According to Arbitrum Beat monitoring, Huang Renxun officially confirmed during his keynote at GTC Taipei 2026: "Vera Rubin is in full production." He expressed special thanks to the Taiwan supply chain partners present, stating that Vera Rubin's supply chain scale is twice that of the previous generation Grace Blackwell, and the assembly time for a single rack has been compressed from 2 hours in the Grace Blackwell era to just 5 minutes.
Huang Renxun showcased a full video of Vera Rubin's manufacturing process on site. The entire system starts with TSMC's 3nm process, including 7 new chips, system-level aggregation of over 6 trillion transistors, single-board over 18,000 components, and a total of 1.3 million parts (third-generation MGX rack design). The HBM4 memory is sourced from Micron, SK Hynix, and Samsung. The system features a wireless cable PCB midplane design, with all ConnectX-9 SuperNICs and BlueField-4 DPUs integrated on board to ensure AI-grade factory-level reliability. The liquid cooling manifold can withstand over 5,000 amps of current, equivalent to full throttle acceleration of 20 electric cars simultaneously.
He also confirmed that Microsoft, Dell, and CoreWeave have been the first to complete the deployment and operation of the Vera Rubin NVL72 engineering sample. Millions of square feet of capacity have come online to support Grace Blackwell shipments, and preparations are currently underway in parallel for Vera Rubin ramp-up, with large-scale shipments set to ramp up in the second half of 2026.
In addition, Huang Renxun also showcased the Vera CPU rack (a single liquid-cooled rack with 256 Vera CPUs, responsible for model orchestration and memory scheduling) and the new Groq 3 LPX low-latency inference rack (256 Groq 3 LPUs, 40 PB/s SRAM bandwidth). NVL72 focuses on high-throughput token generation, while Groq 3 LPX specializes in the lowest latency token generation, complementing each other.
