Physical AI has become one of the most heavily funded corners of venture capital, with companies raising billions to apply large language model techniques to robotics. That enthusiasm drove a landmark IPO for Unitree, China’s leading robot maker, which hit a $66 billion valuation on its first day of trading. But this week, the stock lost nearly half of its value, a sharp correction that analysts attribute to a simple problem: robots are getting better at moving, but they still lack the intelligence to do useful work.
The disconnect between hype and reality was on full display at last week’s Actuate conference in San Francisco, a gathering of developers building the ‘brains’ for robots. The event has tripled in size since its 2023 debut, drawing 1,500 attendees, according to organizer Foxglove, which builds data management tools for physical AI. Yet amid the optimism, a booth for Avala, another infrastructure player, advertised a solution to ‘the robotics data crisis’ — a sign of the industry’s central bottleneck: the shortage of high-quality training data.
Also read: Stability AI raises $76M from Universal Music, Sony Music, and EA to expand creative AI tools
Why robot brains are stuck in the GPT-2 era
Developers are trying to replicate the breakthroughs of frontier AI labs by finding more diverse datasets, experimenting with training regimes, and improving reinforcement learning scenarios. But as Harry Mellsop, co-founder of Antioch, a startup building simulation tools, put it, physical AI is in its ‘GPT-2 era’ — the pre-ChatGPT stage where models show promise but lack the polish for mainstream use. ‘More data and compute will be needed to get over the hump,’ he said, particularly GPUs optimized for ray tracing, which are essential for creating high-fidelity simulations.
The field’s most advanced segment is autonomous vehicles, which benefit from abundant real-world data and a simpler primary task: avoiding contact rather than manipulating objects. Much of the tooling for robot model-building has come from AV companies; Foxglove, for instance, was founded by former employees of Cruise, GM’s erstwhile self-driving unit. Now, carmakers and AV firms are betting their machine learning infrastructure will give them an edge in humanoid robotics. Tesla is already deploying its Optimus robot, while Wayve and Uber have launched robotics labs focused on humanoid form factors.
Also read: OpenAI's Jalapeño chip delivers major inference gains in first benchmark results
‘I think you need to start in vehicles…manipulation robotics is like self-driving five years ago,’ Wayve CEO Alex Kendall told TechCrunch. ‘The data infrastructure, the simulation, ML ops infrastructure, will probably be shared, but the specific world model for the simulator will be a different post-training.’ Kendall argues it’s too early to commit to any single hardware platform, as sensors and components are evolving quickly.
Not everyone agrees. Théophile Gervet, CEO of Genesis AI, a vertically integrated humanoid robotics company that raised a $105 million seed round this year, countered: ‘We’re too early in this wave for a brain strategy to work; there’s lots of opportunities to co-design hardware and AI.’
Vertical focus vs. general-purpose ambition
The debate extends to business strategy. Some companies are finding success by targeting specific tasks — Gritt is building solar farms, Agility is deploying robots in industrial settings, and Bedrock is operating excavators autonomously. Meanwhile, general-purpose humanoids remain largely confined to labs.
‘No customer cares about the general purpose robot that works at 80% success rate,’ Gervet said. ‘We see a lot of other players go general, but there is no value provided because there’s no vertical focus…but then, if you’re building [for a narrow] vertical on top of GPT-2, you’re going to get crushed by the company building on GPT-4.’
Focusing on a vertical offers immediate revenue and real-world deployment data, even if that data may not be diverse enough to advance general models. Bedrock CTO Kevin Peterson noted that his company started with excavation to understand ‘manipulation in the wild,’ with plans to build an intelligence layer spanning various construction machines.
Managing the sheer volume of visual and lidar data is a major challenge. This week, Foxglove announced a new product built on Nvidia’s Cosmos open-weight world model that lets engineers search robotics data using natural language queries, speeding up evaluation and debugging. The goal is to help model builders iterate faster.
What would a ‘ChatGPT moment’ look like?
When asked what a breakthrough might look like, Kendall pointed out that the largest robot deployment in the world is still consumer vacuum bots. A true ‘ChatGPT moment’ would excite consumers, not just investors. ‘One example would be when you get eyes-off autonomy for less than $1000 [worth of hardware] in a car,’ he said. His company is licensing models to automakers to achieve that, a business he sees as a multi-billion dollar opportunity.
Gervet envisions a different milestone: ‘Manipulation that just works out of the box. You can talk to a robot in natural language and have it do any basic task…pushing, pulling, closing a laptop, cleaning up a table…and it works to some level of reliability, let’s say 80% plus out of the box — that’s roughly your ChatGPT experience.’
Adrian Macneil, Foxglove’s CEO, offers a more grounded perspective. ‘There will not be a ChatGPT moment for robotics,’ he told TechCrunch. ‘The thing that made ChatGPT a moment was distribution — they went from zero to a million users in a week. Distribution in the real world is way harder. I’d be very excited for the Apple II moment or the IBM PC moment — when can I buy a home robot that starts doing useful and fun stuff?’
For now, the industry is caught between investor enthusiasm and the hard reality of data and compute limitations. As the Unitree stock slide shows, the market is beginning to price in the gap between promise and performance. The next few years will likely determine which approach — vertical or general, hardware-first or brain-first — can deliver the first genuinely useful robots.
This article is for informational purposes only and does not constitute financial advice. The cryptocurrency and robotics markets are volatile and uncertain; always conduct your own research before making investment decisions.
