Kai Yu, Horizon Robotics: The Scientist-Founder Journey from Zero to One | 5Y View
"Be roughly right, not precisely wrong."
"If you have a vision and mission ahead of the market, if you're ambitiously committed to changing how people think and how industries operate, and if you decide to influence everyone around you through evangelism — you might become a minority in people's eyes. At the same time, you'll face ridicule and doubt, because genius and madness are often separated by only a thin line."
This was the situation Kai Yu faced when he founded Horizon Robotics in 2015. Now, eight years later, market feedback has step by step validated his far-sighted vision and accumulated strength. Looking back at the first five years of entrepreneurship, Yu described "too many dark moments." During 5Y Capital's 15th Anniversary CEO Summit, Kai Yu, founder and CEO of Horizon Robotics, shared his entrepreneurial story.
How does a scientist-entrepreneur make it from zero to one? How does an entrepreneur survive the long dark phase? We've selected some excerpts that might offer you some inspiration :)
The following are excerpts from Kai Yu's sharing at 5Y Capital's 15th Anniversary CEO Summit:
Horizon Robotics was founded in 2015. At that time, it was probably the first company in China to move from software algorithms into deep neural network chips, and also the earliest Chinese company to build dedicated deep neural network chips. In a sense, we created a new track. I think the type of entrepreneurial company that Horizon Robotics represents was perhaps rare in China before, but is becoming more common now — scientist-founded startups.
Scientist-entrepreneurs have a typical characteristic: they have grand visions for the future. Scientists are particularly endowed with endgame thinking and very firm convictions — otherwise they wouldn't spend long years doing research in obscurity. But when it comes to commercialization, product development, where the customers are, what the demand is — often this type of entrepreneur has given little thought to these questions.
I'd like to share Horizon Robotics' story. One thing that might be helpful for tech entrepreneurship today is how scientist-entrepreneurs climb over the pitfalls of the 0 to 1 phase. Horizon Robotics went through too much blood and tears in the middle, and these lessons might be helpful for everyone.
Let me first share some background. Before returning to China in 2012, I mainly did research and wrote papers — over 100 publications. When I came back, I was probably among the Chinese scholars with the highest citation rates in artificial intelligence. If you know AI, you may have heard of ImageNet, known as the Olympics of deep learning and AI. I once led the NEC Labs research team to win the first ImageNet competition in 2010.
New York Times reporter Cade Metz wrote a book called Genius Maker, about the history of deep learning's development over the past decade. The first chapter tells a story that I initiated.
In April 2012, I left NEC Labs in Silicon Valley and returned to Beijing to join Baidu, responsible for AI technology R&D. At that time, GPUs weren't yet widespread, and computing power was very constrained, so we tried to find methods that didn't consume too much computing resources. We used supervised learning with convolutional neural networks to achieve industry-leading results, but there wasn't much breakthrough the following year — basically continuing along the first year's approach, as other teams were doing too.
In 2012, Geoffrey Hinton and his two students achieved a massive breakthrough. One of those students was Ilya, now Chief Scientist at OpenAI. I was extremely sensitive to this development and immediately wrote to Hinton expressing interest in collaboration. This triggered a secret auction — I represented Baidu in bidding against Google, Microsoft, and DeepMind.
What's interesting is that the companies didn't know about each other at the time — it was a secret auction. DeepMind was then only a one-year-old startup, but already showed tremendous ambition in AI, daring to bid against three tech giants. Our first bid was $12 million. In the end, Google successfully acquired Hinton and his two students' research team for $44 million.
This secret auction could be called the "starting gun" for deep learning driving transformation across the global tech industry. After this, Microsoft, Google, Baidu, and other tech companies, seeing how much their rivals were willing to invest, all simultaneously kicked off a deep learning boom. Before this, deep learning was largely academic research in an ivory tower; afterward, it became a race among tech companies. Today's DeepMind and OpenAI were both participants in that auction.
Although we didn't win the auction, everyone benefited. Baidu also recognized the importance of AI and established the Baidu Institute of Deep Learning (IDL), where I served as Executive Deputy Director. IDL pioneered the establishment of frontier AI research institutions in Chinese companies, and has long been called the Whampoa Military Academy of Chinese AI. This is quite an interesting piece of history, and I feel particularly connected to it. This was my background, and I went into entrepreneurship with this foundation.
Anti-Consensus Is a Startup's Only Opportunity
In the summer of 2015, I left Baidu to start my own company. Before the company was even established, I chatted with Richard at Beijing's Four Seasons Hotel from 8 p.m. until past 3 a.m., until the entire hotel had turned off its lights. We talked about the future of robots and chips, extremely excited — though looking back, many of those ideas weren't very practical.
At that time, I went to "recruit" Richard with this thinking: I said we would make brain chips for robots, giving every appliance and every vehicle the capabilities of environmental perception, human-machine interaction, and decision control. In short, we wanted to be the Intel of the robotics era, or even Intel plus Microsoft.
Today, eight years later, the vision for the future seems like a kind of gravitational pull — you may do many other things in between, but somehow it still draws you back to that place. Today, Horizon Robotics is China's largest supplier of automotive intelligence chips, and probably also China's largest supplier of robot vacuum chips — we've returned to our original direction.
Of course, we had many deliberations along the way. Some colleagues suggested making cloud computing chips, mobile phone chips — but we rejected these ideas. I think one thing is particularly important: a startup must bet on something that seems small now, that big companies look down upon, but that may have huge potential in the future. If you go after businesses where the market is already very mature, competing head-to-head with established companies on their home turf, you probably have no chance. On someone else's home battlefield, even if they make mistakes, it's only a tiny probability. You must persist in being anti-consensus, doing non-mainstream things — this is a startup's only opportunity.
When I first started out, as a founder, my core capabilities were technology and having a long-term vision for the future, but I didn't know where the customers were or how to build products. It took me five years — until 2019 — to suddenly have an epiphany. The first principle of business is not what technology you have, but who your customer really is. At that time, I asked myself this question repeatedly every night.
Over the past decade, there have been many AI-related startups in China, but most eventually became mediocre. It wasn't because their technology wasn't great, their team wasn't great, or they couldn't raise funding. They lost because they failed to understand demand. If your business is built on irreversible long-term trends in human society, you should be very certain. However, if you're just catching a short-term fad, when a bigger wave comes later, it's hard to ride it.
Schumpeter believed that true commercial innovation is actually the combination and application of different technologies. His meaning was that recombinations of existing mature technologies oriented toward demand can give birth to new products and new service models. For example, QR codes were invented by a Japanese auto parts supplier, but Tencent uses them best. Large language models may not have been invented by us, but the great commercial companies will certainly be those that use large models best — they are in the process of being born.
How to find the precise matching point between technology maturity and commercial closed loop is extremely critical. From the day you start a company, the clock starts ticking. At each stage, you must complete what that stage requires — for example, finding product-market fit at a certain point in time. You have to manage commercial and organizational uncertainty while also managing too much technological uncertainty. Carrying so much uncertainty in a small startup is extremely challenging. So at different stages, you need to build different organizations and cultures — many companies may die at this point.
Be Roughly Right, Not Precisely Wrong
Let me share what I think is a reusable framework for strategic and organizational evolution.
From a company's early startup phase to maturity, it can roughly be divided into four stages. The first stage, 0 to 0.1, is the strategic experimentation phase. Faced with a grand vision, you need to find an entry point and require extensive trial and error. Of course, China has many opportunity-driven startups where demand has already been validated or there's a benchmark product. But if you're vision-driven, creating a new category with no customers in sight, no demand, no benchmark company or product — what should you do?
You must maintain your vision for the future. Be roughly right, not precisely wrong. I remember discussing application scenarios like toys, furniture, and home appliances with Richard at the time, and later found them all unviable. In the early stage, you definitely need constant trial and error, and passion for the vision will sustain you to keep going despite mistakes, pushing forward until you find the entry point.
At this stage, there's not much organizational building to speak of — processes serve efficiency. But when the direction is wrong, the better your processes, the more wrong you may become. Often it's even hard to set KPIs. For example, if you set a KPI and don't achieve it by year-end, you'd feel embarrassed to yell at your sales lead — because it was the CEO who pointed the wrong direction.
In the dark phase, the core organizational capability is the founding team. You must constantly iterate your own understanding and survive through the darkness. At this point, you rely on faith and mental strength, not professional capability. At this stage, you must constantly encourage and motivate the team. This is the innovation trial-and-error phase, until you basically find direction — perhaps when you have your first paying customer.
This phase took Horizon Robotics about four years, from 2015 to 2019.
From 0.1 to 1 is the strategic formation phase. In 2019, we underwent major strategic iteration, cutting 500 people from 1,200. Looking toward the future, what we were most certain about was the automotive direction. If we only had one bullet in our chamber, we had to aim at the bullseye. So we cut all non-automotive businesses entirely. The auto industry was also in a very dark period that year, but we were very certain about this direction. Changan Automobile was already our customer, and we knew this project had to be carried through — so we persisted.
