by Tan Aik Keong (AK)
Once AI starts growing arms and legs, the competition stops being just about parameters and compute — it becomes about efficiency, cost and industrial dominance in the real world. What's often called physical AI, or embodied intelligence, is putting a model into arms, legs, wheels and sensors, so a machine can see, judge and move reliably in a real environment. When AI on a screen says something wrong, at worst it's embarrassing. When a robot in the real world takes a wrong step or grabs the wrong thing, it can mean a production-line stoppage, damaged equipment, or worse, injury. So the real bar for physical AI was never about whether it can talk — it's whether it can act, reliably, over the long run, and fail safely, controllably and accountably when it doesn't.
That's also why the most interesting storyline right now is how the US-China race is splitting into two distinct playbooks. The US looks more like it's building the brain — betting on foundation models, simulation and platforms. China looks more like it's building the body — betting on supply chain, manufacturing efficiency and the price curve. How these two approaches collide will decide whether humanoid robots end up as expensive showpieces, or something as replicable, scalable and cost-justifiable as industrial equipment.
China's edge: cost brings scale
The Chinese camp's killer move is how fast falling cost is driving adoption. Unitree's 2025 R1 brought the starting price of a bipedal humanoid down to RMB 39,900 — well below its previous generation's price range. Price isn't a vanity metric here — it's the ticket to actual deployment: once a robot moves from a "research budget" line item to an "equipment budget" one, businesses start running real ROI numbers and scale from one or two pilots to dozens, then hundreds. At the same time, China's complete component supply chain and manufacturing base means cost reduction isn't limited to the finished robot — it flows down into motors, gearboxes, actuators, structural parts, wiring and assembly, driving a faster iteration cycle.
UBTECH represents a different, more shop-floor-oriented approach, emphasising delivery and usability. Its Walker S2 targets scaled deployment in settings like factories, and the company has disclosed progress on mass production, batch deliveries and order volume. For the industry, that kind of information matters more than any backflip video, because it signals that quality control, after-sales support, spare parts, on-site integration and staff training are all being tested for real. The more grounded point is that real value usually doesn't come from "as dexterous as a human" — it comes from "reliable enough at a limited set of tasks": moving things, sorting, inspection, simple assembly, loading/unloading, following a process alongside people — turning repetitive labour into stable output.
China's approach carries a familiar risk too: homogenisation and a bubble. A lot of companies are crowding into similar hardware designs and use cases, and in the short term can grow on price and funding alone; long-term, they'll need to differentiate on reliability, energy consumption, maintenance cost, software capability and depth in a specific use case. A shakeout is inevitable, and the survivors will likely be the handful of players who can execute hardware engineering, supply-chain management and a real data feedback loop together, properly.
The US edge: the brain and the platform
The US approach is concentrated more on the brain and the platform. NVIDIA's 2025 release of Isaac GR00T N1 sent a clear signal: humanoid robotics is replaying the large-model era's playbook — use a more general-purpose model as the foundation, then transfer that capability to specific tasks and specific machines through simulation, synthetic data, and a smaller amount of real-world data. The US is good at turning a toolchain into an ecosystem: training frameworks, simulation environments, data generation, deployment acceleration, developer interfaces — the goal isn't selling one robot, it's getting many robots running the same, increasingly capable brain.
Capital markets are also betting on the platform story for a "general-purpose robot model": if future robot form factors end up wildly varied, the most valuable asset might not be any particular piece of hardware — it might be the model and software stack that can transfer across form factors and generalise across use cases. At the hardware level, the US has its own headline names too: Figure's partnership with OpenAI brings generative AI capability directly into a humanoid robot; Tesla positions Optimus as general-purpose labour for dangerous, repetitive and tedious tasks, and has repeatedly signalled plans to gradually trial it in its own operations. The US narrative's key words are generality and scalability — making the robot feel more like a software platform you can install new skills onto.
Physical AI still isn't pure software
Physical AI, in the end, isn't pure software. The US approach's real-world constraint tends to come from supply chain and manufacturing: critical materials, critical components, production ramp-up, cost control, quality consistency — all of these can determine whether something can actually be delivered at volume. Especially for a mechanically and electrically complex product like a humanoid robot, even a mature model can still see scaling slowed if hardware cost and supply stability can't keep up. In other words, a fast brain still needs a body that can keep pace.
Looking at China and the US on the same map, the outcome won't be decided by whose robot looks more human — it'll be decided by who can get the full loop running faster: the brain (model and toolchain), the body (hardware and reliability), the data (continuous feedback and iteration), and manufacturing (scale and cost) — none of which can be skipped. China is more likely to push humanoid robots toward "buyable, affordable, deployable" in the near term; the US is more likely to build a stronger software standard and ecosystem lock-in over the medium-to-long term through foundation models and platforms.
The metrics that will actually decide the industry's shape are fairly unglamorous: how fast per-unit cost falls, continuous uptime, failure rates and safe-stop mechanisms, changeover time on deployment, maintenance and spare-parts systems, the data feedback and iteration cycle, and how stable the supply of critical components stays. Physical AI's real finish line isn't a launch event — it's the steady payoff across thousands upon thousands of repeated actions. When AI genuinely grows arms and legs, what changes isn't just factory efficiency — it's the cost structure and industrial map of the real global economy.
Part of the AK AI Corner column. Originally published in Oriental Daily (东方日报) on Dec 20, 2025.