Onyx and Bria Put World Models, GPU Memory, and Serving Limits Around One Table

World models can expand the ambition of an AI system faster than the hardware underneath it can expand the budget. Onyx and Bria are bringing the people living inside that gap around one table Friday morning.

The coffee is designed for teams training or serving world models on production GPUs. The official description keeps the agenda intentionally narrow: the cost of the memory ceiling, the point where the serving stack gives out, and the infrastructure assumptions that changed during the last year.

That is enough for a serious room. Model capability tends to own the public narrative. Operators inherit the consequences through memory pressure, latency, utilization, orchestration, inference cost, and architectural decisions that can age before the workload stabilizes. The useful comparison is not who has the largest cluster. It is which constraint appeared first, what the team changed because of it, and whether that change still holds at the next scale.

No formal speaker lineup is listed. The knowledge sits with the teams at the table, particularly the places where a promising world model met a very physical limit and the stack had to answer back.

Event details

• Date: Friday, October 9, 2026
• Time: 08:30-11:30 PT
• Location: Civic Center, San Francisco Bay Area, CA
• Format: Breakfast, Brunch or Lunch, Matchmaking, Networking
• Host: Leverage, Tesla
• Registration: Official SF Tech Week event page

Registration status, capacity, meeting details, and program information can change. Use the official event page for the latest organizer updates.