But India made a calculation: focus on investment, elevate the use cases, give everyone a platform. The result was a summit that traded governance ambition for broad-based access and widespread optimism. We left not in despair—there were genuine wins—but with real concerns about the state of the global AI governance conversation.

Here are our thoughts on what happened, and where the conversation needs to go.