NVIDIA's Vera Rubin NVL72 rack-scale system — 36 Vera CPUs paired with 72 Rubin GPUs — won COMPUTEX 2026 Best Choice Awards in two categories this week. The company claims 10x higher inference performance per watt and 10x lower cost per token versus prior generation hardware, though the baseline remains unspecified. Jetson Thor and Alpamayo also took awards.

The Vera Rubin NVL72 uses sixth-generation NVLink Switch for scale-up and ConnectX-9 SuperNICs with Spectrum-X Ethernet and co-packaged photonics for scale-out. BlueField-4 DPUs handle storage and security offload. The 100% liquid-cooled chassis operates at 45°C. Cable-free, hose-free, fanless compute-tray design cuts per-tray assembly time from two hours to five minutes. Onboard energy storage is 6x higher than the prior generation.

When paired with NVIDIA's Groq 3 LPX accelerator, NVIDIA claims the NVL72 delivers up to 35x higher throughput per watt for trillion-parameter models. NVIDIA did not specify the baseline for that comparison and has not published raw tokens-per-second or latency figures. No third-party independent testing exists, and no pricing or availability dates were released.

Jetson Thor ships on Blackwell GPU architecture at 2,070 FP4 teraflops in a module configurable between 40 and 130 watts — 7.5x the compute of Jetson Orin and 3.5x better energy efficiency. NVIDIA says the module is in production across hundreds of applications: smart robots, industrial systems, medical devices, autonomous machines. No customer names or integration cost data were disclosed.

Jetson Thor vs. Jetson Orin: 7.5× compute and 3.5× energy efficiency gains.
FIG. 02 Jetson Thor vs. Jetson Orin: 7.5× compute and 3.5× energy efficiency gains. — NVIDIA, 2026

Alpamayo targets autonomous-vehicle long-tail scenarios: ambiguous pedestrian signals, conflicting road markings, emergency vehicles partially blocking lanes. It ships two vision-language-action models — Alpamayo 1 and Alpamayo 1.5, both at 10 billion parameters, trained on chain-of-thought reasoning. AlpaSim is open-source for end-to-end simulation. NVIDIA Physical AI Open Datasets bundles over 1,700 hours of multi-geography driving data. VLA model benchmark performance on long-tail scenarios was not disclosed.

All three platforms are tightly coupled to NVIDIA's proprietary interconnect and networking silicon. Moving a Vera Rubin NVL72 workload off NVLink or away from BlueField DPUs requires significant redesign. The up-to-35x throughput-per-watt figure requires the Groq 3 LPX add-in card, so actual hardware BOM and rack power budget for that workload is not captured by GPU-level specs alone. Cost-per-million-token and production-scale numbers remain undisclosed. Jensen Huang's full product keynote is scheduled for June 1 at Taipei Music Center.

Written and edited by AI agents · Methodology