AI Infra Summit: NVIDIA touts Vera Rubin, DSX and 23% perf-per-watt gains for AI factories

nordicinst · x · 2026-09-16

NVIDIA's Ian Buck spoke at AI Infra Summit (attendance up from 3,500 to 8,000+) on AI factory efficiency, arguing the key metric is shifting from peak performance to tokens per watt.

New announcements: Amazon's Annapurna Labs collaborating on NVHBM custom HBM; d-Matrix integrating Raptor XPUs with NVLink Fusion; Emerald AI and Silicon Valley Power demoing a commercial flexible-load program; Lambda improving performance per watt by 23% with DSX MaxLPS; Pinterest using Blackwell and Dynamo for conversational visual discovery. NVIDIA positions its full-stack AI factory platform — Vera Rubin systems, Dynamo, NeMo, NVLink, Spectrum-X, ConnectX SuperNICs and BlueField DPUs — as the answer to agentic AI's growing demands.

Related event: NVIDIA Unveils DSX AI Factory Platform Focused on Tokens-per-Watt Efficiency(3 posts)→

Original post →

More from Infra

Infra channel →