2026-09-05 · Ian Lapham
Nobody Buys a Robot From NVIDIA
NVIDIA supplies the compute, simulation, models, and tools behind much of the humanoid market, making robot makers its customers rather than competitors.
Read the humanoid coverage from the last year and the scoreboard is fixed. Tesla Optimus, Figure 03, Unitree's G1 and H2, UBTech, AgiBot, Galbot, and the rest of the Shenzhen and Shanghai cluster. NVIDIA appears in these stories as a supplier, usually one line down in the bill of materials, sometimes as the company whose GTC stage the robots walk across. That framing treats NVIDIA as a contestant with a bad entry. It has no robot, so it must be losing.
Count the layers instead. A humanoid that leaves the lab needs onboard compute that fits in a torso, a simulator to train and test in, synthetic data and a world model to fill the gaps real data leaves, a base policy to fine-tune from, and a way for a warehouse worker to talk to it. Five layers. Name the owner of each one and the humanoid race stops looking like a race and starts looking like a customer list.
PersonaPlex is the freshest tell. NVIDIA's Applied Deep Learning Research group published it as a full-duplex speech-to-speech model, the kind that listens and talks at the same time and handles interruptions like a person does. It descends from Kyutai's open Moshi architecture and the Helium language backbone. The code ships under MIT. The weights ship under the NVIDIA Open Model License (https://research.nvidia.com/labs/adlr/personaplex/). A company that wanted to sell voice would have charged for it. NVIDIA gave it away because it is the human-facing top of a stack whose lower layers it already sells. Nobody buys a robot from NVIDIA. Every robot buys from NVIDIA. That is the pickaxe business, and it is worth taking seriously as a business rather than as a footnote to someone else's demo.
The stack a humanoid actually needs
Layer before logo. Start at the bottom.
Onboard compute. Jetson AGX Thor is the module: a Blackwell GPU, 128 GB of memory, roughly 2,000 FP4 TFLOPS in a power envelope that tops out around 130 watts (https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/). This is the one layer NVIDIA does not give away. It is a physical board with a price, a lead time, and an allocation. Every other layer exists to make this board the obvious thing to put in the chest cavity.
Simulation. Isaac Sim is built on Omniverse and speaks USD, the scene format Pixar open-sourced. Isaac Lab sits on top for reinforcement learning and imitation learning at scale. Isaac ROS wraps the perception stack for ROS 2 (https://developer.nvidia.com/isaac). Isaac Sim itself went open source in 2025. What NVIDIA sells here is the RTX workstation or the DGX node the simulator runs on, and the fact that the whole pipeline assumes CUDA underneath.
Synthetic data and world models. Cosmos is a family of world foundation models: Predict for generating video futures, Transfer for turning a simulation render into something that looks like a real camera feed, Reason for embodied planning (https://www.nvidia.com/en-us/ai/cosmos/). Open weights, again under the NVIDIA Open Model License. The money is in the GPU hours it takes to run them and the training runs it takes to post-train them.
Base policy. GR00T N1 and its successors are open humanoid foundation models, a vision-language-action system with a slow reasoning module and a fast motor module, shipped with the post-training recipe and the simulation environments used to build it (https://developer.nvidia.com/isaac/gr00t). An OEM downloads it, collects a few hundred hours of teleoperation, fine-tunes in Isaac Lab, and deploys to Thor. Free to download. Expensive to run anywhere except the hardware it was tuned for.
Human interface. PersonaPlex. Seven billion parameters, full-duplex, persona-conditioned by a text prompt and a voice sample. The layer a floor supervisor touches when the robot pauses mid-task and asks which pallet. Free.
The Robot Report's rundown of physical AI infrastructure platforms puts NVIDIA in the accelerated compute and simulation slot and treats everyone else as building on or around it (https://www.therobotreport.com/5-physical-ai-infrastructure-platforms-to-watch/). That is the outside read, and it matches the map. The logos in the humanoid race sit downstream of every one of these five boxes.
Give away the model, sell the module
The first mechanism is simple once you separate the artifact from the thing that runs it.
A trained model is a file. Copying it costs nothing. Charging for it invites a fork, and in speech the fork already exists because Kyutai published Moshi under open licenses in 2024. So NVIDIA does not charge. It tunes the model for its own silicon, documents the inference path on Jetson, and publishes it. The value moves from the file to the place the file has to live.
And that place is constrained in ways a data center is not. A humanoid torso has a thermal budget set by battery chemistry and by the fact that a person is standing next to it. A conversational speech model has a latency budget set by human patience, somewhere under 300 milliseconds before the exchange feels broken. Running a 7B full-duplex model at that latency, on battery, alongside a vision-language-action policy and a perception stack, is a hardware problem with a short list of answers. Jetson Thor is on the list. Most other boards are not.
So the paraphrase test for this layer reads: download the voice model for free; the board costs money and has a waitlist. The same sentence works for GR00T and for Cosmos. Generosity has nothing to do with it. NVIDIA prices the stack at the one layer that cannot be copied, and it makes the free layers easiest to use on exactly that layer.
Simulation is where the lock-in lives
The second mechanism is the one the "commodity supplier" read misses, and it matters more than the board.
Switching cost in this business lives above the compute. Boards get swapped. The cost is the pile of work an integrator has already paid for on top of the simulator: USD asset libraries for the customer's factory, hundreds of Isaac Lab environments with tuned reward functions and domain randomization ranges, ROS 2 bindings that assume Isaac ROS message types, and the sim-to-real validation runs that took a policy from a render to a machine that stacked the right tote on a pilot line without knocking over the wrong one. That validation is the product the OEM actually delivers to its customer. Everything else is scaffolding for it.
