Robots came to the keynote.
Not concept renders, not a sizzle reel of simulated warehouses. At GTC 2026, Jensen Huang stood on stage and showed off hardware from Boston Dynamics, Caterpillar, Franka Robotics, Humanoid, LG Electronics, and NEURA Robotics. Alongside that, Nvidia released new open models, frameworks, and AI infrastructure aimed squarely at physical AI. Shares got a boost on the robotics story, and the testing-partner corner of Nvidia’s ecosystem is riding the same wave.
I build bots for a living, mostly software agents, occasionally things with wheels. So let me tell you what this announcement actually means for people like us, as opposed to what it means for a stock chart.
Open models are the part that matters
The robot parade is the headline. The open models are the substance. If you have ever tried to get a manipulator to pick up an unfamiliar object, you know the problem is not the motors. It is perception, policy, and the gap between the two. Every team ends up building its own stack of grasp detection, pose estimation, and some hand-tuned controller that works beautifully in your kitchen and falls apart in someone else’s.
Open foundation models for physical AI change the starting line. Instead of training a policy from scratch on a dataset you collected yourself over three months, you start from something pretrained and adapt. That is the same shift we already lived through with language models. In 2019 you fine-tuned BERT. By 2023 you were writing prompts. Robotics is somewhere around the fine-tuning stage right now, and Nvidia is pushing hard to make that the default workflow.
Six companies, six very different robots
Look at the partner list again, because the spread is the interesting signal:
- Boston Dynamics — legged and mobile platforms, the hardest control problems in the field
- Caterpillar — heavy industrial machinery, where a mistake costs more than a dropped coffee cup
- Franka Robotics — research-friendly arms that show up in a lot of university labs
- Humanoid — general-purpose bipeds, the long bet
- LG Electronics — consumer and appliance scale, which means volume manufacturing
- NEURA Robotics — cognitive robotics aimed at collaborative work
Those are not six variations on one product. Those are six separate markets with different safety requirements, different duty cycles, and wildly different budgets. One model family touching all of them tells you Nvidia is building a platform, not a robot.
The industrial half nobody tweets about
Huang also called out partnerships with Siemens, Cadence, and Synopsis to pull AI, simulation, and agentic systems into the industrial life cycle. This is the least glamorous part of the announcement and possibly the most consequential.
Here is why I care as a builder. Simulation is where robot policies get trained now, because collecting real-world data is slow and breaking real hardware is expensive. If simulation tooling connects directly to the design and planning software that factories already run, your training environment stops being a toy physics sandbox and starts being a digital copy of an actual production line. That closes a loop. Design the cell, simulate the cell, train the policy in simulation, deploy to the cell.
Cadence and Synopsis sitting in that list also means chip design workflows are in scope. Agentic systems designing the silicon that runs the agents is a loop I expect we will be talking about for years.
Someone has to run all this
The enterprise side brought its own crowd. Hitachi Vantara, Hewlett Packard Enterprise, Lenovo, and VAST Data all showed up with solutions for building and managing what Nvidia calls AI factories. IBM announced an expanded collaboration on March 16, 2026, focused on helping enterprises operationalize AI at scale.
Unsexy, necessary. Physical AI generates absurd amounts of data: video from every camera, telemetry from every joint, simulation rollouts by the millions. Storage and orchestration are not afterthoughts in that pipeline, they are the pipeline. If you have ever had a training run die because your data loader could not keep the GPUs fed, you understand why VAST Data is in this conversation.
What I would actually do with this
If you are building bots and wondering where to spend your next month, my read is simple. Get comfortable with simulation-first development. Learn to fine-tune a pretrained policy rather than training one. Treat your data pipeline as a first-class part of the project, not plumbing you bolt on later.
The stock movement is a side effect. The actual story is that the tooling for physical AI is consolidating fast, and the gap between a research demo and a deployed robot is narrowing from years to months. That is good news if you are building. It is also a reminder that the skills worth having are shifting from writing controllers to curating data and shaping behavior.
Six robot companies on one stage, running one model family. That is the detail I keep coming back to.
đź•’ Published: