The benchmark gap is small enough to be noise, but the signal for operators is which foundation models will be available to non-US robot makers. RoboArena was co-built by Nvidia, Stanford, and Berkeley, and a Chinese lab now sits at the top of Nvidia's own scoreboard. That gives Unitree, Xpeng, Agibot, and the rest of the Hangzhou-Shenzhen humanoid cluster a credible domestic policy model to slot in next to (or instead of) Cosmos, which matters as export-control pressure on AI chips bleeds into model and toolchain access.
The more interesting tell is Nvidia's same-week push to partner with Unitree and Sharpa. Jensen is hedging: Cosmos needs Chinese embodiments and Chinese hand hardware to stay relevant as a generalist policy layer, even while a Chinese competitor is beating Cosmos on its own benchmark. For anyone underwriting humanoid software stacks, assume the embodied foundation model layer fragments along the same lines as the LLM layer, with Spirit, RoboBrain, and whatever Huawei ships becoming the default inside China.