Unitree’s bubble bursts and the hardware-intelligence gap is exposed

The IPO of Unitree on China’s STAR exchange lifted the valuation of the leading robot maker to $66 billion, but this week almost half of that value was erased. The reason is clear: the robots can move, but they still cannot perform work that creates real economic value.

At the Actuate conference held last week, an event for developers of “brains” for robots, 1,500 participants were recorded, three times the attendance at the 2023 launch, according to organizer Foxglove, which provides management and visualization tools for physical AI models. One booth displayed the slogan “solve the robotics data crisis,” defining the core problem: there is not enough high-quality data to train general models, and end-to-end learning for specific tasks has not yet yielded products with reliable commercial performance.

Harry Mellsop, founder of Antioch, which builds simulation tools for model developers, describes the situation as “the GPT-2 era,” the stage before the breakthrough of ChatGPT. He says far more data and compute are required, especially graphics processors adapted to ray tracing to generate high-resolution simulations that would replace costly, slow physical data collection.

The most advanced segment is autonomous vehicles, partly because relevant data can be collected from cars driven by humans, and because the primary task is avoidance of contact rather than manipulation of the environment. A large share of the infrastructure for model development originated there: Foxglove itself was founded by alumni of Cruise, General Motors’ autonomous-vehicle venture that was shut down.

Car manufacturers now bet that the capabilities they have built will let them compete with dedicated humanoid makers. Tesla is already doing this with Optimus, and alongside it Wayve launched, focusing on autonomous vehicles, while Over Robotics Labs is adapting its robotics labs to a humanoid format. Alex Kendall, CEO of Wayve, told TechCrunch that “manipulation robotics is where autonomous driving was five years ago.” He expects data infrastructure, simulation and MLOps to be shared, but the world model in the simulator will require different post-training for each embodiment. Kendall argues it is too early to commit to a single hardware platform, because sensors and components evolve rapidly, and a truly general model must be hardware-agnostic.

An opposite view is championed by Théophile Gervet, CEO of Genesis AI, which raised a $105 million Series D this year. He says, “we are too early in this wave for a brain-only strategy; our view is that there are many opportunities to co-design hardware and AI.” Gervet points to another practical difference: companies targeting specific tasks already deploy robots in the field—Gritt builds solar farms, Agility runs robots in industrial settings, and Bedrock operates autonomous excavators—whereas developers of general humanoids remain stuck in the lab.