NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Yu Kai left Baidu to build self-driving chips nobody wanted — now one in three Chinese smart cars runs his silicon

How a deep-learning scientist built Horizon Robotics into China's No.1 automotive AI chip supplier in a decade.

2026-11-15 · 1039 words · NeuroAI
Yu Kai left Baidu to build self-driving chips nobody wanted — now one in three Chinese smart cars runs his silicon

On a conference stage in Chengdu in September 2026, Yu Kai told a room of car executives that his company would outsell Nvidia in China's high-end self-driving chips within a year. Most people in the industry would call that reckless. The number behind the claim is why it isn't.

Yu Kai is the founder and CEO of Horizon Robotics (地平线), the company whose Journey (征程) series of edge AI chips now powers the driver-assistance systems of roughly one in three Chinese smart cars. Ten years ago, he was a research scientist who had never shipped a chip. The story of how he got from one to the other says something useful about where China's automotive silicon actually stands.

The academic who walked away from Baidu

Yu's résumé reads like a tour of modern AI's institutions. He earned bachelor's and master's degrees from Nanjing University in 1998 and 2000, then a PhD in computer science from the University of Munich in 2004. He spent the next decade inside the labs that defined the deep-learning era: Siemens' neural-computing group, then NEC Labs in the United States, where he ran a media-analysis team and even lectured part-time at Stanford in 2011.

In 2012 he returned to China to join Baidu, where he built the multimedia department and, in 2013, founded the Baidu Institute of Deep Learning (IDL) — one of the first frontier-AI research labs inside a Chinese internet company, with Robin Li as its nominal head and Yu as executive vice president. He seeded Baidu Brain and the company's earliest autonomous-driving program. By 2015, with deep learning already reshaping search and advertising, he left.

What he did next surprised his peers. Instead of joining another platform company or starting yet another algorithm startup (创业公司), he decided to design silicon.

Why a chip, and not another app

The logic came from a frustration he has repeated in interviews: the same deep-learning model ran dozens of times faster on a GPU than on a CPU. If algorithm efficiency depended so heavily on the hardware underneath, he reasoned, then the highest-leverage place to invest was the chip itself — not another layer of software on top of someone else's processor.

Horizon, founded in 2015, set an unusually blunt mission: to become the "Intel and Qualcomm of the robotics era." Colleagues offered him easier paths — phone chips, cloud chips — and he turned them down. He bet on a then-cold category: low-power, automotive-grade AI chips for advanced driver-assistance (ADAS) and autonomous driving, plus a future in robots. The bet looked lonely for years.

From zero to one in Chinese cars

The payoff came slowly, then all at once. Horizon's Journey (征程) family moved through several generations — Journey 2, 3, 5, and the Journey 6 series launched in 2024 — each tuned for the specific math of perception, fusion, and planning inside a car.

The company's own figures, echoed by industry trackers, tell the scale story:

  • In 2024, Horizon held a 33.97% share of China's autonomous-driving compute solutions for domestic-brand passenger cars — the No. 1 position, the equivalent of "one in three."
  • Its systems had been adopted by 27 carmakers across roughly 290 vehicle models by late 2024.
  • Revenue reached RMB 2.38 billion (about US$330 million) in 2024, up 53.6% year on year.
  • The Journey 6 series alone shipped more than 7 million units in its first year on the market.

In October 2024, Horizon listed on the Hong Kong Stock Exchange (stock code 09660.HK), raising HK$5.4 billion (about US$696 million) — one of the largest technology IPOs in Hong Kong that year. Cornerstone investors included Alibaba and Baidu. By the company's own account in 2026, cumulative mass production of Journey chips had surpassed 15 million units.

What the Journey chip actually changed

The technical point is worth separating from the fundraising hype. A carmaker can either buy a general-purpose accelerator and struggle to map its software onto it, or buy a chip whose architecture was designed around the exact workloads of driving — the perception, fusion, and planning steps inside the large models (大模型) that run a car. Horizon's pitch is the second path: a BPU (brain-processing unit) architecture paired with a software stack, Horizon SuperDrive (HSD), built as an end-to-end "world model" for driving.

For Chinese car brands, the appeal is practical. A domestic supplier means shorter design cycles, local support, and — increasingly important — a supply chain that does not run through export controls on high-end GPUs. Yu has framed this as "widening the boundary" rather than fighting Nvidia head-on: win the car first, then extend the same stack to robots.

That is the real meaning of his Chengdu promise. He is not claiming to beat Nvidia's data-center GPUs globally. He is claiming to win China's in-vehicle high-end autonomous-driving segment — a narrower, very large, and strategically sensitive market.

Honest limitations

This profile draws on company press releases, Horizon's IPO filings, and reporting from Bloomberg, Global Times, and Chinese business outlets. I have not interviewed Yu Kai, and I have deliberately left out his personal life and any speculation about his politics — neither is relevant to his technical contribution.

Two caveats matter for readers. First, the "one in three" figure depends on how you define the market; third-party trackers (such as Gaogong Intelligent Vehicle) measure "compute solutions for self-brand passenger cars," not all vehicles, and methodologies differ. Second, the "15 million cumulative units" claim comes from the company and a single media report; it is a stated production milestone, not an independently audited shipment count. The comparison with Nvidia is specifically about China's automotive AD chips, not the global AI-accelerator market where Nvidia remains dominant.

What readers can do now

  • If you are shopping for a Chinese EV, check whether the driver-assistance system is built on Horizon's Journey (征程) platform. It affects over-the-air upgrade potential and the manufacturer's long-term support path.
  • For engineers and students, Horizon publishes a BPU toolchain and reference designs. Studying them is a concrete way to learn automotive-edge AI, not just cloud AI.
  • Track one number through 2027: whether Horizon actually passes Nvidia in China's high-end in-vehicle autonomous-driving chip share. It is the clearest public scoreboard for China's claim to chip autonomy in cars.

Related coverage

Cite this article

NeuroAI (脑机网), "Yu Kai left Baidu to build self-driving chips nobody wanted — now one in three Chinese smart cars runs his silicon," 15 November 2026, https://neuroai.site/a/na-people-yu-kai-horizon

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Builders” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories