In some sense, this is terrible business — you know, heavy investment, very slow return. But in my view, sometimes a terrible business means a great business, because that's the entry barrier for all the other newcomers is extremely high.
Everybody, I'm Nicolai Tangen, the CEO of the Norwegian Sovereign Wealth Fund. Ten years ago, we all assumed self-driving cars would come out of Silicon Valley. But today, the most advanced cars on the road are Chinese, and Western car makers are going to China to buy the technology. Nobody sits closer to the shift than my guest Yu Kai, who's the founder and CEO of Horizon Robotics, which supplies the chips and software that let Chinese cars drive themselves. He was one of China's leading AI scientists before he became an entrepreneur, and few people can explain better how China got there and where he goes next. So, Kai, welcome.
Let's just start with a simple question: What does Horizon Robotics do?
Well, essentially what we are doing is the computing platform for all generic robots. So including self-driving cars, but of course our current business is probably, you know, 95, nearly 100% focusing on cars, passenger vehicles.
So if I sit in a car that you supply into, what does your chip do?
Well, essentially it's running the software for all the way from the camera perception to trajectory planning and control. So essentially it's kind of like the brain for all the cars.
Now your chips run roughly one out of four cars in China. Just tell me about the...
One out of three, I think. That's our latest track record.
Thank you. Well done. Who are your customers?
Well, you know, the major Chinese OEMs like BYD, Li Auto, you know, Chery, Changan, Great Wall, and also some of the global OEMs like Volkswagen, Toyota, Suzuki. So quite some local, domestic and global players.
Oh, very good question. I think like Mobileye, you know, and also Huawei and quite a few domestic players like Black Sesame — I'm not sure if you've ever heard of that — and Qualcomm and Nvidia, Texas Instruments, you know, Renesas, NXP. So actually a lot of global and China technology companies.
Oh, well, a very good question. I think it's really because we are very focused — you know, we focus on this single one thing to enable autonomous driving and all the way to robotics. We are not doing anything else. So I think the intensity of focus is one thing. The other thing I would say is fundamentally our vision and mission. Because if you look at the name of our company, you know, 11 years ago when we founded this company, we named our company Horizon Robotics. Right back to 11 years ago in China, there was even no such a business at all. And even the AI business was just starting.
Right, and we had this vision which was not an existing opportunity. So we envisioned in the future the world will be full of all kinds of different robots, you know, working in a farming field, manufacturing assembly line, and the logistic warehouse, you know, every house, home service and elderly care, and also of course self-driving cars. So we envisioned there should be a unifying platform just like, you know, Windows and Intel did for personal computing, and just like, you know, ARM and Android, what they are doing for mobile computing. So we envisioned in the future there's going to be such a kind of computing platform which is generic for all kinds of robotics working across different vertical domains.
When exactly did the vision hit you? Were you in the shower, or were you walking in the mountains, or just what?
Well, that's a very good question. I think exactly you are right. I think the vision actually hit me — not that I found this opportunity or this vision. Wow. I don't know, you know, I have been working in machine learning for over 30 years, you know, since my early days as a junior student in college. And by chance I read a book about machine learning, and I was so amazed that it's so fantastic to develop a computer algorithm one day that can be like a human brain, getting smarter and smarter over experience, by learning more and more from data. So then I decided to devote my whole career into this single subject: machine learning. And that was for my thesis for my undergraduate, and then my thesis for my master, and also for my PhD, and also for my first job back to my days in Siemens, and later to Silicon Valley, joined SRI Labs in Silicon Valley, and then later joined Baidu, and then later founded Horizon Robotics. So I changed my locations and I changed kind of different roles from, you know, individual researcher to a team leader to a head of research labs, and later became a founder of a technology company. So I changed my roles, but there's one thing, you know, remains never changed, which is my mission and my vision to develop machine learning algorithms to make machines getting smarter and smarter.
Two companies that we've had on the podcast, XPeng and Xiaomi, they also make phones, cars in addition to chips. So why did you decide never to build a car?
