The World Artificial Intelligence Conference opened in Shanghai this week, and the humanoid robots were everywhere: more than 30 of them from AGIBOT alone wandered the venue guiding visitors, while the company unveiled four new products and a factory line where its G2 robot already works a live quality-inspection shift. Strip away the choreography, though, and the more interesting story is not any single machine. It is that three of the world's most advanced economies are placing three incompatible bets on what a humanoid robot is for, and this year you can see all of them at once.
China's bet is on volume. As the analysts at TrendForce describe it, the country is running the humanoid playbook it wrote for electric vehicles: localise the supply chain, drive down the cost of motors, reducers and batteries, then flood the market to reach economies of scale. AGIBOT scaled from 1,000 units to 10,000 in barely more than a year. The wager is that shipping thousands of imperfect robots into real factories and warehouses generates a torrent of real-world data, and that the data eventually closes whatever intelligence gap exists today. Deploy, collect, update, redeploy.
America's bet is on the brain. US firms have largely stopped competing on hardware specifications and moved up the stack to intelligence: NVIDIA's physical-AI platforms, Google DeepMind's Gemini Robotics, OpenAI's embodied research, and the vision-language-action models that let a robot generalise across tasks rather than repeat one. The idea is that as robots move into messy environments, the durable advantage is understanding, not actuators. It is a plausible bet with an awkward flip side, which is that Chinese firms are now accumulating exactly the real-world data those models are hungry for.
Japan's bet is the strangest and, on a longer view, maybe the shrewdest. The country that built the first full-scale humanoid in 1973 brought almost nothing to its own robotics expo this year; its one human-shaped entry spent the show sinking basketball shots. That is not decline so much as a deliberate exit. Honda wound down Asimo back in 2018, and Japan redirected its robotics muscle toward jobs nobody livestreams: a 22-metre arm extracting melted fuel from inside Fukushima, robotic labs running drug experiments overnight, arms tested on the space station. Its national plan, via a new consortium called Noetra backed by SoftBank, Sony and Honda, is not to out-manufacture China but to build a foundation model that can run on any robot body at all.
A useful splash of cold water comes from Jeff Burnstein, who has led America's automation trade group for nearly two decades. Most humanoids in factories today are pilots, he notes, landing at 20% to 50% effectiveness when customers need 99-point-something. The real gate on adoption is not intelligence or dexterity but safety, and there is still no humanoid safety standard for machines working alongside people. So the three bets are being placed on a market that, for all the billions and all the dancing robots, has barely started to prove itself. China is wagering that scale buys intelligence, America that intelligence beats scale, and Japan that the smart move is to skip the parade and quietly build the part everyone else will eventually need. They cannot all be right, and the expo floor is not where we will find out.