NVIDIA GTC Taipei at COMPUTEX: Live Updates on the Next Phase of AI

NVIDIA GTC Taipei at COMPUTEX, held in Taipei, Taiwan, brought together the latest trends around AI factories, scaling infrastructure, agentic AI, and physical AI. CEO Jensen Huang's keynote will take place at 11:00 AM (Taiwan time) on June 1, alongside the pre-event program GTC Live at Taipei 2026.

At the COMPUTEX Best Choice Awards, NVIDIA's major product lineup won multiple awards in succession. Vera Rubin NVL72 received the Golden Award and the Sustainable Tech Special Award, Jetson Thor received the Golden Award, and Alpamayo won in the Vehicle Technology and Smart Cockpit category.

Vera Rubin NVL72 is a rack-scale AI supercomputer that links 36 Vera CPUs and 72 Rubin GPUs via NVLink Switch, combined with ConnectX-9 SuperNIC, Spectrum-X Ethernet Photonics co-packaged optics switch, and BlueField-4 DPU. NVIDIA stated that this system delivers 10x improvement in inference performance per power, 10x reduction in cost per token, and, when combined with Groq 3 LPX, a 35x improvement in throughput per power for trillion-parameter models.

The design also brings major changes.

• A cable- and hose-free modular tray structure reduces assembly time from 2 hours to 5 minutes.
• Onboard energy storage capacity for power smoothing has been increased 6x.
• A 100% liquid-cooled design operates even in 45 degrees Celsius environments, drops directly into existing liquid-cooled data centers, and can reduce cooling overhead.

Jetson Thor is an edge AI platform aimed at physical AI and autonomous robots, delivering up to 2,070 FP4 teraflops. This represents a 7.5x improvement in compute performance and a 3.5x improvement in energy efficiency compared to the previous Jetson Orin generation, and it can be configured within a 40W~130W range.

Alpamayo is an open reasoning platform for autonomous driving development, focused on handling long-tail driving scenarios such as ambiguous hand gestures, intersections where signals and lane markings conflict, and passing emergency vehicles that partially block a lane. The platform includes the 10-billion-parameter chain-of-thought vision language action models from Alpamayo 1.5 and Alpamayo 1, the high-precision simulation framework AlpaSim, and the NVIDIA Physical AI Open Datasets, containing over 1,700 hours of driving data.