A third of Chinese cars run on a deliberately terrible business

1. Wintel for robots — ▶ 4:07

   Horizon was named for a market that did not exist. When it was founded 11 years ago, China had no robotics business, so the bet was a unifying computing platform for every kind of robot: farms, assembly lines, warehouses, home service, elderly care and cars. The template is Windows and Intel for PCs, ARM and Android for mobile. Today the business is 95-100% passenger cars, and its chips run roughly one in three Chinese cars, for customers including BYD, Great Wall, Volkswagen, Toyota and Suzuki.

2. Never build a car — ▶ 7:30

   Unlike Xpeng and Xiaomi, Horizon will not make cars. Yu Kai wants to be the "Microsoft and Intel" for robots. His argument: Windows and Intel did not just help Lenovo or Dell, they raised white-collar productivity for everyone, whereas a carmaker helps one brand become attractive. He wants to enable all brands and all cars. Asked how long it would take Horizon to build a car, he agreed it is difficult but gave no timeline, calling the platform a far bigger dream.

3. A market gone crazy — ▶ 9:39

   China has 60-70 car brands, and more than 580 new models launched in the first half of this year. Competition drives it, but so does consumer demand for automatic parking, autonomous driving and fancier infotainment. Horizon's stance is that it competes with no carmaker and enables them instead, building partnerships with carmakers and suppliers such as Bosch and Denso. Making friends and long-term trust is, he says, "in the gene" of the company, which has little competitive culture.

4. Terrible business, great moat — ▶ 12:20

   A new chip generation takes 3-4 years to be ready for carmakers. Winning a nomination then means another 3-5 years to reach production: about 2 in China, typically 5 for global OEMs. That is at least 5-8 years of heavy investment before a slow return. Yu Kai's view: "terrible business means a great business", because mission-critical, safety-qualified silicon cannot be entered by a few people with VC funding. In China the field has consolidated to Nvidia, Huawei and Horizon. A good business is one where you sleep well.

5. Hands, eyes, mind — ▶ 14:27

   The stages are framed for his mother, not the SAE levels, which consumers do not understand. Announced in January last year as 3, 5 and 10 years, the dates are now: hands off by end of 2028, eyes off by end of 2030, mind off by end of 2035, when you can sleep in the car. Horizon SuperDrive on the Journey 6T chip already feels completely hands-free, which he calls a "smarter version of automatic transmission": you only watch ahead.

6. Proof is the bottleneck — ▶ 17:04

   The gap to eyes off is evidence, not capability. Hands off is close, but eyes off needs far more test mileage and better statistics to show it is safe enough. Progress comes from three factors together, "a typical formula of today AI business": much bigger compute running locally in the vehicle, ever more data that is meaningful rather than merely large, especially corner cases, and models with more parameters. Horizon builds its own self-driving models, and designs chips while outsourcing manufacturing.

7. No clocks, 2,000 engineers — ▶ 19:02

   Horizon has around 2,000 people, mostly software and hardware engineers. There is no 8 a.m. start; most arrive around 10 a.m., nobody must stay until a set hour, and some work from home. He says he has no clue how long they work, unlike Chinese companies that measure hours. The faster model cycle at Chinese OEMs, 2 years against 5 for global OEMs, he attributes to fierce competition and hard work, including skipping the 30-40 working days of summer vacation.

8. Vision, not opportunity — ▶ 22:00

   Horizon's culture is driven by vision rather than opportunity or competition, and it avoids what other companies do. That, he says, explains why its business is uncrowded in China, where anything even somewhat interesting draws hundreds of companies, as embodied AI now does. Innovation is rationed: chip architecture, compiler and algorithms must be new. Elsewhere Horizon follows tradition religiously, choosing the best partner, because a crash is a life-and-death public safety issue, not just a problem for the driver.

9. China as one village — ▶ 24:52

   A German friend calls China a "wonderland" for supply chain: whatever the new idea, you can find a partner or supplier, and the speed is incredible. Over two decades the component and technology base became so rich that you do not email somewhere far away. You make a phone call, send a WeChat message, or take a three-hour flight from Beijing to Shenzhen. He and his friends treat all of China as one village, which he says gives it a fantastic environment to innovate.

10. Win-win with Europe — ▶ 26:36

    Asked about Volkswagen cutting 100,000 jobs, he answers that Horizon is a partner, not a competitor. European chipmakers Infineon, NXP and STMicroelectronics focus on analog chips and MCUs, while Horizon does computing for cars. He argues both sides need a new mindset. Global firms making money from China is now difficult, but Chinese firms moving into Europe to take share is "not healthy" either. The alternative he proposes is working together to explore new markets and launch more competitive products.

11. Cars become robots — ▶ 29:14

    By end of 2028 he expects every passenger vehicle in China to be hands-free, already a semi robot. Cars and robots will converge on one computing platform, and the supply chains, from actuators to lasers and radars, are "almost the same". Car volume drives cost and size down, as mobile phone winners later won in IoT. Horizon claims top-two in China's generic robotics computing, supplying chips only. Meaningful humanoid volumes are 5 years away "if optimistic"; he calls himself conservative. He rejects US-China zero-sum thinking.