At this stage, understanding of market demand and customers gradually converges, and the team forms consensus on direction. We found an inflection point — no longer just the first customer, but second, third, and fourth customers. We found real demand, found the market to attack, and got through the entire 0 to 1 phase.
Today, Horizon Robotics is in the 1 to 10 strategic expansion phase, where business is relatively clearer. At this point, we need large-scale shipment, so there are higher standards for supply chain management, inventory, quality management, and financial management. The most painful and difficult stage in strategic management for companies is mostly in the 1 to 10 phase. Many companies climb out of the first pit of 0 to 1, then immediately fall into the 1 to 10 pit — and sometimes the latter is even more deadly. Because at this stage, the competitors you face are all top players in the industry. Tolerance for error is lower — for example, if inventory management has problems, you might immediately face billion-level issues.
What's needed now isn't guerrilla troops, but professional teams that can play in the World Cup final. Some talent that grew in the 0 to 1 guerrilla phase may not grow further, because the competency model has changed — this is very frightening.
At this point, the CEO must first evolve themselves, which is of course extremely difficult. Because before it was mainly about innovation, now it's mainly about operations and management, while carrying innovation within operations and management. You must care about direction, efficiency, and quality. Silicon Valley has a very good model — it has genius coaches like Bill Campbell who take founders from 0 to 1 and 1 to 10, but some founders actually can't adapt to 1 to 10 after making it through 0 to 1.
The strategic expansion phase has high demands for professionalism, processes, and organization. When a company starts with 1 to 10 people, basically everyone can figure things out over one dinner table. At 100 to 200 people, one meeting to communicate can still work. But what about 2,000 people? Many problems may emerge — silos, lack of collaboration. At this point, you need culture, processes, mechanisms, rewards and punishments, and performance to ensure coordination in a complex organization. This is very difficult, and many companies die at this stage.
This model has been very useful for us. Even at the micro level, every new project in a company will go through 0 to 1, 1 to 10, and 10 to N phases. Matching appropriate resources, organization, and strategic goals to different stages is extremely critical. I can use this model repeatedly, and hope it will be helpful for all of you too.
Compete Where There Is No Competition
You may form some mental frameworks in the 0 to 1 phase. I mainly want to talk about the mindset of non-competition.
Every entrepreneur may naturally want to win, or be combative, or like competition. But when truly making strategic trade-offs, competition is a very poor strategy. Like in the TV series "Drawing Sword," where the brave win when two armies clash — Li Yunlong's strategy is very unwise.
A truly good strategy must choose to compete where there is no competition — choosing an area that most people look down upon, don't understand, or aren't willing to do, but that will have huge opportunities in the future. Or, even if everyone sees this point, you have 10x resource investment. If there's an opportunity, pour in 10x resources and charge ahead. Truly excellent strategy doesn't gamble.
For organizations in the 0 to 1 phase, I think the founding team is extremely important. Some young entrepreneurs ask me before starting whether they need co-founders. I say it's best to have them. This isn't just about complementary capabilities — in the 0 to 1 phase, you'll go through many things, experience long dark periods, need someone to tell you the truth, and need somewhere to speak the truth. Founders trust each other. If everyone has great格局 (vision/character), each can sacrifice local and personal interests for the company's development, providing strong psychological support for each other.
On the other hand, the founding team's learning ability must also be strong. First, you must keep growing yourself. At the same time, you must have the胸怀 (breadth of mind) — when there's a capability you can't grow yourself, you must find someone better than you. At this point, having great格局 is very important — this is what co-founders must do.
In leadership at this stage, I believe mental strength is greater than intellectual strength, and intellectual strength is greater than manpower. In the 1 to 10 phase and 10 to N phase, professional capabilities become increasingly important. You must fill in your weak spots and find world-class players to play the World Cup final together.
I think Horizon Robotics has been very fortunate. We entered a field that may not have been very large initially, but by 2030 or 2035, we believe the core main computing chip in every vehicle will exceed $1,000. $1,000 times 100 million vehicles — this will be a market larger than today's smartphones.
The automotive field concerns human life safety, with its own thresholds and product iteration cycles. From day one, this field hasn't been a country club league, but the World Cup final. We've now received our ticket to the World Cup final. At every stage going forward, we'll face many risks and challenges, and we will continue to maintain a sense of awe.
5Y Capital seeks out, supports, and inspires lonely entrepreneurs, providing support from spirit to all business operations. We believe that if the "crazy you" in others' eyes begins to be believed in, the world will become a different place.
BEIJING · SHANGHAI · SHENZHEN · HONG KONG