The scarcest thing in AI infrastructure is not a GPU, a substation, or a water permit.
It is a room where the people who sign for capacity and the people who build it say what they think, with the logos left at the door.
We brought sixteen guests together for four hours in Hayes Valley under Chatham House rules.
Each was admitted based on a single sentence typed into the registration form answering one question:
What is one thing you know about the next fifteen years of AI infrastructure that most people get wrong?
Fifteen gave a real answer.
Zero named the chip.
Instead, they named:
- thermal physics
- interconnect bandwidth
- edge inference for robotics
- real-world operational data
- trust
- distribution
- heat reuse
- grid interconnection queues
Out of thirteen distinct bottlenecks raised, not one is a physics problem that money alone fixes—they are all coordination problems between humans and institutions.
Key Takeaways from the Table
The Company is the Product: When asked to choose between focusing on the self-running building or the self-running operator, the room overwhelmingly chose the operator.
The challenge isn’t just an autonomous data center; it’s an operating company whose head-count does not scale linearly with megawatts.The 90-Day Test, Redefined: Running ninety quiet days without a human inside proves very little.
The true kill-bar for autonomous operations is verifiable recovery from a real failure with a clean audit log and zero human escalation required.Inference at the Edge: Robotics and real-time agents require small GPU regions situated close to physical activity, shifting infrastructure demands from giant centralized campuses toward dynamic, API-driven regional capacity.
As intelligence advances faster than the physical grid, success belongs to operators who build lean, publication-ready audit logs and deploy autonomous fleet operations.