Unitree's 73-day IPO approval and LingBot-Vision going Apache-2.0 in the same fortnight: two numbers I can't square

I spent the last couple weeks tracking the China robotics pipeline and two events in basically the same fortnight have been bouncing off each other in my head. On July 3, Unitree got final CSRC registration for its STAR Market IPO at roughly $618 million, filed March 20, cleared in a record 73 days. That clears it to become China's first listed pure-play humanoid company. The hardware side has been accelerating hard. AgiBot hit its 10,000th mass-produced robot on March 30, going from 5,000 to 10,000 in about three months per Xinhua. Stanford's 2026 AI Index, citing IFR, has China installing nearly nine times as many industrial robots as the US in 2024. Unitree shipped over 5,500 humanoids last year with humanoids already more than half its revenue through the first nine months of 2025, per Caixin. Meanwhile Tesla Optimus fell far short of its \~5,000-unit 2025 target, with only hundreds built per The Information, and has walked back firm timelines around the V3 redesign.

Then around July 6, Robbyant (Ant Group's robotics company) dropped LingBot-Vision on Hugging Face under Apache-2.0: four backbone sizes, 21M, 86M, 0.3B and 1.1B parameters, all open weight. The company self-reports that its 0.3B model matches Meta's 7B DINOv3 on NYU depth estimation while training on less than a third of the data. The downstream product that actually sees glass and mirrors, LingBot-Depth 2.0, stays closed. But the foundation layer itself is now free, which reads to me like a straight commoditization of the perception stack.

So here's what I keep circling back to. The classic bull case for Western robotics, the thing that justifies the software multiples, is that perception and control software is the moat. The metal was never what you were paying for. The stack was. But LingBot-Vision is now free. And on the hardware side, the leading pure-play maker is about to hit public markets with actual unit economics while the best known Western name in the space can't hit production targets.

To be clear, none of this means China has won the category. The US still has more venture dollars flowing into frontier hardware, deeper systems integration talent, and NVIDIA still sits under most of the simulation tooling this field trains on. China has not solved general manipulation or reliable bipedal locomotion at scale either, and a 73-day approval sprint tells you nothing about how the stock trades once it actually lists. The geopolitical overlay could cut the other way entirely.

But if I'm trying to model where Western robotics margin actually lives five years out, I keep trying to square those two facts and I can't. If the perception backbone is free and the hardware is commoditizing, where does the investable moat live?