The Technologies Behind NVIDIA’s Four COMPUTEX 2026 Best Choice Awards

Jetson Thor: Edge AI Power for Robotics and Physical AI

Jetson Thor represents NVIDIA’s next generation of edge AI computing, designed to power intelligent machines such as robots, industrial systems, and autonomous devices.

Built on the NVIDIA Blackwell GPU architecture, the Jetson AGX Thor developer platform delivers up to 2,070 FP4 TFLOPS of AI compute with 128 GB of memory, operating within roughly a 130‑watt power envelope. Compared with the previous Jetson AGX Orin platform, it provides up to 7.5× more AI compute and about 3.5× better energy efficiency.

This performance jump enables a new class of applications often described as “physical AI.” Instead of relying heavily on cloud servers, robots can run advanced models directly on the device, including vision‑language and reasoning models that interpret sensor data in real time.

Key improvements include:

• High‑performance edge inference for generative and reasoning models
• Real‑time processing of multi‑sensor inputs such as cameras and lidar
• Power‑efficient operation suitable for mobile robots and industrial systems

For robotics developers, this means machines can perform more complex perception, planning, and interaction locally—reducing latency and dependence on remote compute resources.

Vera Rubin NVL72: Rack‑Scale Infrastructure for AI Factories

While Jetson Thor focuses on the edge, Vera Rubin NVL72 targets the opposite end of the spectrum: large‑scale AI infrastructure.

The system is designed as a rack‑scale AI supercomputer integrating 36 NVIDIA Vera CPUs and 72 Rubin GPUs connected through sixth‑generation NVLink so that the GPUs operate as a unified accelerator.

This architecture is optimized for the next generation of AI workloads, including trillion‑parameter models, post‑training, large‑scale inference, and long‑context reasoning tasks.

NVIDIA reports several major efficiency improvements compared with the previous Blackwell platform:

• Up to 10× higher inference performance per watt
• Up to 10× lower cost per token during inference
• Large model training with roughly one‑fourth the number of GPUs

These gains come from deep system‑level integration. The platform combines multiple specialized chips—including the Rubin GPU, Vera CPU, NVLink 6 switch, ConnectX‑9 SuperNIC, BlueField‑4 DPU, and Spectrum networking—into a single architecture designed to operate as one AI system.

Another reason the system received the Sustainable Tech Special Award is its infrastructure design. Modular trays and liquid‑cooling systems improve deployment efficiency and data‑center energy performance.

In effect, Vera Rubin NVL72 reflects a shift in the AI race: performance is no longer just about individual GPUs but about entire racks—or even AI factories—engineered as unified computing systems.

Alpamayo: Reasoning‑Based AI for Autonomous Driving

The third award‑winning technology, Alpamayo, addresses one of the hardest problems in AI: safe and reliable autonomous driving.

Alpamayo is an open development platform and family of models for autonomous vehicles, designed to move beyond traditional perception‑focused driving systems. Instead of only detecting objects, the models use vision‑language‑action (VLA) reasoning to interpret road situations and decide how to respond.

For example, the Alpamayo 1 model can process video input and generate:

• driving trajectories (steering, braking, acceleration)
• reasoning traces explaining the decisions made

This approach is intended to make self‑driving systems more transparent and easier to analyze, which is crucial for testing safety in rare or complex driving scenarios.

Another important feature is the training pipeline. Alpamayo models are trained using a combination of:

• real‑world driving demonstrations
• large synthetic datasets generated through simulation
• carefully annotated driving examples

Combining real and simulated data helps expose models to far more edge cases than road testing alone could provide.

A Full‑Stack Strategy for the AI Era

Taken together, the three award‑winning technologies represent NVIDIA’s broader strategy for the AI era.

• Jetson Thor brings powerful AI reasoning directly to robots and edge devices.
• Vera Rubin NVL72 provides the massive infrastructure needed to train and run trillion‑parameter models at scale.
• Alpamayo applies reasoning‑based AI to autonomous vehicles using data pipelines that combine real and synthetic environments.

The combination suggests a future where AI systems operate across multiple layers: training in massive AI factories, deployment in data centers, and real‑time decision‑making on edge devices and autonomous machines.

That full‑stack approach—spanning chips, racks, and intelligent agents—is a major reason NVIDIA emerged as one of the most prominent winners at COMPUTEX 2026.