Who owns the robot's brain

The tell arrived at GTC Taipei in early June 2026. NVIDIA announced what it called an open humanoid robot reference design: a division of labor in which NVIDIA supplies the brain — Jetson Thor onboard compute plus its open Isaac GR00T models — while Unitree builds and sells the body, the H2 Plus, fitted with Sharpa five-fingered hands, due to ship from Unitree in late 2026.

Read the roles carefully. The company with the deepest interest in selling robots chose to publish the reference design for the physical machine and keep the model-and-compute layer for itself. You can argue about motive — NVIDIA still sells the silicon the brain runs on, so this is not charity — but the shape of the deal is the clearest single statement of where the industry now believes the durable value sits. Not in the chassis.

GR00T is not one model but a line. NVIDIA released Isaac GR00T N1 in March 2025 as the first open, downloadable humanoid foundation model, then N1.6, then the current N1.7 — cross-embodiment vision-language-action models published open-weight on Hugging Face and GitHub. The early-access list read as a who's-who of other people's robots: 1X, Agility, Boston Dynamics, NEURA. None of them NVIDIA's own machine. That is what a hardware-agnostic software layer looks like.

Underneath sits a data strategy that is itself the asset. N1.7 is pretrained on roughly 20,854 hours of egocentric human video alongside robot demonstrations, and NVIDIA describes a scaling law for robot dexterity — more video hours, more completed tasks. Around it sit the Cosmos world models — Transfer 2.5, Predict 2.5, Reason 2 — released to generate physically based synthetic training data and to evaluate policies in simulation before a robot ever moves. The defensible thing here is the data pipeline and the model trained on it, not the arm.

The model layer has two poles. Physical Intelligence sits at the open, independent-lab end: its pi0 policy is built on a pretrained roughly three-billion-parameter vision-language backbone with a flow-matching action head, trained on internet-scale data plus the shared Open X-Embodiment set plus its own dexterous data from eight distinct robot platforms. Google DeepMind sits at the other end: Gemini Robotics-ER 1.6, out around April 2026, brings Gemini's reasoning and spatial understanding to robots but ships only as a closed, hosted API — a model code you call, with no weights to download.

There is a standards move in between. In late May 2026 NVIDIA said it would place its physical-AI families, GR00T and Cosmos among them, under the Linux Foundation's OpenMDW-1.1 license. That matters to anyone betting on a shared layer. But the honest caveat is that open is doing heavy lifting in all of this. GR00T weights ship under NVIDIA's own Open Model License, and OpenMDW-1.1 is a Linux Foundation license — neither is OSI-approved. This is open enough to build on, not open in the strict sense the word once carried.

Follow the capital. In January 2026 Skild AI raised about USD 1.4 billion at a valuation above USD 14 billion — up from roughly USD 4.5 billion only seven months earlier — in a round led by SoftBank with NVIDIA's venture arm, Bezos Expeditions and a line of strategics. Skild builds no robot. It sells Skild Brain, which it pitches as an omni-bodied foundation model able to control any robot without prior knowledge of its body.

Discount the superlatives — the company calls it the largest robotics-AI round on record and its model the industry's first of its kind, and both are the kind of claim a press release makes. Strip those away and the durable fact remains: this much money flowed to a pure-software robot brain, decoupled from any machine. It is the single strongest financial signal that the value in this field is drifting up, off the hardware and into the layer that decides what the hardware does.

Moduloa's bet is that the durable moat in physical work is the framework around the robots — standards, data, certification, training, routing — and not the machines themselves. The mid-2026 evidence for a value shift into the layer above the hardware is genuinely strong, and it is arriving faster than the deployments are. That is encouraging for the direction of the thesis.

Two honest qualifications keep it from being a victory lap. First, the layer being captured here is the model-and-compute layer, and NVIDIA still sells the compute — so value moving off hardware is directional, not absolute, and an open model layer plausibly exists to sell more chips. Second, and more important, the model layer is not the certification-and-routing layer Moduloa is aiming at. A cross-embodiment policy decides how a robot moves. It does not decide which certified, portable production capacity a customer should trust with a real order. Those are different moats. The model layer filling in first is useful groundwork, not a competitor for the same ground. This note scores nothing on the register; it is evidence about the shape of the industry, not a decisive test of any dated claim. See the register →

This is a synthesis of current public reporting and primary announcements, not a primary source itself. Key references:

NVIDIA — open humanoid robot reference design · NVIDIA — new physical-AI models (Cosmos, GR00T N1.6) · NVIDIA — Isaac GR00T N1 announcement · arXiv — GR00T N1 technical report · GitHub — NVIDIA Isaac-GR00T (N1.7) · Linux Foundation — OpenMDW-1.1 and NVIDIA adoption · Physical Intelligence — pi0 · Google — Gemini Robotics-ER API overview · Business Wire — Skild AI raises USD 1.4B · The Robot Report — Skild AI and the omni-bodied brain

Version numbers and figures in this field age within weeks — GR00T moved through four releases in about fifteen months, and several facts here date to May and June 2026. Independent, non-vendor market-share numbers for foundation-model adoption remain thin, and claims about Figure's and Tesla's in-house models could not be verified to this standard, so they are left out. Corrections are welcome: shk@moduloa.com.