The biggest bug in scientist-founded startups: not knowing who the customer is
For its first five years Horizon sprayed 360 degrees across air conditioners, toys, home appliances and cars; in 2019 it cut half its people and bet everything on automotive, turning from a company firing in all directions into a strategy-driven one.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Fundraising is a funnel, and the first meeting must never be in the investor's office
Yu Kai breaks an investor's decision into a funnel model: from never heard of it, to aware, to interested, to doing research, to visiting the team, to forming a view, to playing a card. If you enter this funnel early, you have to spend a long time persuading the other side. So he set himself an iron rule: the first meeting with an investor is absolutely never in his office, it must be in your own office — if the other side comes knocking, it means they are already at the bottom of the funnel, close to playing a card. He used "no time" and "the project is very busy" to filter out the people still at the card-looking stage, and in the end only four institutions were seriously negotiating, each willing to sign an exclusivity agreement.
— Yu KaiThe common disease of scientist founders is spraying 360 degrees
Yu Kai says Horizon at the start only knew it "wanted to build a hardware platform and a software platform for the robotics era," but "which business scenario to enter from, we had absolutely no idea." So automotive, AIoT, toys, home appliances — all were tried, resources spread across so many directions, and "every point was very hard to break through." He attributes this to the typical problem of scientist founders: triggered by a very big vision, believing they have technology, but having no concept at all of the business scenario, so it becomes all kinds of spraying rather than focusing — aiming at one hill and pounding it with all the firepower.
— Yu KaiCutting half the people was a deliberate strategic focus
In November 2019 Yu Kai figured out he had to "give up nine to take one," cutting the AIoT direction. The company had 1,200 to 1,300 people at the time, and he cut nearly half, about 600 people. He stresses this was not the "Horizon can't survive" that outside reports said — there were over three billion in cash on the books at the time. He calls this focus the one that remade Horizon: from a company firing in all directions with no strategic thinking into a truly strategy-driven startup team with clear goals and a clear customer definition.
— Yu KaiSplitting the layoffs into three rounds was to make himself comfortable
The original plan was to cut a bit before the New Year, a bit after, and a bit more mid-year. One night Yu Kai woke from a dream and realized that doing it this way "was essentially to make ourselves comfortable, not to be good to the employees": employees would go home happy for Spring Festival and come back to work only to be told they were laid off. He changed it to finishing all the layoffs within one month, from December 15, 2019 to January 15, 2020, and personally talked to a dozen or twenty people. His judgment: when facing a tough decision, do the right thing.
— Yu KaiRaising in reverse during the bubble, without raising the valuation a cent
From the second half of 2020 to the first half of 2021 the capital markets were in a frenzy. Yu Kai judged that "the bubble would definitely burst," so he went the opposite way: the shareholders' meeting gave a cap of about 800 million US dollars, and after raising that quickly, he applied to expand it and immediately took another 800 million, 1.6 billion in total, without raising the valuation a single cent. At the time most companies raised two or three hundred million and then raised the valuation, then raised again at a higher valuation. Later the market turned sharply down and most companies couldn't raise money at all; Horizon, with a hundred in hand, was not anxious at all.
— Yu KaiAutonomous driving is a standard feature; OEMs shouldn't build it in-house
Yu Kai's counter-consensus judgment: the user experience standard for autonomous driving is unified — from A to B, safe, comfortable, efficient, no difference between men and women, old and young, Chinese or Japanese; you can't create differentiation. He uses phones as an analogy: no phone maker builds its own calling function, even Apple doesn't build its own baseband. So OEMs should spend their energy on things like cameras, which have no common standard, vary from person to person, and carry emotional value, and leave standardized autonomous driving to suppliers. He admits the vast majority of OEMs disagree with this view.