Look at deployment tempo. An OEM that validated a GR00T-derived policy in Isaac Sim and then proved it on a factory pilot does not go back and re-validate in MuJoCo to save a few hundred dollars on a board. The numbers do not work. A Jetson Thor developer kit lists around $3,500, and production modules come in lower at volume. One re-validation cycle at a factory pilot is a few robotics engineers for a quarter, plus the customer's line time, plus the risk that the new simulator disagrees with the old one about contact physics in some corner case that only shows up on the fortieth pallet. That is engineer-months, not parts. The board is a rounding error against it.
This is also why NVIDIA open-sourced Isaac Sim itself. Once the pipeline is the moat, the simulator being free does not cost anything. Every USD scene an integrator builds is a scene that runs best on the GPUs NVIDIA sells, and every validated policy is a reason to keep the target board the same next year. Robots don't compound. Pipelines do.
The miners are in China
Now name where demand sits, because it is the strength and the exposure in one sentence.
Reporting derived from the Wall Street Journal in late August 2026 described Chinese humanoid and mobile-manipulation makers as the most eager buyers of Jetson, Isaac, and Cosmos, put China at a majority of humanoid shipments in the first half of 2026, and attributed a figure of around $10 billion in physical AI annualized revenue to Jensen Huang's own comments (summary at https://aiweekly.co/). Treat that last number carefully. It is attributed, it is secondary, and nobody outside NVIDIA has audited it. But the shape of the claim matches what anyone walking the floor at the World Robot Conference in Beijing would have seen: Jetson modules in torsos, Isaac logos on partner slides, GR00T in the training pipeline of companies most Western readers have never heard of.
This is good for a pickaxe seller. More miners, more picks. Unitree, AgiBot, Galbot, UBTech, Fourier, LimX, and a long tail of Shenzhen shops are the largest single camp of customers for the stack described above, and they are shipping in volume while Figure and Tesla are still counting units in the low thousands or fewer.
But a pickaxe seller is only as healthy as its miners, and the largest camp sits under U.S. export controls that have already moved the line twice, and under a domestic-substitution push that Beijing treats as industrial policy rather than preference. Huawei's Ascend line exists for exactly this reason. Chinese simulators exist for exactly this reason. If the compute layer becomes a policy variable, the whole stack's revenue from its largest customer geography becomes a policy variable too, and no amount of open licensing on the upper layers fixes that. Concede it fully. The Chinese OEMs are the customers, they are building real machines, and the risk here belongs to the supplier, not to them.
Objection: pickaxes commoditize
An integrator or an OEM engineer reading this will push back, and the pushback is strong. In their voice:
"Pickaxe sellers get squeezed once the miners are big enough. Unitree and the Shenzhen cluster will design Jetson out for Huawei Ascend or a domestic board the moment volume justifies it. Isaac Sim competes with MuJoCo, which DeepMind maintains and gives away, and with Genesis, which is faster on some workloads. PersonaPlex descends from Kyutai's open Moshi, so the speech layer is already commons. Every layer on your map has a free or domestic substitute. The map proves the opposite of what you want it to."
Take the part-level point. It is correct. Any single layer is substitutable. Boards get swapped, simulators get swapped, and in China substitution is national policy backed by procurement, not just an engineering preference. Anyone pricing NVIDIA's robotics business as five unassailable monopolies is wrong.
But the map was never about parts. The substitute for one layer exists. The substitute for the validated path through all five does not. Re-validating a GR00T-derived policy on a new simulator, a new board with different memory bandwidth and a different inference runtime, and a new speech front end with different latency characteristics, is a multi-quarter project at fleet scale, and it has to be repeated for every SKU and every customer site. An OEM shipping fifty robots a year can absorb that. An OEM shipping five thousand cannot, at least not while also trying to ship. Integration cost is the moat. The parts are just where it happens to accrue.
There is a scope condition, and it should be stated plainly rather than waved past. The thesis holds while sim-to-real is the bottleneck. Right now it is. Humanoids do not have enough deployed hours to learn from their own fleets, so synthetic data and simulation do the heavy lifting, and whoever owns the simulation pipeline owns the compounding. If that flips, if deployed-fleet data starts outweighing synthetic data for the tasks that matter, the simulation layer weakens and the moat thins toward compute alone. Compute alone is a commodity with a lead-time premium. That is the watch item. Track the ratio of real to synthetic training hours in the OEMs' own disclosures, although most of them are not disclosing it yet.
And the Moshi point deserves a steelman rather than a dodge. Yes, PersonaPlex is built on open lineage. That is the strategy working in both directions. NVIDIA absorbed the commons, tuned it for its own inference path, and licensed it open again with a persona-conditioning layer Kyutai did not ship. The commons got better. The board got a new reason to exist. A company that only sold closed models could not play this move, because it would be competing with the commons instead of feeding on it.
Three tests before you build on the stack
For an OEM or integrator deciding how deep to build on NVIDIA, three tests. They are meant to be run, not admired.
If your validation budget exceeds your hardware budget, you are already locked in. Stop describing Jetson as a swappable part in internal decks. Price the dependency as a line item, negotiate allocation like it is one, and put a number on what a forced migration would cost you in engineer-months so the board can see it before a supply shock does.
If your primary market is China, treat the compute layer as a policy variable. Keep a domestic board path alive in simulation even if you never ship it. That means at minimum an Ascend or Cambricon inference target for your policy, exercised quarterly, so that an export-control change is a schedule problem and not an existential one. The cost of keeping that path warm is small next to the cost of discovering you have none.
If your fleet data is growing faster than your synthetic data, the simulation moat is thinning for you first. You are the leading indicator for the scope condition above. Renegotiate what you pay for on the way down. The compute layer will still matter; the simulation and world-model layers will matter less, and your spend should follow that.
The pickaxe seller asks whether the miners keep showing up, and where. Who wins the race is somebody else's problem.