The body is a commodity. The brain is the company.
Robot bodies are getting cheap fast, mostly in China. The value pools in the models, the training data and the narrow jobs where a robot can work today.
Neil Gaikwad
The size of the prize
What the consensus misses
The market is buying humanoid bodies. I think the body is the part that commoditizes. China installed more than half the world's industrial robots in 2024, a Chinese humanoid's parts already cost about $35,000 and are heading under $17,000, and Unitree sells a humanoid for about $5,900 and turns a profit. Competing on hardware cost against that supply chain is not a venture bet.
Where I see the value is the brain and its fuel. Foundation models for robots are improving fast, and the scarce input is real-world data: hours of robots and people doing actual tasks. A company that sells robot training data is reportedly near a $500M valuation. That tells you where the bottleneck is.
And the revenue arrives through narrow jobs before general ones. Teleoperated picking, laundry, inspection and hotel work run today with a person in the loop, earn money, and generate exactly the data that moves those jobs toward autonomy. The home robot is the last mile, not the first.
Why now
Robot models started generalizing. Physical Intelligence published a model that cleans kitchens and bedrooms in homes it has never seen, Google DeepMind launched Gemini Robotics, and NVIDIA released an open humanoid model.
Real deployments produced real lessons. Figure's robots handled 90,000+ parts across 30,000+ BMW X3s over eleven months, and BMW published what broke. Agility's Digit has worked 65,000+ hours across nine customer sites.
Public markets opened. Unitree listed in Shanghai at about 61 billion yuan, and Agility agreed to go public at a $2.5B pre-money value.
Where the value pools
| Layer | My call | Why |
|---|---|---|
| Robot bodies | Commoditizes | China leads on cost and volume, and prices are falling by half within five years. |
| Robot foundation models | Most value | The general skill layer that every body will need, if the leaders can turn models into reliable work. |
| Training data and simulation | Underpriced | The binding constraint today; data vendors are being valued like model companies. |
| Narrow deployments | Durable value | Teleop-first jobs earn now and own the task data for later. |
| Components: hands and touch | Most value | Dexterity and durability are where pilots fail. Good hands are scarce. |
Who is winning so far
| Company | What it does | Stage and last round |
|---|---|---|
| Physical Intelligence | General robot foundation models, partly open-sourced | $600M led by CapitalG at $5.6B Nov 2025 · news.bloomberglaw.com |
| Skild AI | A general robot brain that learns tasks | ~$1.4B Series C led by SoftBank at $14B+ Jan 2026 · news.bloomberglaw.com |
| Figure | Humanoid robots and their own factory | Series C, $1B+ at $39B Sep 2025 · figure.ai |
| Apptronik | Apollo humanoid; Gemini Robotics partner | $520M extension at about $5B Feb 2026 · unite.ai |
| Agility Robotics | Digit humanoid in logistics and manufacturing | Going public at a $2.5B pre-money value Jun 2026 · robotics247.com |
| Unitree Robotics | Low-cost quadrupeds and humanoids, profitable | Shanghai IPO at about 61B yuan Aug 2026 · caixinglobal.com |
| Generalist AI | Robot model trained on large-scale real-world data | $400M led by Radical Ventures Jun 2026 · robotics247.com |
| Dyna Robotics | Robots working in hotels, restaurants and laundromats | $120M Series A Sep 2025 · prnewswire.com |
| Genesis AI | Robot model trained in its own physics simulation | $105M seed co-led by Eclipse and Khosla Jul 2025 · tipranks.com |
Where I would write a seed check
Teleoperation and first-person human video at scale. The model labs say data is the bottleneck, and they pay for it.
Early examples: Vision Lab, XDOF, Mecka AIVideo-only learning misses force and touch, and pilots fail at the forearm. A good hand is a company.
Early examples: Tangent RoboticsPicking, sorting, laundry and inspection with a person in the loop earn today and generate the data for autonomy.
Early examples: Avatar RoboticsThe standard for robots that balance themselves is still a draft. Whoever can prove a robot is safe near people decides how fast every deployment happens.
How it plays out
- 2027
Home robots reach first customers, and the share of tasks that still need a remote operator becomes the number to watch.
- 2028
Industrial robot installs pass 700,000 a year, and the safety standard for walking robots is published, opening mixed human workplaces.
- 2031
My expectation: humanoids ship by the hundreds of thousands a year, mostly Chinese-built, running models from a handful of labs.
- 2036
My expectation: robots that learn a new warehouse task in a day are normal, and data rights matter more than hardware.
As of Oct 2026
Risks
- Demos are ahead of production: Tesla missed its 2025 Optimus forecast and would not give a 2026 target. electrek.co
- Hardware durability: in BMW's pilot the forearm was the top failure point and cabling snapped on 10-hour shifts. humanoidsdaily.com
- Safety: the standard for dynamically stable robots is still a draft, and experienced roboticists advise staying three meters from walking robots. rodneybrooks.com
- China's cost and volume lead: 54% of 2024 installs and 57% of its home market held by domestic suppliers. ifr.org
- Valuations far ahead of volume: Figure at $39B and Skild above $14B against about 18,000 humanoids shipped in 2025. caixinglobal.com
Sources
- IFR, World Robotics 2025
- Fortune, BofA humanoid forecast (Mar 2026)
- Humanoid Guide on Morgan Stanley's humanoid model
- Jiemian on IDC humanoid shipments (Jan 2026)
- The Manufacturing Institute, workforce outlook
- arXiv, pi0.5
- Google DeepMind, Gemini Robotics
- NVIDIA, Isaac GR00T N1
- Humanoids Daily, BMW Spartanburg pilot
- Caixin, Unitree IPO (Aug 2026)
- TechCrunch, Mecka AI and robot training data (Sep 2026)
- Rodney Brooks, Why humanoids won't learn dexterity
Nothing here is investment advice: these are my own notes on public information, written to sharpen judgment, not to recommend buying or selling anything.