2070 TFLOPs in 130 Watts: NVIDIA Crams a Data Center into a Book-Sized Box
Is Cloud AI Buried Forever?
Is Cloud AI Buried Forever?
NVIDIA is moving generative AI from the "cloud" to real-world machines: the company has announced the availability of the Jetson AGX Thor modules and devkit for serial production—from humanoids and surgical assistants to industrial robots on the assembly line. The platform is designed for local operation without constant reliance on data centers and can run multiple AI models in real-time simultaneously.
At its core is the Blackwell graphics architecture. On paper, the Jetson Thor delivers up to 2,070 TFLOPS of FP4 compute performance with a power consumption of around 130W and comes equipped with 128GB of memory. Compared to the Jetson Orin, it boasts an 7.5x increase in compute performance and a 3.5x gain in energy efficiency. This is critical for tasks where a single robot simultaneously "sees," "understands," and "acts": the device runs large language and vision models in parallel, as well as vision-language-action (VLA) models—including NVIDIA's proprietary Isaac GR00T.
NVIDIA's concept of "physical AI" refers to machines that perceive their surroundings, reason, and react in fractions of a second, without wasting time on network delays. The Jetson Thor is built for precisely this: local generation and inference without constant trips to the cloud. Hence the primary use cases—humanoid platforms, industrial manipulators, robotics for precision agriculture, as well as assistants in surgery, where decisions must be made instantly.
Developers have access to the full NVIDIA stack: simulation and development in Isaac, video analytics in Metropolis, and sensor stream processing in Holoscan. Compatibility with the entire toolkit allows for rapid assembly of pipelines—from training to deployment on hardware.
Major players have already joined the ecosystem. Among the early users are Amazon Robotics, Boston Dynamics, and Caterpillar. John Deere, Meta, OpenAI, and Medtronic are testing it. For logistics, this is a chance to accelerate the adoption of autonomous systems in warehouses; for heavy machinery—to move decision-making on-board for vehicles in quarries and construction sites, saving fuel and reducing errors; for medicine—to ensure the stable operation of assistants where the network is unstable or unavailable.
The Jetson Thor significantly raises the bar for embedded AI: it has enough performance to run several generative models simultaneously, its power budget fits within 130W, and its software ecosystem covers the entire cycle—from simulation to serial deployment. It is precisely this combination that allows complex AI workloads to be moved to the edge and enables robotic systems that can independently navigate a living, unpredictable world. However, it's important to remember that industrial robots require special attention to safety, especially in the context of growing competition in the AI technology market.
