NVIDIA Thor-X: Empowering the Next Generation of Humanoid Robotics

At GTC 2026, NVIDIA CEO Jensen Huang stood alongside a line of autonomous humanoids and declared the beginning of the "Physical AI" era. The centerpiece of this announcement was Thor-X, a system-on-a-chip (SoC) specifically designed to run Vision-Language-Action (VLA) models with the low latency required for real-world interaction.

Thor-X: 2,500 TFLOPS in the Palm of a Hand

Thor-X is a radical departure from NVIDIA’s data center GPUs. While it shares the Blackwell architecture at its core, Thor-X introduces a specialized Proprioceptive Engine—a dedicated silicon block for processing real-time feedback from high-frequency tactile and torque sensors.

Delivering a staggering 2,500 TFLOPS of FP8 performance, Thor-X enables humanoid robots to run "foundation models for motion" locally. This means the robot can perceive its environment, plan its trajectory, and adjust its grip in response to physical resistance—all without a round-trip to the cloud. The SoC also features Safe-Motion Hardware Interlocks, which physically disconnect the motor drivers if the AI’s trajectory exceeds predefined safety envelopes.

The "Humanoid-First" Design Philosophy

NVIDIA isn't just building chips; they are defining the Humanoid Reference Architecture. Thor-X supports the GR00T 2.0 framework, which allows robots to learn from human demonstration in Omniverse and transfer those skills to the real world.

By integrating Silicon Photonics for inter-chip communication, Thor-X can scale across multiple nodes within a robot's body—for example, placing a Thor-X "Lite" chip in each limb to handle local reflexes while the central Thor-X "Pro" manages high-level mission planning. This decentralized nervous system mimics biological organisms and significantly improves the robot's resilience to localized damage.

Strategic Partnerships: Tesla and Figure

The impact of Thor-X is already visible in the industry. Tesla has confirmed that the Optimus Gen 3 will utilize a customized version of the Thor-X SoC for its vision-based navigation system. Similarly, Figure AI has integrated Thor-X into its latest fleet of logistics humanoids, achieving a 40% improvement in sorting speed through enhanced Spatial Reasoning.

The Road to General-Purpose Autonomy

With Thor-X, the bottleneck for robotics has shifted from "compute" to "data." As millions of robots powered by Thor-X begin to interact with the world, the resulting multimodal sensor data will be used to train even more capable VLA models, creating a virtuous cycle of robotic intelligence.

Jensen Huang’s vision of a "robot in every home" may still be years away, but Thor-X provides the essential silicon foundation to make general-purpose physical AI a technical reality.