If the target market is real time robotics, the Jetson Thor system, with 128 GB of memory, but only 273 GB/s of bandwidth, seems over-provisioned with expensive memory. You want the model evaluating at tens of fps, so it can't use more than 10 GB of weights at most. More memory could be used if the model is a wide mixture of experts, or a large strategic planning model is operating in small time slices over a longer period of time, and more memory always makes development life easier, but for cost optimized systems, you should be able to get by with much less.

Excited to share: NVIDIA Cosmos 3 Edge is openly available, bringing frontier world models to NVIDIA Jetson Thor and other local devices for physical AI. Read the blog: Learn more about it at today’s SIGGRAPH keynote—in person or via livestream at 3:45 p.m. PT:

Big day for robotics + AI! 🤖🧠 Huge congrats to @NVIDIARobotics on the official launch of Jetson Thor! ⚡ We’ve had the opportunity to get early access to the NVIDIA Jetson Thor AGX Developer Kit over the past few weeks, and we can confidently say it’s a game-changer for Apollo and for humanoid robotics

San Francisco startup @lightberry has unveiled Lumi: a $40k social humanoid packing an 8-mic Brooklyn-built array, stereo vision, and NVIDIA Jetson Thor compute onto an upgraded Unitree chassis. We chatted with CEO @aliattar to unpack the split-brain architecture, $20k/year software economics, and navigating sweeping US import bans:

Ready to take robots from development to deployment? Join our livestream on August 26 at 9 a.m. PT to learn how to: 🧠 Develop and deploy robots with NVIDIA Isaac GR00T and Jetson Thor 🏭 Apply sim to real workflows for industrial robotics 🦾 Turn teleoperation data into real robot actions 🎙️ See @NobleMachines and @seeedstudio demonstrate practical robotics workflows across humanoids and robotic arms. Watch live:

Robotics deep dive: HUMANOID ROBOT EMBEDDED COMPUTE Every robot company that says anything about its onboard compute quotes a TOPS number. Almost none of them are quoting the same thing. NVIDIA's Jetson Thor markets 2,070. The footnote on NVIDIA's own page: sparse FP4. The previous generation, AGX Orin, markets 275, under a column header reading sparse INT8. Divide one by the other and you get the headline claim for the generation, up to 7.5x more AI compute. Put both on the same footing and it is 3.76x. Dense FP8 against dense INT8, 517 against 137.5. The other half of the speed-up is the number format. NVIDIA states 3.5x better energy efficiency, which reconciles as 2,070 over 130 watts against 275 over 60. At matched precision it is 1.74x. Both headline claims collapse by exactly 2x, by the same mechanism. Sparsity is real, FP4 is real, and NVIDIA prints both in its spec table. Thor is a large generational step. It is roughly half the step the marketing arithmetic describes, and the difference lives in a footnote most people never resolve. Unitree's page for the H2 PLUS states its chip correctly: FP4 2,070 TFLOPS. Search-indexed writeups of the same robot render it as a 2,070 TOPS chip. Floating point silently became integer and the precision qualifier vanished. The spec survived the vendor and died in the coverage. Qualcomm's Dragonwing IQ10 markets up to 700 TOPS and states no precision. Horizon's Journey 6P markets 560 and states none either. Either could be INT8 or INT4, a factor of two, with no document that settles it. Neither publishes a power figure, so neither can be placed on a per-watt axis at all. Horizon does publish one thing nobody else does. The footnote under its TOPS figure reads TPP under 4800. Total Processing Performance is the threshold metric in US export control. That number is bounded by a regulation rather than by the silicon. Figure, 1X, Agility and Apptronik publish no compute specification at all. The only two humanoid vendors that name their chip in a public spec table are Chinese.

Japan’s ugo just unveiled Nova, a wheeled dual-arm robot with mass production planned for 2027. Nova has a 22-DOF humanoid upper body on mecanum wheels, with a 4 kg payload per arm and up to 10 kg using both. Runtime is about six hours. Operators can record demonstrations using force-feedback or VR controllers for AI training. Trained models can run onboard on NVIDIA Jetson Thor. ugo is also offering help with what comes after buying the robot: collecting task data and adapting models for a factory’s handling, picking and assembly work.

Weekly NVIDIA Update JAPAN LAUNCHES NATIONAL AI INFRASTRUCTURE $NVDA will partner with Noetra Corp. to build the world’s first national AI infrastructure for physical AI, featuring 13,750 Vera CPUs and 27,500 Rubin GPUs. Backed by Japan’s METI, the 140MW AI factory will power the FRONTia Project, accelerating multimodal AI models for robotics, manufacturing, logistics and healthcare. NOKIA, NVIDIA LAUNCH AI-POWERED RAN $NOK and $NVDA unveiled an AI-driven radio access network platform that could double telecom data capacity over existing spectrum by 2028. The software-defined system, available next year, aims to improve spectrum efficiency, accelerate 6G readiness and expand AI capabilities at the network edge. NVIDIA EXPANDS EDGE AI PORTFOLIO $NVDA introduced Jetson Thor T3000 and T2000 modules, powered by Blackwell architecture, to accelerate mainstream robotics and edge AI. The new systems deliver scalable on-device AI performance, memory optimization tools and support for embodied AI models, with commercial availability expected in Q1 2027.