Robot bodies are getting good. Robot brains are not. That gap just wiped out tens of billions in market value, and according to TechCrunch AI, it’s the defining tension in physical AI right now. China’s leading robot maker, Unitree, IPO’d at a $66 billion valuation on China’s Nasdaq equivalent, then lost nearly half its value in a week. Analysts pointed to one problem: the machines can move, but they still can’t reliably do work anyone will pay for.
That’s the story underneath the hype. Billions are pouring into companies trying to apply the large-language-model playbook to robotics, and the energy is real. TechCrunch AI reports that the Actuate conference, a gathering for developers building AI ‘brains’ for robots, has tripled in size since 2023 and drew 1,500 attendees this year. But one booth summed up the anxiety in the room. Avala, an infrastructure player, promised to solve ‘the robotics data crisis,’ the shortage of high-quality training data that keeps these models from getting smart.
The GPT-2 problem
Harry Mellsop, founder of the simulation startup Antioch, gave the sector its sharpest label: physical AI is in its ‘GPT-2 era.’ That’s the OpenAI model from before ChatGPT, capable but not yet useful to normal people. What stands out here is how honest that framing is. It means the breakthrough isn’t a tweak away. It needs far more data, far more compute, and specifically GPUs tuned for ray tracing to build the high-fidelity simulations these models learn from.
Right now, autonomous vehicles are furthest ahead. Two reasons: cars driven by people generate mountains of relevant data, and the core task is avoiding contact rather than manipulating objects. A lot of the tooling flows from that world. Foxglove, which helps model builders manage their data, was founded by ex-Cruise engineers. Now those car companies want in on humanoids. Tesla has Optimus, and both Wayve and Uber have spun up robotics labs.
Two camps, one big disagreement
The experts TechCrunch AI spoke with split cleanly on strategy.
- Build the brain first. Wayve CEO Alex Kendall argues you should start in vehicles, because ‘manipulation robotics is like self-driving five years ago.’ His bet: the data, simulation, and ML ops infrastructure will be shared across robot types, so it’s too early to lock into any one piece of hardware.
- Co-design brain and body. Genesis AI CEO Théophile Gervet, whose humanoid company raised a $105 million seed this year, disagrees flatly. ‘We’re too early in this wave for a brain strategy to work,’ he said, arguing there’s real value in designing hardware and AI together.
Gervet also named the trap that’s catching general-purpose players. ‘No customer cares about the general-purpose robot that works at 80% success rate,’ he said. His warning to founders chasing a narrow vertical too early is blunt: build on GPT-2 and ‘you’re going to get crushed by the company building on GPT-4.’
Why it matters now
The companies actually shipping robots are the focused ones. Gritt builds solar farms. Agility runs robots in industrial settings. Bedrock operates excavators autonomously. General-purpose humanoids, meanwhile, mostly stay in the lab. Task-specific work brings in revenue and, just as valuable, real deployment data, even if that data isn’t diverse enough to push general models forward.
Here’s the practical read for anyone building or investing in this space:
- Pick a vertical that pays. Real-world deployment funds the research and generates data you can’t buy.
- Own your data tooling. Foxglove just launched a product built on Nvidia’s Cosmos world model that lets engineers search dense visual and lidar data with plain-language queries. Faster iteration is the moat.
- Don’t marry your hardware. Sensors are improving fast. Flexibility is worth more than a bet on today’s best body.
Sam Altman recently said physical AI’s ChatGPT moment is only a few years out. Kendall offers a sharper test for it: the largest robot deployment on Earth is still the vacuum bot, so the real moment arrives when consumers get excited, not investors. His example is ‘eyes-off autonomy for less than $1,000 of hardware in a car.’ Watch that number. When it lands, the GPT-2 era ends. You can find the full report at the original source.