NVIDIA’s Jensen Huang declared a “ChatGPT moment” for robotics at CES 2026 on January 6. The headline writes itself—except during the actual keynote, Huang said this moment is “nearly here,” not “is here.” Sound familiar? At CES 2025, he said it was “around the corner.” The timeline keeps shifting, but this year’s announcements suggest NVIDIA robotics is getting closer to something real.
What actually got announced? A lot. NVIDIA unveiled the Vera Rubin platform promising 5x inference gains over Blackwell, the $1,999 Jetson Thor computing module delivering 2,070 TFLOPS, the open-source GR00T N1.6 foundation model for humanoid robots, and partnerships with Boston Dynamics and Caterpillar for real production deployments. The hardware specs are substantial. The deployment timeline remains speculative.
What NVIDIA actually announced
The Vera Rubin platform is NVIDIA’s next-generation AI supercomputer arriving in the second half of 2026. It comprises six new chips: the Vera CPU, the Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet switch. The Vera Rubin NVL72 system combines 72 GPUs into a single rack delivering what NVIDIA claims is 5x greater inference performance than Blackwell with 10x lower cost per token.
The Rubin GPU itself packs 336 billion transistors with eight stacks of HBM4 memory delivering 288GB capacity and 22 TB/s bandwidth. That’s 50 PFLOPS of NVFP4 inference performance and 35 PFLOPS for training. NVLink 6 provides 3.6 TB/s per-GPU fabric bandwidth, with nine switches delivering 260 TB/s total scale-up bandwidth across the system.
For robotics specifically, NVIDIA announced two Jetson Thor options. The flagship delivers 2,070 FP4 TFLOPS with 128GB of memory—that’s 7.5x higher AI compute than Jetson AGX Orin with 3.5x better energy efficiency. The more affordable Jetson T4000 module offers 1,200 FP4 TFLOPS and 64GB memory for $1,999 at 1,000-unit volume, delivering 4x the performance of the previous generation within a configurable 70-watt envelope.
On the software side, NVIDIA released Isaac GR00T N1.6, an open-source vision-language-action (VLA) model purpose-built for humanoid robots. It’s available on Hugging Face and uses NVIDIA’s Cosmos Reason VLM variant supporting flexible resolution and native aspect ratio encoding. The model was trained on diverse robot data including bimanual, semi-humanoid, and humanoid datasets to enable cross-embodiment learning.
NVIDIA also unveiled new Cosmos world foundation models—Cosmos Transfer 2.5 and Cosmos Predict 2.5 for physically-based synthetic data generation, plus Cosmos Reason 2 for vision-language reasoning. These aren’t just research demos: Agility Robotics, Figure AI, Skild AI, and Uber are already using Cosmos to generate training data at scale.
The partnership play: NVIDIA robotics meets production
The most concrete validation came from Boston Dynamics. The company’s fully electric Atlas humanoid robot—featuring 56 degrees of freedom with fully rotational joints and human-scale hands with tactile sensing—will ship to Hyundai’s facilities in 2026. All Atlas units for 2026 are already committed, with fleets headed to Hyundai’s Robotics Metaplant Application Center and Google DeepMind in the coming months.
Hyundai’s broader plan is ambitious: production deployment at Metaplant America in Savannah, Georgia by 2028, with capacity to manufacture 30,000 robot units annually. Boston Dynamics is using NVIDIA’s AI infrastructure and simulation libraries to accelerate physical robotics training, while Google DeepMind is integrating foundation models to give Atlas greater cognitive capabilities.
Caterpillar is piloting AI-powered excavators using NVIDIA’s platform. LG Electronics showcased next-generation robots at CES. Humanoid startups like Agibot and RLWRLD have integrated Jetson Thor into existing platforms. This isn’t vaporware—it’s real hardware shipping to real customers for real use cases.
Huang positioned NVIDIA as building the “Android of robotics”—a full-stack platform with open models, standardized compute, and developer tools. Over 2 million developers are already using NVIDIA’s robotics stack. That’s the kind of ecosystem lock-in that defines platform dominance.
Why the “ChatGPT moment” is both right and wrong
Let’s be fair: there are legitimate parallels to ChatGPT’s breakout moment. NVIDIA is releasing open foundation models that developers can customize, just like GPT-2 and GPT-3 enabled the LLM ecosystem. They’re providing full-stack infrastructure from chips to simulation to deployment tools. They have production partnerships with major manufacturers. The pieces are falling into place.
But physical AI has constraints language models don’t. ChatGPT could scale instantly because it lives in the cloud—tokens are cheap, iteration is fast, and failures are low-stakes. Robots operate in the physical world where inference must happen in real-time, training requires expensive hardware, and mistakes can break things or hurt people.
The safety challenges alone are substantial. A language model hallucination is embarrassing. A robotics hallucination in a factory is a liability event. Boston Dynamics is starting Atlas with “processes with proven safety and quality benefits, such as parts sequencing”—not the general-purpose manipulation tasks Huang demonstrated on stage with BD-1 droids from Star Wars.
Huang claimed robotics will become a “$50 trillion market.” That’s marketing hyperbole, not economic analysis. For context, the entire global GDP is roughly $105 trillion. Does he expect robots to capture nearly half of all economic activity? The number sounds impressive until you think about it for five seconds.
The pattern is clear: at CES 2025, the ChatGPT moment was “around the corner.” At CES 2026, it’s “nearly here.” Next year it might be “almost arrived.” NVIDIA is making real progress on hardware and partnerships, but the timeline keeps sliding right. That’s not necessarily bad—it’s just honest about how hard this problem is.
What this means for the AI industry
NVIDIA is positioning itself as the infrastructure layer for physical AI the same way it dominated training and inference for LLMs. Vera Rubin launching in H2 2026 gives them another 6-12 month lead over competitors. The Jetson Thor pricing at $1,999 makes robotics compute accessible to startups and researchers, not just deep-pocketed enterprises.
The open-source play matters. By releasing GR00T N1.6 on Hugging Face and GitHub, NVIDIA is inviting the developer community to build on their platform rather than proprietary alternatives. This mirrors how world models differ from language models in requiring physical grounding—you can’t train a robot in pure text space.
For companies like DeepSeek pushing efficient AI architectures, NVIDIA’s robotics push is a reminder that inference optimization matters more in constrained environments. DeepSeek’s mHC architecture focuses on computational efficiency—exactly what robotics needs when you can’t scale to a data center.
The real test will be Rubin availability in the second half of 2026 and Atlas production deployments at Hyundai in 2028. Those are the milestones that matter. If Boston Dynamics ships thousands of Atlas units that actually perform useful work in factories without constant human intervention, then we’ll know the ChatGPT moment has truly arrived. Until then, it’s “nearly here.”
NVIDIA’s hardware announcements are impressive. The partnerships are real. The $50 trillion market projection is absurd. The truth is somewhere in between: physical AI is progressing faster than skeptics expected but slower than the marketing hype suggests. That’s not a criticism—it’s just reality when you’re building robots instead of chatbots.
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.
