Unitree's Post-IPO Halving Exposes Physical AI's Missing Software

Analysts blamed the plunge on a specific gap: robots that move better every month but still cannot be trusted to do paid work without close supervision ^[4]. That gap is why Actuate, the conference for people building AI brains for robots, tripled in size since 2023 to roughly 1,500 attendees this year, according to organizer Foxglove ^[5]. It is also why a booth at that same event, run by infrastructure startup Avala, advertised a fix for what it flatly called the robotics data crisis ^[6], the shortage of high-quality training data that keeps end-to-end learning from producing commercially reliable products ^[7][8].

Antioch founder Harry Mellsop frames that gap as a stage: physical AI, he argues, is stuck in its GPT-2 era, the awkward stretch before OpenAI's language models became good enough to build a product around ^[9]. Getting past it will take more data and more compute, especially GPUs tuned for ray tracing, the kind used to generate the high-fidelity simulations that stand in for real-world training runs ^[10].

That is where autonomous-vehicle companies think they have an edge. Cars driven by people already generate labeled training data at scale, and driving is mostly about avoiding contact rather than manipulating objects ^[11]. Foxglove itself was started by former employees of Cruise, General Motors' former self-driving unit, and much of the tooling used across the sector traces back to AV companies ^[12]. Tesla is testing whether that data advantage transfers to Optimus, and Wayve and Uber have each opened humanoid robotics labs on the same bet ^[13][14][15]. Wayve's Alex Kendall told TechCrunch that the infrastructure, simulation, and ML-ops layers should carry over between vehicles and humanoids, even if the underlying world model needs separate training for each body ^[16], and that it is too early to lock in a hardware platform while sensors keep improving ^[17].

Genesis AI's Theophile Gervet, who raised $105 million this year to build a vertically integrated humanoid company, takes the opposite view: hardware and software need to be designed together now, not decoupled ^[18]. He also names the trap. A narrow vertical gets a robot into revenue and real-world data faster than a general-purpose one, since no customer wants a robot with an 80 percent success rate ^[21]. But a vertical built on today's immature models risks getting overtaken the moment a rival ships the equivalent of GPT-4 ^[21]. Gritt, Agility, and Bedrock are the current answer to that trap, narrow enough to already be working in solar construction, industrial sites, and excavation respectively ^[19], while general-purpose humanoids are still confined to labs ^[20]. Unitree makes exactly that kind of general-purpose hardware, which is why the market's reaction to its IPO reads less like panic and more like repricing ^[2][4].