NVIDIA’s Computex week has not yet reached the main keynote, but its rolling GTC Taipei page already contains the hardware signal worth watching.
The company says three products won Computex 2026 Best Choice awards: Vera Rubin NVL72, Jetson Thor and Alpamayo. Strip away the awards framing and the product choices are revealing. NVIDIA is pushing at three layers at once: rack-scale AI factories, edge robotics compute, and autonomous-vehicle development stacks.
Vera Rubin NVL72 is the data-centre headline. NVIDIA describes a rack-scale AI supercomputer connecting 36 Vera CPUs and 72 Rubin GPUs through sixth-generation NVLink Switch, ConnectX-9 SuperNICs, Spectrum-X Ethernet photonics switches and BlueField-4 DPUs. The company claims up to 10x higher inference performance per watt and 10x lower cost per token, with 100% liquid cooling and tray designs intended to cut assembly time.
Jetson Thor is the physical-AI part. NVIDIA says the Blackwell-based module delivers up to 2,070 FP4 teraflops, 7.5x the compute and 3.5x the energy efficiency of Jetson Orin, configurable between 40 and 130 watts. That is aimed less at chatbots than robots, medical devices, industrial machines and local autonomy.
Alpamayo is the software-and-data complement for autonomous vehicles: reasoning vision-language-action models, simulation and open datasets for long-tail driving cases. This is NVIDIA bundling hardware with the development environment needed to make it useful.
The caveat is that this is vendor-provided positioning, not independent benchmarking. The keynote is scheduled for 1 June, so more detail may arrive later. Still, the direction is clear enough: the next AI hardware cycle is about watts, racks, cooling, interconnects and deployment environments as much as raw GPU names.