Five years ago, capital markets were abuzz with excitement as new forces clashed with established players. Five years later, the end-to-end ecosystem thrives, with autonomous driving and cockpits undergoing rapid transformation, fueled by undiminished enthusiasm.

Looking back and forward, this is a grand narrative unique to China's autonomous driving industry.

Author | Nianqiu

Editor | Piye

Produced by | Industry Entrepreneur

April 2012, from Silicon Valley to Beijing.

Yu Kai, a renowned figure in machine learning, boldly left NEC Laboratories America and ventured from the forefront of AI theory into its practical applications.

Scientists entering the industrial world is not uncommon. Two years later, Han Xu, a tenured professor at the University of Missouri, also made the transition from academia to industry.

The influx of talent and capital made Baidu the ideal incubator for autonomous driving. Joining Yu Kai and Han Xu were two veteran Google engineers, Peng Jun and Lou Tiancheng.

In 2013, Baidu's autonomous driving project took off, gathering China's top software engineers to propel the rapid advancement of the industry over the past decade.

Perhaps even Baidu didn't anticipate that, besides building the Apollo platform and serving diverse clients in the autonomous driving industry, it would also sow the seeds of hope in this competitive landscape, akin to China's "Whampoa Military Academy" for autonomous driving.

In 2015, Yu Kai left Baidu, and Horizon Robotics emerged as China's first company to independently develop AI chips and achieve mass production.

"To truly democratize AI and enhance its efficiency, relying solely on software is insufficient. We must be more aggressive and design dedicated chips for AI," Yu Kai explained his vision when founding Horizon Robotics. After three years of deepening their roots in autonomous driving, scientists gained a deeper understanding of theory and application.

Practitioners like Yu Kai chose to accelerate into the entrepreneurial race—Peng Jun co-founded Pony.ai with renowned Lou Tiancheng in 2016, while Han Xu founded WeRide in 2017, ushering in a flourishing era for autonomous driving.

Today, diverse paths have converged. Horizon Robotics, Pony.ai, and WeRide have all reached the IPO stage.

As the bell tolls, amidst the festivities, lies China's transformation story in autonomous driving.

I. 2024: Autonomous Driving 'Finally Bears Fruit'

Let's start with the present. In the first half of 2024, over 60% of new vehicles delivered were equipped with ADAS.

This signifies autonomous driving's penetration into the automotive industry, no longer a mere theoretical construct. The explosive growth of assisted intelligent driving over the past five years has made it a non-negligible dimension in new car evaluations, with prices becoming increasingly affordable amidst domestic auto industry competition.

A year ago, from July 2023, WeRide, AutoX, and Pony.ai's Robotaxis graced the streets of Guangzhou, Shanghai, Wuhan, and other major cities, bringing self-driving taxis from fiction to reality.

In the eyes of the market, autonomous driving has finally crossed a significant milestone—moving from product validation (0 to 1) to expansion and transformation (1 to 10).

The relentless efforts of leading players have fueled autonomous driving's rapid development over the past five years, but the hardships are seldom shared. Automotive-grade AI chips represent the first hurdle for China's autonomous driving enterprises.

From a training perspective, autonomous driving technology's development and testing are computationally intensive. Complex road environments and multi-sensor fusion necessitate exponentially growing computational demands in deep learning. For actual deployment, tens of thousands of training and validation sessions are required. L3 autonomy demands 20-30 TOPS, L4 requires over 200 TOPS, and L5 exceeds 2,000 TOPS, posing a significant challenge to chip specifications at the time.

Horizon Robotics was born to address this issue and has navigated this challenging terrain for a long time.

"The second half of 2015 was a funding winter, yet Horizon secured a substantial investment. I joked with my team about how to spend it. Later, once our chip team started work, we realized it wasn't nearly enough," Yu Kai recalled to the media during Horizon's early AI chip days. Little did he anticipate that each tape-out of automotive AI chips would exceed USD 50 million in R&D investment.

