When Unitree's Wang Xingxing told a Beijing conference earlier on Thursday that humanoid robots are nearing their "ChatGPT moment," he was reaching for the most seductive analogy in technology.
But it could be the wrong analogy. The comparison implies that robot brains will improve the way large language models did, by ingesting ever more data until a capability threshold is crossed.
In short, the two problems are not alike, and the difference means there's a bottleneck with the new tech.
ChatGPT trained on roughly 13 trillion tokens of text scraped from the internet, while the entire world's robot manipulation data amounts to about 100,000 hours.
Put simply, physical interaction data cannot be harvested from the web, because a machine must actually pick up the wrench, drop it, and try again to record the attempt.
That data scarcity, alongside the "sim-to-real" gap, remains constrained (acutely).
This gap is apparent in testing: Stanford research reported by Fortune in May found that robots scoring nearly 90% in controlled simulations succeeded at just 12% of real-world tasks.
Wang's own target, a robot completing 80% of household jobs from voice commands, sits far above what current hardware delivers outside a studio.
But the reality doesn't gel with the claim. For example, on a factory line, gripper failures produce a 5% to 15% drop rate on objects outside a robot's training data, meaning dozens of failed grasps every hour.
And of course, hardware imposes its own ceiling. No commercially available humanoid can finish an eight-hour shift on one charge, with leading models managing two to four hours.
Wang, to his credit, was more sober than the frenzy around Unitree's sixfold debut suggested, conceding that a major software leap could take five to 10 years.
Investors betting on a near-term inflexion should note where the serious money is already going. Around $6 billion flowed into world model companies in the first quarter of 2026, because whoever controls the simulation infrastructure controls the ability to scale training without scaling robot fleets.
The "ChatGPT moment" may still arrive.
It will not arrive the way ChatGPT did.
