Welcome to the first issue of On the Loop, an English-language newsletter on physical AI and robotics, focused on the Shenzhen-Hong Kong corridor.

As an expat here, 15 hours ahead of my roots in San Francisco and Silicon Valley, it's energizing to see the builder spirit alive and well on this side of the Pacific. My first startup was Awair, an IoT air-quality company (Internet of Things, for those who remember). A decade later, IoT has morphed into physical AI and robotics.

This won't be a news aggregator. In each edition, I'll give you my read on what I see on the ground, you tell me where I'm wrong, and we'll debate from there. Please push back—otherwise it'll just be me speaking into the void.

On September 8, I walked the factory floor at Lumos Robotics in Shenzhen and visited Robotuo, a working showcase in Nanshan's "Robot Valley"—a 10-kilometer stretch between two mountains that Xinhua says houses over 200 robotics firms. Robotuo connects global teams to China's supply chain, with 50-plus companies in its showroom and over 500 in its network.

Others have already covered how fast China's humanoid supply chain moves—including Tech Buzz China's field report earlier this year—and the macro numbers back it up: Shenzhen's robotics output topped 240 billion yuan (about US$35 billion) in 2025, up over 20% in a year. One founder I met in Nanshan is moving back from San Francisco because there’s less friction to build in China. As one local university president told Xinhua, a paper published in the morning can be tested on a neighboring factory line by the afternoon.

What struck me at Lumos and Robotuo wasn't just that you could buy a robot. It was how you could buy one: as a finished humanoid, as a fleet, or completely à la carte—gears priced by torque rating, actuators, pincers, cameras, and sensors sold separately off the shelf. It feels like being a Lego builder inside a Lego factory.

If you were around China’s tech scene in the late 2000s, you’ve seen this playbook before. While Nokia and Motorola built phones with big in-house R&D teams, Taiwanese chipmaker MediaTek bundled a “turnkey” phone chip with the reference software that cut a phone’s time to market to as little as four months. Suddenly, a 10-person team in Shenzhen’s Huaqiangbei market could buy a shared board design (gongban), snap on a screen, battery, and casing from the parts bazaar, and ship a working shanzhai (山寨) phone in weeks. By 2008, an estimated 3,000 to 4,000 shanzhai businesses had sprung up, CNN reported.

Robotics is in its MediaTek moment right now. At the Nanshan showroom, dozens of startups looked nearly interchangeable because they pull from the exact same component bazaar. That bazaar is driving hardware prices through the floor: in early 2024, Goldman Sachs found humanoid manufacturing costs had dropped 40% in a year (to $30,000–$150,000), far faster than the 15–20% analysts expected. Today, Unitree's R1 humanoid lists from $4,900.

Where this analogy could break (the steelman): A fair counterargument is that robotics doesn't have a MediaTek chip yet. In smartphones, the chipset was the turnkey core; in robotics, the hardest part isn't the gear or actuator—it's the foundation model (the vision-language-action model, or VLA) and the real-world demonstration data needed to train it. Neither ships as a reference design off a Shenzhen shelf. If the eventual integrator is a pure model company licensing brains to every hardware maker, looking for a hardware winner is the wrong lens.

Even so, once physical components commoditize, hardware assembly alone isn't a moat. Apple beat the shanzhai phone bazaar by refusing it—building a closed, trusted ecosystem that captured high-value buyers (even in 2025, iOS took only a quarter of app downloads but roughly 70% of consumer spending, the exact reason we launched the Awair app on iOS first).

In physical AI, what does a closed garden actually look like? It isn't an App Store. The garden is the depot—how easily you can service, calibrate, and replace parts when hardware inevitably breaks.

China is already testing multi-brand versions of this. Shenzhen's Longgang district opened a robot "6S" store in July 2025—taking the car dealership's "4S" model (sales, spare parts, service, and surveys) and adding leasing and customization. Beijing's Robot Mall followed in August with a dedicated floor for diagnostics and part swaps. Both carry dozens of brands, making them the Home Depot of robotics. Nobody has built the Apple Store or Tesla Service Center of robotics yet: one brand, its own parts, its own repair counter.

Commercial fleets show where the recurring margin lives. Waymo relies on partners like Moove to run its robotaxis in Phoenix and Miami, with London next; in August, Moove raised a $250 million Series C at a $2.1 billion valuation to build "Nests"—depots where autonomous fleets are charged, cleaned, calibrated, and repaired. Shenzhen's bazaar makers sell parts to everyone and never own the customer long enough to service them.

Could the closed integrator come from DJI, which launched its second-generation robot vacuum globally in September, or Insta360? Or OpenAI, whose first hardware device slipped to 2027, or Apple itself, whose tabletop robot is pegged for 2028? Whoever wins won't win by assembling cheaper actuators—they'll win by owning the depot alongside the robot.

The thesis: The breakout robotics brand in Asia won't win by selling standalone hardware through multi-brand "6S" stores—it will break out of the parts bazaar by building a single-brand service and depot network (its own hardware, its own parts, its own repair counter).

Tesla starts making robots in its car plant (TechCrunch). It already makes the robot and the AI and runs its own service centers: the Apple-plus-depot model, if it carries over.

China’s humanoid rental prices fell over 70% in a year (Seoul Economic Daily, machine-translated). Experts there say insurance and standardized service, not low prices, will pick the winners, which is the depot thesis.

Car lidar maker Hesai walks into the robot bazaar (HK01, machine-translated). When a car supplier with its own chips and factories sells robot parts, those parts get cheaper for everyone.

Tell me I'm wrong: The weakest link here is the commodity claim. If you've swapped one maker's actuator into another maker's robot and it just worked (or broke everything), I want to hear about it.

Until next time,
Kevin L