Horizon initially wavered between the automotive and IoT industries, focusing solely on automotive over a year later, making automotive AI chips its sole target. Amid economic pressures and internal turmoil, Horizon's first automotive-grade AI chip, Journey 2, arrived in 2019 after multiple funding rounds. Equipped with BPU2.0, a high-performance computing architecture independently developed by Horizon, it offers equivalent computing power of 4 TOPS at a typical power consumption of just 2W. While still lagging behind NVIDIA's Orin at the time, it offered hope to domestic new energy vehicle makers in deploying ADAS.

Tesla's FSD chip inadvertently accelerated Horizon's progress.

When Tesla's large screens adorned the cramped cabins of consumer vehicles, people might recall Steve Jobs' words at the iPhone 4s launch: "Why do we need a keyboard?" With Nokia's downfall at Apple's hands, leading automakers had no time to ponder the rationale of intelligent systems. Keeping pace became their paramount goal. Even rising stars like Horizon were inadvertently "bet on" by domestic new energy vehicle makers.

The 2020 collaboration between LIXIANG and Horizon garnered attention. While XPENG and NIO chose NVIDIA's Orin chip, LIXIANG placed its bet on Horizon. Given the timeline and their respective business conditions, this was almost a gamble for both—a "new brand + new chip" combination that neither LIXIANG nor Horizon could afford to lose.

In retrospect, the outcome is clear. Journey series' performance in LIXIANG models validated domestic AI chips' computing power in intelligent assisted driving, rivaling that of leading enterprises. Subsequent market moves confirmed this, with XPENG and NIO subsequently pouring significant resources into in-house chip development.

On July 27, 2024, NIO unveiled its 'Shenji NX9031' chip, boasting 50 billion transistors and claiming to be the world's first 5nm intelligent driving chip, marking NIO's answer to autonomous driving scenarios. A month later, XPENG swiftly followed suit. After four years of dormancy, the 'XPENG Turing' chip successfully taped out, covering AI-powered cars, robots, and flying cars. This L4 autonomous driving chip boasts a 40-core processor, two independent image signal processors for driving perception and user-perceptible images, and two neural processing units for neural network data processing, capable of running 30B large models.

LIXIANG adopted a two-pronged approach, collaborating with Horizon while pursuing in-house AI chip development. Currently, the Journey 5 chip in LIXIANG L9 Pro supports industry-standard autonomous driving algorithms (BEV) and high-speed NOA. Meanwhile, LIXIANG has invested heavily in independent R&D of intelligent driving SoC chips. According to insiders, this chip, named 'Shu Ma Ke,' embodies LIXIANG's deep research in Chiplet and RISC-V technologies and is expected to debut by year-end.

The breakthroughs of 'WEI XIAO LI' in AI chips epitomize China's new energy vehicle makers' strides in autonomous driving. China's autonomous driving is emerging from its predicament at the foundational level.

Beyond AI chips, from sensors to perception algorithms, high-precision mapping services to cloud mapping vendors, the autonomous driving industry's supply chain is gradually being streamlined, with costs decreasing, paving the way for autonomous driving services to reach the masses.

Upstream and downstream enterprises are navigating uncharted waters together, with midstream players enduring relentless capital burn to achieve great escapes, fueling the autonomous driving industry's rapid advancement over the past five years and culminating in 2024 IPOs.

A decade ago, mobile phones supplanted PCs as ubiquitous smart devices. In 2024, autonomous driving steps down from its pedestal, shedding its mystique to become an accessible emerging technology. Today, a new protagonist graces the stage—autonomous driving.

II. 2000 Days: Visible Forks, Hidden Intersections

"Achieving autonomous driving is akin to climbing Mount Everest. We have at least two routes—the southern and northern slopes—with pioneers and followers alike. Though the summit is singular, climbers' experiences and tales along the way are diverse and captivating," said Zhang Yaqin, academician of the Chinese Academy of Engineering and Dean of Tsinghua Intelligent Research Institute, during a 2023 speech describing the current state of autonomous driving.

The challenges of climbing Everest do not deter the brave but inspire autonomous driving players to explore diverse solutions.

The past five years have witnessed an unending battle between lidar and vision-based algorithms, with their safety and commercial viability constantly scrutinized.