I always position ourselves going to be like a Microsoft and Intel for — not for PC, but for robots. And I think a platform, an enabling platform, I think it's a much more interesting positioning, not just about like a business model, but also about in terms of the impact you're going to make to the whole world. You know, for example, you know, back to the days of the personal computing era, you know, Windows and Intel, they are not just, you know, helping or enabling like Lenovo or Dell, they actually help the entire mankind, you know, get into the new kind of a sort of a technology civilization or new era of technology, which is personal computer, to enable all the white-collar people to increase dramatically their productivity. Right? It's an amazing impact. So they actually maybe help one brand, you know, to become very, very appealing, very, very attractive to consumers. But I want to enable and help all the brands, all the cars.
Is it difficult to build a car? So, but you don't have the technology — you have the chips. So if I ask you to build a car, how long would it take you?
You know, I wanted to achieve like a Horizon inside, you know, that's — I think that's a much bigger dream, much more interesting for me. I don't want to be a car maker, you know, and I want to help more car brands to become more attractive.
Now the Chinese car market is kind of brutally competitive. What is driving the Chinese car market just now? Is it price, intelligence, just what is it?
Well, the current Chinese car market is really crazy. I think you don't see any market like here. There are, you know, I don't know how many, probably 60 to 70 different car brands, and for the first half of this year, more than 580 new car models launched. And it's an extremely tough market. But I think the market is really driven by competition. But on the other hand, it's also driven by consumer demand. So consumers are really expecting, you know, more and more smart features, you know, automatic parking, you know, autonomous driving, and the more fancy infotainment systems. And I think all of those factors are driving the market — not just the competition, but also very innovative, very innovative.
What does it do to Chinese manufacturers when you have such strong competition at home that when you get out to, let's say, Europe, is it just like easy peasy to compete relative?
I think we position ourselves — first, we are not competing with any car makers. We are actually enabling them. We are helping them. We are building strong partnerships with all the car makers and all the ecosystem partners, including the automotive suppliers, for example Bosch, Denso, and many other global and local suppliers. And our philosophy is really make friends and make a long-term trust with our customers and partners. Essentially, this is fundamentally in the gene of our company. So we don't have too much culture from the competition perspective, but more from, you know, what kind of value we can create. But luckily, if you look at our business, I mean, fundamentally compared to those car makers, we are not so crowded, right? And if you look at China, I think pretty much the three top players for automotive computing, I think nowadays, really consolidated to Nvidia, Huawei, and Horizon Robotics. Not very crowded.
Yeah, partially because we chose the position where this business is having extremely high entry barrier. You know, it's not like, okay, a bunch of guys having VC funding and then easily get into the business, because this is automotive. Automotive means mission critical, and it has to have the highest standard for the quality, for the safety. You know, and let's say we are doing chip design for automotive, right? It's not like a consumer type of silicon, and it has to go through all the very rigorous and tough qualification and testing process. So normally, if we initiate a new project, a new generation of product, it will take three to four years to get the product, the chip, and the solution ready for car makers. And then if luckily enough we win a new nomination from one car maker, then we need to also go through the cycle of the production for the car platforms, and that's going to be another three to five years in China — two years, right — for global OEMs, they are slower, typically five years. So then think about it: it's going to be at least from five years to eight years for the cycle. So in some sense, this is terrible business — you know, heavy investment, very slow return. But in my view, sometimes a terrible business means a great business, because that's the entry barrier for all the other newcomers is extremely high. And then I think a good business is something where you can sleep well overnight, and that's my opinion.
Talking about sleeping well and the autonomous journey. So you split between hands off, eyes off, mind off. These are kind of the three steps that a car will go through. And in January last year, you put dates on them. So we'll have hands off in three years, we'll have eyes off in five, mind off in ten. Just where are we today?
Well, we are in very good progress. I think, you know what, in the industry there's sort of like a standard from like a level two, level three, level four, level five. It's really from the engineering standpoint. For the consumers, you know, they have no clue what is level three. You know, so I always try to connect our technology and the product to something where my mom at least she can understand. So then that's my way to interpret the technology stages from, you know, hands-free, hands off, eyes off, and minds off.
So what do you tell your mom just now?
