Two things happened in the same week at the end of August, and together they say more about the state of humanoid robotics than either does alone. Unitree, China’s dominant robot manufacturer, arrived on the country’s Nasdaq-equivalent exchange to a $66 billion valuation and then, within days, lost nearly half of it. Days later, Skild AI, a company that sells no robots at all, unveiled S1, a foundation model that it says learns entirely new tasks from a single video placed in its prompt.
The market has drawn its conclusion with unusual speed. Bodies, however impressive, are no longer the scarce asset. Brains are. And the price of that conclusion, roughly $14 billion for Skild against a halved Unitree, is now the central fact of the industry’s economics.
What S1 Actually Claims to Do
Skild AI was founded in 2023 and has raised nearly $1.7 billion to build what it calls a general-purpose robot brain. Its homepage states the thesis flatly: “Physical AI should be omni-bodied,” a unified brain “to control any robot for any task,” from quadrupeds to humanoids to static arms.
S1 is the company’s flagship model, and its central claim is a methodological jump rather than a capability jump. Previous robot policies, when facing a new task, had to be post-trained: fine-tuned, retrained, or otherwise reprocessed for each new behavior. S1 instead uses in-context learning, the mechanism that made ChatGPT feel magical in 2022. You put a video of a human performing a task into the prompt, and the model executes it on the robot.
“You just add a video of a human doing something in the prompt, also called the context of the model, and it can just follow it on the robot,” Deepak Pathak, Skild’s co-founder and CEO, told The Robot Report. “The tasks we are showing are extremely complex and long horizon. They are not three-second, four-second tasks, not those tiny, simple tasks.”
The demonstrated tasks run up to ten minutes: repotting a plant, making coffee, cooking pancakes. If the claims hold, the economic significance is straightforward. Post-training is the tax that has kept robot deployment narrow, and in-context learning, if it transfers outside curated demos, abolishes it. A robot that learns from a YouTube clip is a robot whose marginal cost of new capability approaches zero.
Skild is also making a claim about embodiment. The company says the same model runs across quadrupeds, humanoids, and stationary arms, and that an earlier version demonstrated on-the-fly adaptation when a humanoid’s limbs were damaged. The company admits it has not yet scaled the model to humanoids specifically, which is a notable gap for a firm whose valuation rests partly on the humanoid boom.
The Data Problem Has No Golden Path
The most intellectually honest part of Skild’s announcement is its account of training data. Pathak divides robot training data into four types, and his taxonomy is worth quoting because it frames the entire sector’s constraint.
Teleoperation data, humans directly controlling robots, is high quality because it comes from the robot itself, but it is slow to capture and narrow. Human video is abundant and diverse, but far from a robot’s embodiment. Simulation is infinitely scalable, but the sim-to-real gap never fully closes. Data-capture gloves, humans wearing instrumented gloves while performing tasks, sit between teleoperation and video on both axes.
“If you look at any company out there, they are primarily focusing on one of these sources,” Pathak said. “But if we think from first principles, we realize there is no golden path. We have to use all of them because the pros of one source compensates for the downside of another source.”
This is the “robotics data crisis” that a booth sign at the Actuate conference in August promised to solve, an event that has tripled in size since 2023 to 1,500 attendees. Language models could be trained on the entire internet. Robot models cannot, because the internet contains no proprioception. Every brain-builder is forced to manufacture its own corpus, and the manufacturing method is the entire competitive moat. Skild’s answer is blending all four data types in pre-training. Generalist, the DeepMind-alumni firm, makes a similar bet; its Gen 1.5 model learns from video demonstrations as short as 3 to 12 seconds.
Emergence, Carefully Audited
The most interesting anecdote in the S1 rollout is a pancake. When the model flipped one, the Skild team was paranoid enough to audit its own training corpus.
“We went back and checked all of our millions of hours of data to look for any flipping example of any kind, and there was none,” Pathak said. “This flipping basically emerged by looking at how the spatula is moving and the robot could just do it.”
Emergence, a capability appearing in the model that was never explicitly present in the data, is the signature phenomenon of scaling in language models. Claiming it in manipulation is claiming that robotics sits on the same scaling curve as GPT. One audited anecdote is not a scaling law, and the industry’s demo-to-deployment gap is well documented. But the anecdote matters because it is falsifiable: Skild named the check it ran. Most competitors’ emergence claims do not come with an audit trail.
The Market Inverted
While Skild was announcing brains, the body market was repricing. Unitree’s crash, as TechCrunch reported from Actuate, had a one-line diagnosis from analysts: “While the robots’ physical capabilities are improving, they still lack the know-how to actually do value-creating work.”