— Yu KaiLi Auto agrees with the logic, but Li Auto is not a car company
Yu Kai says he discussed this with Li Auto, and Li Auto actually agrees that autonomous driving is a standard feature and from that angle not worth building in-house. But the key is "don't forget what Li Auto's ideal is" — Li Auto would love to compete with Google, Microsoft and OpenAI in the future. So doing autonomous driving is its path to artificial intelligence, to sustaining war with war, training the team, and building capability. Yu Kai says that once he understood this, even if one day Li Auto stops working with him, he can accept it, because making cars is only one step for Li Auto.
— Yu KaiA human brain runs 5000T on 20 watts; the chip draws over 100 watts
Yu Kai says cars and robots are different from cloud computing: the cloud has almost unlimited power and good cooling, while cars and robots depend on their own batteries, with limited power and poor cooling. The human brain can achieve 5000T of compute on 20 watts of power, ten times the compute of Horizon's highest-performance chip, Journey 6P, yet Journey 6P alone draws over 100 watts. To reach 20 watts and 5000T, you must move far away from today's von Neumann architecture and completely merge compute and storage, like the human brain.
— Yu KaiIn their own words · checked verbatim
The first meeting with an investor absolutely cannot be in his office, it must be in my office, because if they're in my office, it means they came knocking, they're definitely at the bottom of the funnel, close to the moment of playing a card.
我跟这个投资人，第一次见面，绝对不能在他办模式，一定要在我办公室，因为在我办公室的，它说明，它敲门来了嘛，它肯定是在漏斗模型的下端，是接近要出牌的这个时候。
Yu Kai1:16:03
Scientist founders, I think, usually have this problem: they're triggered by a very big vision, then they feel they have technology, but they have absolutely no concept of the business scenario, so it becomes all kinds of spraying, rather than focusing — aiming at one hill and pounding it with all the firepower.
科学家创业，我觉得通常会有这个问题，就是说，他做一个很大的愿景的触法，然后呢，他觉得自己有技术，但是他对于这个商业的场景，是完全没有概念的，所以就会，各种的这个扫射，而不是聚焦的，对着一个山头，其中所有的炮火猛轰。
Yu Kai1:22:11
So we thought, this was essentially to make ourselves comfortable, not to be good to the employees.
所以我们想，这个本质上面，是为了让我们自己舒服，并不是对员工好。
Yu Kai1:35:19
Don't forget what Li Auto's ideal is — it is not a car company; Li Auto would love to compete with Google, with Microsoft, with OpenAI in the future.
不要忘了，理想的理想是什么，它不是一个造车公司，理想是恨不得将来跟Google，跟微软，跟OpenAI PK的。
Yu Kai2:19:46
Figures
| Horizon layoffs in 2019 | about 600 people (the company had 1,200-1,300 at the time) | 1:26:16 |
| Cash on Horizon's books in 2019 | over three billion | 1:28:16 |
| Total raised by Horizon in 2020-2021 | 1.6 billion US dollars (without raising the valuation a single cent) | 1:49:26 |
| Number of shareholders/investment institutions before Horizon's IPO | 102 | 2:02:39 |
| Power draw of Horizon's Journey 6P chip | over 100 watts | 2:24:48 |
| Human brain power and compute | 20 watts, 5000T | 2:24:48 |
| Production project cycle for Li Auto and Horizon Journey 3 | 8 months | 1:45:22 |
| Changan UNI first-month sales | over ten thousand units | 1:40:19 |
Glossary
- PMF / Product Market Fit
- Product Market Fit — the product has found a market it can enter.
- Cuda / Nvidia's parallel computing platform
- Nvidia's GPU parallel computing ecosystem, the underlying standard for AI training.
- Baseband
- The chip module in a phone responsible for processing communication signals.
- HMI / Human Machine Interface
- Human Machine Interface — the interface between the car system and the person.
How to listen
Founders building hard tech or AI startups, investors looking at early-stage projects — especially teams stuck at the stage of "a big vision but no entry point."
The rapid-fire Q&A and extras from 2:35:23 onward can be skipped; the information density is clearly lower than the main body.