The barrier in automotive chips is time, not technology. A new Horizon chip generation takes 3-4 years to be ready for carmakers. Winning a nomination from one carmaker then starts a second cycle, 3-5 years, to get onto a vehicle platform. In China that second leg can be 2 years. For global carmakers it is typically 5. Add it up and money comes back 5-8 years after the first design decision. Yu Kai calls this a terrible business and then argues that "terrible business means a great business". A team with venture funding cannot shortcut the qualification process. Automotive silicon is mission-critical, and unlike consumer silicon it must pass tough testing for quality and safety. So the field has thinned to three players in China: Nvidia, Huawei and Horizon. His test for a good business is personal. It is one where you can sleep well overnight.

The second layer of the moat is a refusal. Horizon will not build a car. Xpeng and Xiaomi make cars and chips. Yu Kai wants to be the Microsoft and Intel of robots. He points to what Windows and Intel did for PCs: they raised productivity for all white-collar workers, rather than making one brand more attractive. A carmaker helps itself. A platform helps every brand and every car. The practical effect is neutrality. Customers include BYD, Great Wall, Volkswagen, Toyota and Suzuki, and the chips run roughly one in three cars in China. Horizon can sell to all of them because it competes with none of them. Asked how long it would take to build a car himself, he did not give a number. He said the platform is the much bigger dream.

Chaos downstream feeds the platform. China has 60-70 car brands, and more than 580 new models launched in the first half of this year. Competition drives it, and so does consumer demand for automatic parking, autonomous driving and richer infotainment. Every one of those models needs a computing brain. Horizon sits underneath the churn. The uncrowded position is deliberate. In China, any business that looks even somewhat interesting draws hundreds of companies, and embodied AI is the current example. Yu Kai says Horizon is driven by vision, not by opportunity or competition, and does not want to do what others are doing. Innovation is also rationed. Chip architecture, compiler and algorithms have to be new. Everything else follows tradition strictly, because a crash is a public safety issue.

China adds speed. Its carmakers run a model cycle of 2 years against 5 for global OEMs. Yu Kai credits competition and hard work, and notes that they skip the 30-40 working days of summer vacation. The supply chain does the rest. A German friend told him China is a "wonderland" for supply chain: any new idea finds a partner or supplier, fast. There is no email to a distant office. There is a phone call, a WeChat message, or a three-hour flight from Beijing to Shenzhen. He treats the whole country as one village. Horizon's own culture mirrors it. It has 2,000 people, mostly engineers, no required start time and no measured hours. It designs chips and outsources manufacturing.

The autonomy dates sit on top of this base. Yu Kai describes three stages in language he can explain to his mother: hands off, eyes off, mind off. His targets are hands off by the end of 2028, eyes off by the end of 2030 and mind off by the end of 2035, when you can sleep in the car. Horizon's SuperDrive software on the Journey 6T chip already feels completely hands-free. He compares it to a smarter automatic transmission, where you only watch ahead. The constraint is evidence. Eyes off needs far more test mileage and better statistics to prove it is safe. The route there is the same formula as the rest of AI: bigger compute in the vehicle, more data that is meaningful rather than merely large, especially corner cases, and models with more parameters. The dates rest on accumulating data, not on a missing invention.

The same stack leads to robots. Yu Kai expects every passenger vehicle in China to be hands-free by the end of 2028, which makes cars semi robots already. He expects cars and robots to converge on one computing platform. A car senses its surroundings, reads what pedestrians intend and understands a policeman's gesture. That is the same pattern as any other robot. The supply chain overlap is almost total: actuators, lasers, radars and other sensors. Car volumes force costs down and form factors smaller, as the winners of the mobile phone supply chain later became strong in IoT. So whoever wins in cars wins in robots. Horizon says it is already one of the top two in China's generic robotics computing. For humanoids he is cautious. Hundreds of companies are chasing it, all with a long way to go, and meaningful volumes are 5 years off if optimistic. He frames this as an objective view from a technology founder. The first use he names is home service, and ageing is hitting China harder than Europe or Japan. He also rejects a US-China zero-sum view. American robotics firms that work with China's manufacturing ecosystem could innovate faster. His advice to Europe is similar: stop thinking of each other as markets to extract from, and work together.

The obvious objection is the Windows analogy. Wintel won through software that locked in users and developers. In robots, Yu Kai says Horizon supplies chips and is not so much into the software. In cars it ships its own end-to-end driving stack. So the platform is chips in one domain and a full stack in another, and he does not say which model carries over. The claim of friendship also sits oddly beside his own list of rivals: Mobileye, Huawei, Qualcomm, Nvidia, Texas Instruments, Renesas, NXP and Black Sesame. Huawei and Nvidia are the other two of his three consolidated players. Asked about Volkswagen cutting 100,000 jobs, he answered with partnership rather than reflection. And the eyes-off date arrives in 2030 while he concedes the mileage needed to prove it is not there yet.

The moat protects Horizon from newcomers. It does not say who owns the platform once cars become robots.