SPECIAL EDITION · EMBODIED AI

ShanghAI SPECIAL — Humanoid Robots: “We were too optimistic”

Hey Lobsters, welcome to a special edition of ShanghAI.

This letter is usually about models, tokens, and the invoice that lands at the end of the month.

This week it is about the same cost curve, one layer down: a machine that moves through a building.

You have seen the videos. A humanoid folds a shirt, jogs upstairs, sorts parts. What the video does not tell you is whether anything in it is the product — or whether a person in a headset is standing just outside the frame.

Last week I met with Kuan Yang. He was Strategy Director at AgiBot, the Shanghai company that has now produced 20,000 humanoids and booked roughly ¥2 billion in the first half of 2026.

Before that: automotive engineering, EVs, AI chips. Now he advises European and American clients on the same question you are asking.

Everything in quotation marks below is his, lightly punctuated. Where I have cut a sentence, there is an ellipsis. Nothing is paraphrased into a quote. The AgiBot numbers come from the AgiBot 29th of October Innovation Day announcement, not from Kuan words.

Kuan has also written five books exploring the following ecosystems: EVs, Autonomous driving, ChatGPT and Humanoid robots. He is now sharing his industry insights in a LinkedIn newsletter accessible here.

START HERE
WHAT EMBODIED AI ACTUALLY IS

Three layers, and most coverage blurs them into one.

  • Body. Motors, reducers, actuators, hands. Actuators alone carry roughly 65% of the mechanical bill of materials. This layer is commoditising fast: Unitree lists the R1 at $4,900, the G1 at $13,500, the H2 at $29,900 today.

  • Policy. The model that turns perception into movement — vision-language-action, or VLA for short. It is not trained the way ChatGPT was. It is trained on teleoperation (a human driving the robot), imitation learning (copying that motion), and increasingly simulation. That is the whole reason a video can look finished while the machine is not.

  • Deployment. Integration, safety, certification, spare parts, and the engineer who answers in month fourteen. Boring, expensive, and the only layer that decides whether the thing survives a night shift.

Most of what reaches your feed is layer two, filmed well. Which is where the interview starts.

WHERE CHINA IS
USE CASE BY USE CASE, NOT HEADLINE BY HEADLINE

I asked him whether 90% of what we see is a demo. He refused the headline and gave me the segmentation instead.

"I think it depends on the use case. You can say it's 99% is demos because there are different use cases."

Research and education is real. Universities, professors, research labs, and high schools buying robots as teaching tools. "This is a real deployment... And this is a real business case."

Entertainment is deployed. Dancing, reception, guiding — "I think this is a deployment," with hundreds of units in theme parks.

Factories and logistics are not. "Almost all the robotics they are for the demo for the PoC. So this is a early stage."

And the number that matters more than any of those: two years. "If you really want to deploy robot into the real production line... it needs almost two years. I think from the pitch and from the early demo POC and then you transfer to some SOP and transfer to mass production."

None of this means the industry is vapour. AgiBot has produced 20,000 units, shipped about 9,700 in six months, and claims 43% of the global humanoid market. Unitree did ¥1.70 billion in 2025, ¥868 million of it humanoids, and shipped more than 5,500. TrendForce puts China's humanoid market at CNY 15 billion in 2026, growing at least 60% next year.

Why, then, do new robots launch every month while factories move slowly? Because vendors run both tracks at once.

"I think they are doing the both. One is to develop the new product... because the product will be developed or renewed very very quickly. And another thing is about the deployment in the industry."

Deployment, he says, needs a customer with "a high tolerance" who is "willing to cooperate with you", opens the scenario, and puts engineering in. That is not a purchase order. It is a joint venture with a factory.

THE THREE-WAY FIGHT
FOUNDERS, FACTORIES, INVESTORS

Ask who controls the market and you get a shrug. Ask what each party believes, and the picture sharpens.

Founders believe speed is the barrier: "If you find something, you need to do it right now. But for the factories... they only care about the ROI. They only care about your success rate... If you have a very product which was generated very fast, they think it's not reliable."

Investors are the third force, and their patience is conditional: "Sometimes you will see that you are patient but sometimes they are not patient... They will totally change."

"This is kind of three powers. They are fighting together."

His summary of where the market actually is: "Everyone want to find something. Where is the money come from? I think this is a million questions or topic right now."

WHAT SHIPS, WHAT STAYS
THE PART OF THE DEAL NOBODY PRINTS

This is the answer a European buyer should tape to the wall before signing anything.

"What they can take is hardware itself, the body. Right. And also there is already some of the skills or policy already be trained based on the hardware. And also some documents."

"What is stayed in China? I think is the iteration speed... It's a supplier chain industry cluster... And also the... scenario data flywheel... because it's more close to the manufacturers."

In other words: you buy the machine and the frozen skills. You do not buy the system that made them faster than anyone else, and you do not buy the data loop that improves them every week next to a Chinese factory.

Then comes the part that is not in the crate at all. "You need to get the certification and also you need to get the spare parts which was in the locally. And also you need to get the local team support after sales service... You need to find a really local partner and who have a very good channel."

Which is why buyers hesitate — and why the hesitation is rational. "They think the China made robotics... can do it in three months or six months. They think the product is not reliable." Especially, he adds, "if they don't come to China, if they don't see the real things."

His description of how that changes is the most useful sales document in the industry: "First they don't listen and second okay I give you a chance. I listen... they will compare different players... in the final step will be okay we will rely on you if you have a really very high speed."

THE HONEST PART
THREE MONTHS, TWO YEARS, AND 90%

I asked him what he got wrong at AgiBot. He did not reach for a lesson about culture.

"We were both too optimistic. We think that the general purpose robotics... can solve every scenarios."

They found a customer who believed it, and put a robot in his factory. "We first set up three months. Okay. There are milestone. But then we delay, we delay, we delay... Right now we have been delayed I don't know two years. We think that this is a simple step task but haven't been fully solved." Then, almost as an aside: "Thanks to the customers patient tolerance for that."

His conclusion is not a pitch. "We need to respect this industry. We should be respect for the automation machines... We will be step by step and to bring value."

And the number that should reframe every video you watch:

"What you have saw that the autonomous is mostly... almost 90% is come from the tele operation and come from the coding. So it's not a fully AI."

"It seems that it's autonomous. Right. But behind that you use a lot of human." He does not spare the Americans: "The big player like figure like Tesla they all facing this kind of issues... It is all the teleoperation to collecting the data... to training, to imitation." Better in roughly three years, in his view.

Which produces the line that should end every procurement meeting:

"The customer don't care about which kind of robot you use. Don't care about which kind of AI they use. They only care about the solution... you reduce the cost."

And its corollary, which nobody selling a general-purpose humanoid wants to hear: "Maybe you have a very simple hardware plus some AI perception and with some solutions. Totally, I think that's good."

🔴 🔴 🔴 A small experiment, then.

In spring we are bringing a cohort to Shanghai and Shenzhen factory floors, to meet the builders themselves and come back with facts and partnerships. There is no page and no price list yet.

If you want the details when they exist, reply to this email with "ground truth" and I will send them to you directly. That is the whole pitch.

Until next week,
The ShanghAI Guy