At a glance
WHAT IT’S REALLY ABOUT
Humanoid robots hinge on manufacturing, data, and open-source commoditization ahead
- Kang argues humanoid robot intelligence may reach “good enough” for many tasks within 2–3 years, but real-world impact will be gated by slow, capital-intensive manufacturing and supply-chain scale-up.
- He explains why full-stack, vertically integrated companies (hardware + model training + deployments + manufacturing) are advantaged because embodiment-specific data and co-optimization between hardware and learning systems drive faster performance gains.
- He projects a multi-trillion-dollar humanoid market by pricing robots against annual human labor costs and highlighting demand expansion when labor becomes cheaper and more abundant.
- He predicts open-source models will compress the gap with frontier systems and eventually commoditize much of the “model layer,” shifting durable value toward deployments, hardware innovation, and production capacity.
- He rejects a simplistic “US vs China race” framing, expecting both ecosystems to mature largely domestically with shared benefit from open research, while noting China’s platform-first hardware distribution strategy as a notable differentiator.
IDEAS WORTH REMEMBERING
5 ideasHumanoid capability is becoming real, not just curated demos.
He cites Figure’s long-running livestream as evidence of robustness versus “best take” videos, even if humans still narrowly win; the bigger point is repeatable performance under fatigue and time pressure.
The real limiter is robot supply, not just smarter models.
Even if intelligence reaches useful levels quickly, robots can’t be “instantiated” like software; scaling requires factories, component supply chains, and lead times for hundreds to thousands of units.
Full-stack humanoid companies have a compounding advantage.
Building hardware, training systems, deployments, and manufacturing together enables co-optimization (sensing, torque control, simulation fidelity) and ensures access to large fleets for proprietary data collection.
Embodiment-specific data is a moat—and it requires fleets.
Models learn best from data collected in the same body they’ll operate in; without many identical robots (including teleop data collection), progress slows and generalization suffers.
Open source is closing fast and will commoditize much intelligence.
He claims the open-source gap has shrunk from ~2 years to ~6 months and will become “good enough” for many physical tasks, making differentiated value migrate to deployment, hardware, and manufacturing execution.
WORDS WORTH SAVING
5 quotesThis is a technological revolution that is different because it turns physical labor into a, in a product that almost anybody can access.
— Andrew Kang
I actually think it's closer to something like two to three years where humanoid intelligence, robot intelligence gets good enough to do most of the tasks that we need on, on a daily basis.
— Andrew Kang
I can spin up a million instances of a chatbot instantly, but I can't do that for robots. I can't produce them out of thin air.
— Andrew Kang
Because if I'm doing a simple task, like for example, restocking shelves or, uh, you know, assembling a computer mouse, I don't need a really high-level intelligence to do that. I don't need an Einstein to be able to do these tasks.
— Andrew Kang
It's pretty underappreciated how hard building a robotics company actually is.
— Andrew Kang
High quality AI-generated summary created from speaker-labeled transcript.
