The reference humanoid is the part to dwell on. Nvidia putting a Unitree H2 plus Sharpa Wave hands plus Jetson Thor into a single open BOM, and routing it through Ai2, ETH, Stanford, and UCSD, sets a default research platform that any US-aligned humanoid startup will be benchmarked against. The quiet implication: the Chinese hardware stack (Unitree chassis, Chinese-sourced actuators) is now the reference body for Western humanoid R&D, while Nvidia owns the compute and the GR00T foundation models on top. That is the supply-chain reality the funding decks have been dancing around.
On the agent tools, the numbers that matter for operators are the synthetic data deltas Nvidia surfaced from named customers: Pegatron cutting training and deployment 67%, Inventec cutting defect data collection 30%, Foxconn lifting first-pass yield 3 points, Delta improving detection 17%. Those are the case studies the Cosmos/Omniverse sales motion will be built on in 2026, and they are aimed squarely at the visual inspection and digital twin budgets that have historically gone to Cognex, Keyence, and Siemens. Li Auto, DeepRoute, and Afari generating 300,000 sim renders per day on NuRec is the AV-side version of the same flywheel. The lock-in shows up later, when your synthetic data pipeline only runs on Brev launchables and your robot only learns on Isaac Lab.