Physical AI is attracting major investor attention, but the sector’s core problem remains practical: robot bodies are advancing faster than the AI systems that can make them useful. TechCrunch reports that companies are raising billions to apply large language model-style methods to robotics, while developers at last week’s Actuate conference were focused on the harder question of how to train reliable robot “brains.” The market’s patience is already being tested. According to TechCrunch, Unitree, described as China’s leading robot maker, reached a $66 billion valuation after its arrival on China’s equivalent of Nasdaq. This week, however, the company lost nearly half its value. Analysts cited by TechCrunch pointed to a basic concern: the hardware may be improving, but the robots still lack the ability to perform value-creating work. Actuate showed both the momentum and the bottleneck. The conference, which gathers developers building AI systems for robots, has tripled in size since it began in 2023 and drew 1,500 attendees, according to organizer Foxglove. Foxglove helps physical-AI model builders manage and visualize data, placing it close to the infrastructure layer that many robotics companies now need. The problem, as TechCrunch frames it, is a robotics data crisis. A booth sign from Avala, another physical-AI infrastructure company, explicitly promised to solve that crisis. In practice, the issue is the lack of high-quality training data for models that need to operate in messy physical environments. TechCrunch reports that generalized robots capable of handling arbitrary tasks remain far off, while end-to-end learning for specific tasks has not yet produced reliable commercial products. That leaves robotics developers trying to borrow the operating model of frontier AI labs without yet having the same kind of data flywheel. The reported playbook is to find or generate more diverse datasets, test different training approaches, and improve reinforcement-learning scenarios. Harry Mellsop, a founder of Antioch, a startup building simulation tools for model builders, told TechCrunch that physical AI is in its “GPT-2 era,” referring to the pre-ChatGPT stage of OpenAI’s model progression. Mellsop’s implication is that the sector still needs more data and compute before it clears the next threshold. TechCrunch specifically notes demand for GPUs optimized for ray tracing, which are used to build high-fidelity simulations. That matters because physical-AI developers often need to train or test robots across many situations that are expensive, slow, or risky to reproduce in the real world. TechCrunch reports that autonomous vehicles are the furthest ahead. Cars can collect relevant driving data from human-operated vehicles, and the core task is often to avoid contact rather than manipulate objects. Much of the tooling used by robotics model builders has roots in autonomous-vehicle work; Foxglove, for example, was founded by former employees of Cruise, General Motors’ former self-driving effort. That overlap is now pulling autonomous-vehicle companies toward humanoid robotics. TechCrunch reports that Tesla is pursuing this path with Optimus, while Wayve and Uber have launched robotics labs focused on humanoid form factors as research-and-development efforts. Wayve CEO Alex Kendall told TechCrunch that “manipulation robotics is like self-driving five years ago,” and said data infrastructure, simulation, and machine-learning operations will likely be shared, even if different robot embodiments require different world models or post-training. The near-term read is not that humanoid robots are suddenly ready for broad deployment. It is that the industry is beginning to standardize around a familiar AI-scaling problem: data, simulation, compute, and training infrastructure. The companies that solve those layers may have leverage even if the final winning robot body is still unsettled. Who benefits: Infrastructure companies serving physical-AI teams, including data-management, visualization, simulation, and machine-learning operations vendors, stand to benefit if robotics developers follow the frontier-AI scaling playbook. Autonomous-vehicle teams may also have an advantage because some of their tooling and data practices transfer into robotics research. Who's exposed: Robot makers with high expectations but limited commercially reliable autonomy are exposed to valuation pressure. TechCrunch’s Unitree example shows how investors may reassess hardware-first robotics stories when the software capability gap remains visible.