Five years ago, a humanoid robot cost about as much as a house. Today you can order one that does kung fu for the price of a used car, from @UnitreeRobotics. @MorganStanley estimates there could be more than a billion humanoids by 2050. If current trends continue, most will be built in China. So what is actually going on? When @ReadFuturist suggested I write about China's humanoid boom for AI Supremacy, I ended up going much deeper than the robots themselves. The result is my attempt at a full account of Chinese physical AI: how decades of industrialization built the supply chain underneath it; the founders and companies leading the field; what the robots can actually do today; who is buying them; where the money is going; how the business models look; and what still has to happen before this becomes a real mass market. It gets into the funding boom and rush toward IPOs, Beijing's sudden effort to cool the hype, the race to make the machines genuinely useful, and the growing US-China tensions around robotics and supply chains. And along the way: robot fight leagues, training grounds, dexterous hands, world models, billion-dollar rounds, factory deployments and national security. In short, a deep dive into the industry as it exists right now, how it got here, and where it may be going next. The full story of how China became the center of the humanoid boom and came to be building what some call humanity's last machine. Full piece 👇🏻

I’m pretty sure the entire world could get behind a Nokia resurgence 🔥

Chinese investors have had a difficult time capturing the AI wave. Compare this to the S&P 500 and it's night and day. Even the robotics hype has been disappointing. Unitree is down >20% from open. Not surprising there's a huge demand to invest in U.S. stocks, but the capital controls and recent brokerage crackdown (Futu, Tiger etc) make direct access much harder.

We are in a golden age for consumer AI that economically cannot last. The subsidies on these coding plans are craaaaazy.

Claude 5.5 be like

The U.S. versus China guys often miss this key feature The difference in capital markets This is American exceptionalism 🇺🇸

I have seen neither Boardwalk Empire nor Dexter. Which one should I watch next?

Art 👇🏻

You know what, bodyguard isn’t actually a bad use case

Weilan Deep Sea (卫澜深海), a Xi'an startup building underwater maintenance robots, has completed 5 funding rounds in 9 months. Its core shareholders include Shenzhen Capital Group, Yuanhe Origin (元禾原点) and Guoce Capital (国策资本), with existing backer Innoangel investing again. 36Kr's report gives no amount for the latest round, labeled “angel++++”. The company pairs its two-armed robot with a seabed base designed to supply the robot with power and communications. The base generates that power from ocean currents with a screw-shaped turbine. The first jobs are inspecting subsea structures, cleaning off marine growth and checking the sacrificial zinc anodes that protect equipment from corrosion; the new money is aimed at oil and gas sites. Founder Zhang Dayu says the growth must be cleared regularly to keep structures safe, and the anodes checked to see whether they still work. For offshore wind farms and oil and gas pipelines, he puts maintenance at about 1% of their value a year over lives of 20 years or more. Today that work needs a ship on site. For divers working deeper than 120 meters, Weilan puts the cost of a specialist support vessel and dozens of staff at over ¥1M (about $150,000) a day. It says standard remotely operated robots are still controlled by cable from a large mother ship. Zhang argues a robot kept at the installation could cut that reliance on ships and crews. Zhang started the company out of a Xi'an Jiaotong University lab team. The university says the accelerator MiraclePlus backed it early. His doctoral research was on drawing power from slow currents; the technology was developed for equipment left underwater for long periods, including submarine-detection sensors. A turbine that cannot start in slow water produces nothing, and a university profile of Zhang says his design starts turning at 0.1 meters a second, less than a tenth of walking pace. In December 2025, Weilan said military demand was its starting point. Weilan says it has delivered its fourth-generation robot and completed autonomous inspections of subsea energy facilities. Its AI is meant to find targets by sonar and guide the arms up close. Training would move from simulation into real jobs under human supervision, with human corrections used to retrain the model. Weilan plans to sell maintenance services before selling robots. Those jobs would give it experience of what fails underwater and the records of completed jobs needed to improve its robots. Weilan's figures don't show how much power the base produces or how much energy a job uses, so it is unclear how often the robot can work. Long idle periods or repeated vessel trips to recover a failed robot would weaken the case. Where inspection or cleaning is frequent, a reliable robot could avoid enough vessel visits to justify staying. Full analysis: research.zengenlabs.com/p/un…

Imagine pairing a fast diffusion model like this and a decision model like JEV. I think we're approaching what will feel like magic.

I had not seen this. This is now the greatest thing I've ever seen.

This is amazing stuff. Impeccable vibes, captivating

how they be describing nathan on youtube

me: hi claude 5.5: python3 - <<'PY' words = ["good", "morning"] print(" ".join(words)) PY

I have long worried about this. Chinese society's respect for, and deference to, anybody who can reasonably refer to themselves as 老师 (teacher) or 师傅 (master) is a huge attack surface and the primary haven for charlatans in the country.

I reckon most teams been working around the Mid-Autumn (September 25–27) and Golden Week (October 1–7) holidays, I wouldn't be surprised if there isn't a flurry soon

I think EUV boils down to a manufacturing problem, precision parts, etc. It’s only a matter of time. And it’s a good thing. Mankind needs lots of high quality and cheap chips

This is the 3rd or 4th tweet I’ve seen where @GoogleDeepMind Argon has had quite an interesting and often better way of reasoning through a problem or approaching it from a very different angle. It goes to show that competition in AI is important, and having different labs produce models with different architectures and training approaches is a huge benefit to society. I find @OpenRouter’s fusion tool very useful. When I use a panel composed of different architectures, I often get much better results than from any one model alone.

The home of the brave

Wild numbers