Physical AI just got a sharp reality check. Unitree, China's leading robot maker, listed on China's NASDAQ equivalent at a $66 billion valuation, then lost nearly half that value this week as analysts flagged the same gap builders keep running into: the hardware is improving faster than the models that tell it what to do. The bottom-line problem is that humanoids still cannot reliably do value-creating work.
The gap was on full display at last week's Actuate conference, the developer gathering run by data-tooling company Foxglove. Attendance hit 1,500, triple the size of the inaugural 2023 event, and the exhibit floor doubled as a diagnosis of the field's central bottleneck. A booth for infrastructure startup Avala advertised a fix for what it called the robotics data crisis — the shortage of high-quality training data needed to push end-to-end learning past demo-quality performance.
Antioch founder Harry Mellsop framed the state of play bluntly, telling attendees that physical AI is in its GPT-2 era — the OpenAI model that preceded ChatGPT and hinted at the shape of what was coming without being useful on its own. Getting to the next rung will take more compute, more diverse data, and GPUs tuned for the ray tracing that high-fidelity simulation demands. Autonomous vehicles are furthest along, in part because human-driven cars generate directly relevant data and in part because the core task is avoiding contact rather than manipulating objects.
Key facts
- 01Unitree debuted on China's NASDAQ equivalent at a $66 billion valuation, then shed nearly 50% of its value this week.
- 02The Actuate conference drew 1,500 attendees last week, tripling in size since it launched in 2023.
- 03Genesis AI, a vertically-integrated humanoid company, raised a $105 million seed round this year.
- 04Wayve CEO Alex Kendall pegs the consumer inflection at eyes-off autonomy for under $1,000 of in-car hardware.
- 05Foxglove shipped a new data-search product this week built on Nvidia's Cosmos open-weight world model.
That head start is now bleeding into humanoid work. Foxglove itself was founded by former Cruise engineers, spun out of General Motors' self-driving effort. Tesla is pushing Optimus, and both Wayve and Uber have opened humanoid-focused robotics labs. Wayve CEO Alex Kendall argues the tooling transfers cleanly even if the models don't.
Kendall's view is that data infrastructure, simulation stacks, and ML ops pipelines will be largely shared across vehicles and manipulation robots, with divergence coming at post-training for specific embodiments. He also warns against committing to any one hardware platform too early, since sensors and components are still improving quickly. A truly general model, in his framing, should be embodiment-agnostic.
Genesis AI CEO Théophile Gervet takes the opposite bet. His vertically-integrated humanoid company raised a $105 million seed round this year on the thesis that it is too early in the wave for a pure brain strategy — that hardware and AI need to be co-designed. Gervet also laid out the sector's uncomfortable trade-off between generalists and specialists.
“No customer cares about the general purpose robot that works at 80% success rate”— Théophile Gervet, Genesis AI CEO
Task-specific players are shipping. Gritt is building solar farms, Agility is deploying in industrial settings, and Bedrock is running excavators autonomously. General-purpose humanoids, meanwhile, remain lab curiosities. Bedrock CTO Kevin Peterson described excavation as a first probe into manipulation in the wild, with plans to extend the intelligence layer across a family of construction machines. The catch Gervet flagged: pick a vertical built on GPT-2-grade foundations today and a competitor building on GPT-4-grade foundations tomorrow will crush you.
Managing the resulting torrent of visual and lidar data is its own problem. Foxglove used Actuate to launch a new product built on top of Nvidia's Cosmos open-weight world model — actually, on Nvidia Cosmos — that lets engineers search training data with natural-language queries, speeding up evaluation and debugging cycles. The bet is that faster iteration, not a single breakthrough model, gets robotics over the hump.
“There will not be a ChatGPT moment for robotics”— Adrian Macneil, Foxglove CEO
That leaves the question Sam Altman recently teed up: when does physical AI have its ChatGPT moment, and what does it even look like? Kendall's answer is consumer-facing and hardware-priced — eyes-off autonomy in a car for less than $1,000 of hardware, which happens to be roughly the business Wayve is licensing models toward, and which he sizes as a multi-billion dollar opportunity. Gervet's answer is manipulation that works out of the box at 80%-plus reliability, so a person can tell a robot to close a laptop or clear a table and it just does it.
Foxglove CEO Adrian Macneil rejects the framing entirely. ChatGPT became a moment because ChatGPT went from zero to a million active users in a week, a distribution feat that has no analog in the physical world where you have to actually ship hardware into homes. What Macneil says he is waiting for is the Apple II or IBM PC moment for robotics — the first home robot people can buy that does something useful and fun.
The Unitree drawdown is the market pricing that gap in real time. A $66 billion valuation assumed the models were close to catching up to the chassis; the correction says investors are no longer sure. That is the tension the entire sector is now working through: hardware demos are impressive enough to raise nine-figure seed rounds and float multi-billion-dollar IPOs, but the software still cannot cross the reliability threshold that turns a robot into a product.
The takeaway for the AI market is that the physical AI trade is decoupling from the language-model trade. Frontier LLM progress does not automatically translate into robots that fold laundry, and the winners in this cycle will be the companies that either lock down a narrow vertical with real deployment data — Gritt, Agility, Bedrock — or control the tooling layer everyone else depends on, which is where Foxglove, Antioch, and Avala are positioning. Everyone building a general-purpose humanoid on today's model stack is running an expensive experiment with a public valuation attached, and Unitree just showed what happens when the market notices.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