Right now we are in very good progress to achieve fully hands off. Yeah, I think — then I said that I made that statement last year, right? So that means by end of 2028 we're going to achieve 100% hands off, and then by the end of 2030 to achieve 100% eyes off, and then by the end of 2035 going to be 100% minds off — basically you can sleep in your car, you know, running car. And I think we are in very good progress. And if you look at our product, our latest product, namely the software stack called Horizon SuperDrive running on our Journey 6T SoC, currently is essentially you can feel completely hands-free.
Yeah, it's just like a new version, a smarter version of automatic transmission. You know, if you remember, you know, over the last 20 years, so the whole industry, the car industry, evolved from like the manual transmission, and you need to control all the gears, and later getting into the automatic transmission, right? But now I call this getting into the smarter version of automatic transmission. Basically you just need to, you know, like stay watching ahead and nothing else you need to take care.
What is still lacking? What part of the technology are you still — what's the critical element which is still left here?
Well, I think we need a much bigger mileage to test, and we need to have better numbers for statistics, and then to prove that the technology is safe enough to claim, you know, eyes off. Right. So I think already we are in very good progress to approach the stage of hands off, but I think we need a much bigger, you know, mileage to test and prove eyes off.
Do you build your own AI models?
Sure, sure. The self-driving model, yeah. You remember we position ourselves as a Wintel, right? Not just Intel — we position ourselves as a Wintel, and I think Wintel is really, I think it's an amazing business model.
And where is the kind of the future breakthrough going to come? Is it the model, the data, the chips?
Well, it's a combination of three factors together. I think this is a typical formula of today's AI business. I think you need to have much bigger compute running locally on the vehicle, and also you need to continuously collect bigger and bigger data, but also meaningful data. It's not just quantity, but also quality. That means you need to collect many, many of the meaningful corner cases, right? And I think also you need bigger and bigger, you know, models in terms of number of parameters.
How many people are you now in the company?
I think we are a very technology-savvy company. So most of our employees, they are engineers. They are either software engineers or hardware engineers, and they are working extremely hard.
Because you don't produce the chips. You outsource the production, right?
Exactly. We are like a typical — on the hardware side, we are a typical fabless company. We design the chip, but we don't manufacture the chip — we outsource to our partners. And also we develop the software, the end-to-end neural network for self-driving.
What does working hard mean?
You know, we are kind of like a typical technology culture. We don't have the rules or the policy that you need to come to work at 8:00 a.m. Actually, typically most of our engineers, they come to work around 10:00 a.m. And they freely organize their working, you know, time. And there's no rules like, okay, before which time you are not supposed to leave office. So there's no such a thing. I think the sense of ownership of our engineers, I think it's very high.
What do you think on average — how long days do you think on average they work?
Actually, I have no clue. And quite some of our engineers, they even work from home, right? Yeah, we don't measure that by like working hours. I do know some of the China companies measure by working hours. We don't — we never measure that. We just, you know, caring about, you know, what kind of — how far we can achieve.
When you say that the model time in China is, you know, the cycle time is two years versus five years in Europe, what explains the difference?
As far as I know, because I work with so many Chinese OEMs, they are working extremely hard. They have a very strong sense of competition. That urges them to try all the ways to improve the process, you know. And yeah, but they work really hard. Yeah. For example, they don't have summer vacations like 30 or 40 working days off during the summer, right? So I think they are working really hard.
What's the key to innovation? How do you create the really innovative culture?
I think first of all, we are a company, you know, that is really driven by the vision. And we are kind of — we are not like opportunity-driven, right? We are not competition-driven. We work on that because we really want to push the boundary of the entire society and the industry. I think this self-motivation is very important. That means we don't want to be the same as any other company, you know. We don't want to do sort of things other companies are doing. So that's why also explains — that's why our business essentially is not so crowded in China, right? This is actually a very real phenomenon, because in China, if any business looked even somewhat interesting — not even profitable, but somewhat interesting, right — typically hundreds of companies jump in to compete. It's crazy. Look at today the kind of, you know, AI and the embodied AI business, right? So many companies, and also so many car companies. I think we are kind of working on something, you know, which we believe, but not necessarily, you know, other people love to jump in.
The fact that you have all these companies jumping into these sectors, what does it mean in terms of the supply chain and the various choices you have as a manufacturer? How does it all go together?