The contrast with the brain-makers’ valuations is stark. Skild, valued at roughly $14 billion and SoftBank-backed, sells software and has just begun talking about production deployments. Physical Intelligence is reportedly valued at $11 billion. Generalist closed a $200 million extension led by 8VC in August on top of a $400 million Series B led by Radical Ventures, lifting it to a $3 billion valuation and a $600 million total round. Genesis AI, barely past a $105 million seed, was in talks to raise at $3 billion. The hardware champion of China halved; the software firms with no shipping products tripled.
Consolidation is running the same direction. Mobileye acquired Mentee Robotics for $900 million in January. Amazon bought Fauna Robotics in March. Meta acquired Assured Robot Intelligence. Automakers and AV firms are opening humanoid labs, with Wayve and Uber both launching robotics R&D efforts. When incumbent giants buy, they are buying brain teams, not actuator supply chains.
Three Strategies, One Unresolved Question
Underneath the valuations is a genuine strategic disagreement about sequencing, articulated most clearly at Actuate.
The brain-first camp, Skild, Physical Intelligence, Generalist, and arguably Wayve, builds the general model and lets any hardware inherit it. Alex Kendall, Wayve’s CEO, argues it is too early to commit to any hardware platform: “Manipulation robotics is like self-driving five years ago. The data infrastructure, the simulation, ML ops infrastructure, will probably be shared… but there’s going to be some differences for different embodiments.”
The co-design camp argues the opposite. “We’re too early in this wave for a brain strategy to work,” Théophile Gervet, president and co-founder of Genesis AI, told TechCrunch. “Our take is there’s lots of opportunities to co-design hardware and AI.”
The vertical camp is the only one already billing customers. Gritt builds solar farms, Agility deploys in industrial settings, Bedrock runs autonomous excavators. Gervet’s critique cuts at both other camps: “No customer cares about the general-purpose robot that works at 80% success rate,” while “if you’re building [a narrow] vertical on top of GPT-2, you’re going to get crushed by the company building on GPT-4.”
That “GPT-2 era” framing, from Antioch co-founder Harry Mellsop, is the honest descriptor of the moment. The tooling is maturing: Foxglove, founded by Cruise alumni, shipped a product in August that lets engineers search robot data with natural-language queries, built on Nvidia’s Cosmos world model. But the field is still waiting for its scaling-law moment, the point where more data and compute reliably buy more generality.
What Would a ChatGPT Moment Actually Be?
Asked directly, the practitioners do not agree, and their disagreements are revealing.
Pathak is the most conservative about his own release: “Is it completely ready to be rolled out to people’s homes? Not quite. But this is the first sign of what we believe might come.” Gervet defines it as “manipulation that just works out of the box… say 80% plus out of the box.” Kendall defines it economically: eyes-off autonomy on sub-$1,000 of hardware.
The sharpest answer belongs to Adrian Macneil, Foxglove’s CEO: “There will not be a ChatGPT moment for robotics. The thing that made ChatGPT a moment in time was the distribution, they went from zero to like a million active users in a week. Distribution in the real world is way harder than that. I would be very excited for the Apple II moment in robotics or the IBM PC moment in robotics.”
Macneil is right about the mechanism even if the metaphor changes. ChatGPT rode an existing distribution channel, the browser, at zero marginal cost. Robots must be manufactured, shipped, installed, insured, and maintained. Whatever discontinuity arrives in capability will be slowed by the physics of deployment, which is precisely why Tesla’s Optimus timeline keeps slipping and why Unitree’s valuation could halve in a week on a capability disappointment.
What to Watch
Three concrete signals will test the brain-split thesis over the next quarter.
First, Skild’s production claims. “You will see in the coming weeks, we’ll show how S1 is already helping in production,” Pathak said. The company’s acquisition of Fetch Robotics assets from Zebra was a deployment play, buying field-service talent and an installed customer base. If S1 runs real customer work, in-context learning exits the demo category. If the weeks pass with party tricks, the $14 billion premium starts to look like the top of the brain bubble.
Second, whether the model scales to humanoids. An omni-bodied brain that has not yet scaled to the form factor every customer is buying is a brain with an asterisk. Skild says humanoid performance is coming; the schedule is unspecified.
Third, whether Unitree stabilizes as a body-platform. If the world’s best-positioned hardware maker becomes the commodity arm supplier to Western brain-makers, the industry’s structure resolves into an Intel-versus-PC configuration, with all the margin concentration that implies.
The split is the story. Capital has decided that the scarce resource in robotics is no longer the actuator, the battery, or the balance controller. It is the general policy that makes any of those things useful. Skild’s S1 is the strongest claim yet that in-context learning is the bridge from the GPT-2 era to whatever comes next. The claim now has a countdown clock attached to it, and it is measured in weeks.
Sources: The Robot Report, TechCrunch, TechCrunch, Skild AI, The Robot Report