Jensen Huang’s physical AI vision: what changed from 2024 to 2026?

Jensen Huang was talking about physical AI well before the latest wave of robot announcements. At COMPUTEX in June 2024, he called it the next wave of AI. That matters when reading a fresh headline: the broad ambition is a continuing strategy, not an idea that appeared this year. NVIDIA’s account of the address

“The next wave of AI is physical AI.”

Jensen Huang, June 2, 2024, as quoted by NVIDIA.

The question worth following is how that ambition turns into a usable development process. A robot needs more than an answer to a question. It has to choose an action, carry it out with real hardware and cope when the world does not match its training examples.

From an idea to a development toolkit

The January 2025 launch of Cosmos gave the story a clearer mechanism: world foundation models and synthetic data intended to help developers train and evaluate physical systems. NVIDIA presented it as a way to reduce barriers to robotics development. That was a statement about tools available to builders, not evidence that every builder could ship a general-purpose robot. Cosmos launch

In its June 1, 2026 investor announcement, NVIDIA described Cosmos 3 as bringing visual reasoning, world generation and action prediction into one system. The progression is toward connecting more of the work between understanding a scene and deciding what to do in it. The performance and productivity claims in that announcement remain NVIDIA’s claims. Cosmos 3 announcement

The missing link is the task

Imagine a robot sorting parts beside a production line. Generating varied training scenes could help it encounter an unfamiliar part before that part arrives at the factory. But a convincing simulated scene does not establish that the gripper can hold a slippery object, recover from a failed pick or work at the required pace.

For a buyer, the comparison is with the process already in use. A more flexible robot may be valuable if it handles changes that previously required expensive reprogramming. If it needs frequent rescue, the cost of that support belongs in the comparison too. These are the practical tests we would apply; they are not measured results for Cosmos.

September evidence: what Skild adds

In a September 10 account published by NVIDIA, Skild AI CEO Deepak Pathak described Isaac Lab and Cosmos as tools for creating varied robot training experiences. NVIDIA outlined their roles in simulation and training-data preparation. Skild was already named among Cosmos adopters in the January 2025 announcement. The useful addition is a more specific account of the work, rather than a newly discovered relationship. These are statements from the companies involved. NVIDIA’s September 10 account The earlier adopter list

Skild’s own S1 research makes the distinction between learning and reliability clearer. Its pretrained model takes a video demonstration as a prompt for a new task without another training run. In an internal comparison using 100,000 hours of pretraining data, Skild reports 66% for video prompting versus 9% for language prompting, measured as average cumulative success across task steps. The evaluation allows human recovery after failures; those percentages are not uninterrupted whole-task completion rates. S1 research and evaluation method

Our reading: this gives Huang’s platform argument a concrete example of use, while leaving its commercial strength open. These accounts do not isolate Cosmos 3’s contribution to the results or quantify the NVIDIA revenue attributable to this work. Repeated customer results and disclosed spending would move the argument further.

What Isaac ROS 5.0 adds

NVIDIA’s September 22 announcement adds a practical development step to the platform story. Isaac ROS 5.0 is available software, with reusable workflows for AI assistants helping developers set up and adapt robot applications. That is a more specific claim than a promise of future general-purpose machines. NVIDIA’s announcement

The version notes carry a September 21 date. They identify environment setup and code migration as uses for the new agent skills. Existing applications that call NVIDIA’s older NITROS interfaces directly require source changes. The notes also flag image-preprocessing performance regressions on listed platforms and compatibility limits affecting some motion-planning examples. A team still has to check its own robot, sensors and software together. Version notes and known limitations

NVIDIA says Magna uses Isaac ROS for perception, data collection and robot-model deployment, alongside Isaac Sim testing connected to real hardware. This identifies a participant and a workflow; the account does not provide a before-and-after measure of factory output or costs. It also does not establish that every named partner has already adopted version 5.0. The partner account

Our reading: these tools could reduce the work between a model demonstration and a usable application. Evidence that would strengthen that case includes measured integration time, reliable operation after migration and repeated customer use. Released code is progress. Its effect on a working production line still needs to be measured.

A builder’s October account: simulation still needs real boxes

Ariana Eisenstein, Pickle Robot’s CTO, describes a more specific use of NVIDIA’s tools in an interview published October 7. The conversation happened in September, not on publication day. She names Isaac Sim and Isaac Lab, and says Pickle recreates real warehouse problems in simulation, then keeps the repaired scenarios as tests. Cardboard deformation and humidity are among the details that need realistic parameters. The interviewer, Justin Hodges, discloses that his employer CoreWeave does business with Pickle. This is participant testimony, not an independent performance study. The interview and disclosure

Pickle’s own architecture page describes a learned policy bounded by classical control, with separate monitoring and a supervisor that can request help. It also retains a teleoperation interface. That makes the engineering around the model part of the adoption story. Pickle’s architecture description

This adds a builder’s explanation of how simulation enters testing. It does not isolate NVIDIA’s contribution to customer output or costs. A useful next comparison would measure accepted packages and interventions across comparable shifts before and after a specific change. For NVDA readers, tool use and incremental spending remain separate questions.

October: controlling the agent is a different test

Huang’s London Build-a-Claw appearance was on October 14, according to NVIDIA’s dated event account. TechRadar’s October 15 report connects his agent argument to robotics: a machine’s arms and motors become tools used by software. His emphasis is on restricting permissions and monitoring actions. The publication date is not the event date. NVIDIA’s London entry Mike Moore’s report

OpenShell’s documentation describes controls over files, networks and processes. That explains the software boundary. It does not establish that a moving robot will stop before hitting someone. The documented runtime controls

There is a narrower physical-action experiment to examine. In an August 7 research note, NVIDIA’s Alex Watson describes a simulated pick-and-place task: a separate checker can reject a proposed path and return constraints for a revised plan. The prototype checks human distance before motion; continuous monitoring during execution is listed as future integration. It does not model full-body geometry, braking distance or perception uncertainty, and the author explicitly says it is not safety-rated. This is historical prototype evidence, not a newly certified robot product. The experiment and its limits

Our reading: the platform story now includes who can authorize an action, alongside how a model learns it. A useful next disclosure would show those controls working on real hardware when sensor information is wrong or a person moves unexpectedly, with measured stopping behavior and human interventions. A software permission rule alone cannot answer that operating question.

What an investor can take from this

NVIDIA is directly connected to the theme through its computing and software platform. Its stock is identified as NVDA in the cited company materials. That establishes a business relationship with physical AI; it does not tell us how much additional revenue any one robotics announcement will create.

The next useful disclosures would show customers using the tools repeatedly, a shorter development cycle measured against a stated baseline, or a deployment that performs reliably outside a demonstration. Broad adoption across different robot builders would support the platform argument more strongly than a long list of people standing on a stage.

For near-term readers, a model launch is a reason to examine the details. A verified change in customer spending or deployment economics would be a different—and more financially specific—kind of evidence. Until those details exist, translating a technical announcement directly into a profit forecast goes further than the sources support.