The lidar approach, with a lower difficulty coefficient, transforms autonomous driving challenges into cost issues. By integrating high-precision maps with real-time distance and speed calculations, lidar combines prior and real-time information to assess driving conditions.

In such solutions, high-precision maps of cities are preloaded, recording static building shapes and distances. Lidar or millimeter-wave radar generates real-time 3D point cloud data, capturing dynamic information. Simultaneous Localization and Mapping (SLAM) technology constructs environmental maps, providing planning and decision-making information for vehicles. As high-precision maps provide most prior information, this solution is deemed safer and more reliable, becoming the industry's preferred choice. Waymo in California and AutoX in Shanghai Pudong adopt this approach.

As domestic prices drop, lidar solutions are gaining momentum.

Leading map providers like Gaode Maps, Tencent Maps, Baidu Maps, and others occupy over 80% of the high-precision mapping market, serving multiple top automakers. As the domestic self-driving market surpasses RMB 10 billion, high-precision map costs are decreasing, with leading vendors offering prices in the hundreds of yuan per vehicle, with costs for map collection, creation, and maintenance declining annually.

Price wars have benefited lidar solutions, but industrial upgrades are not without hiccups. In 2021, several car accidents involving various brands eroded public trust in lidar's autonomous driving safety.

On August 12, 2021, a NIO ES8 was involved in an accident on the Shenhai Expressway while its NOP pilot assist function was engaged. The same day, an XPENG G3, with ACC activated, rear-ended a stationary vehicle at 70 km/h.

These black swan events chilled autonomous driving financing. In 2022, there were 125 autonomous driving investment events involving over RMB 20.5 billion. While the total number of events matched 2021, disclosed funding fell short of one-third of 2021's level. As accidents piled up, capital markets grew wary of autonomous driving's future.

However, safety is but the superficial sugar coating; the core issue lies in profitability, prompting investors to halt funding.

Overall, lidar-centric product industrialization progresses faster than vision-based algorithms, but commercialization remains elusive. Waymo, the pioneer in Robotaxi operations in California, reported Q2 2024 revenue of USD 365 million, up from USD 285 million YoY, but losses widened to USD 1.13 billion from USD 813 million YoY. This suggests that even as lidar industry players continually compress costs, "burning money" will persist.

Against this backdrop, vision-based autonomous driving emerges as a new arena for tech giants.

Vision-based algorithms, unlike lidar, resemble a "rich man's game." Abandoning high-precision maps' prior information and relying solely on cameras and sensors for comprehensive environmental perception poses significant challenges for vision algorithms. This path necessitates a decade or more of funding, attracting top talent and equipment, with technical complexity surpassing lidar.

Amid financial pressures, only well-capitalized giants can remain in this race. Domestic heavyweights like Baidu and Huawei, along with overseas player Tesla, are leading the charge in vision-based autonomous driving.

In the long run, if vision-based algorithms achieve actual deployment, their lower maintenance costs could facilitate quicker commercialization than lidar-centric products.

However, Tesla's 9% stock price drop on October 10, 2024, reveals capital's dissatisfaction with Tesla's performance. Fundamentally, investors seek clear business models and detailed technical specifics, values that cannot be conveyed at a party. This demand extends beyond Tesla to every autonomous driving player.

As described by Academician Zhang Yaqin, autonomous driving players on different paths converge toward the same summit in 2024.

This summit represents robust autonomous driving products and viable commercialization models.

A decade of heavy investment and five years of rapid growth; who will be the first to scale this peak, transforming relentless capital burn into profitability, will reign supreme.

III. The Long-Term Vision of End-to-End Autonomous Driving and the 'AI Agent' Node

Two decades ago, autonomous driving's steady progress resembled a marathon. Today, five years of rapid growth do not signify the finish line.

GPT's emergence Endowed with AI " entity " form ， Provided direction for natural language processing ， Make big models a cross industry productivity tool 。

'End-to-end' is quietly emerging as the new consensus for autonomous driving's new paradigm.

Past modular deployments focused on perception, prediction, planning, and control. For different industry players, excelling in one or even a subset of these areas, linked through protocols and interfaces with other modules, yielded mature, deployable autonomous driving solutions. This industrial upgrade path is achievable and enhances technical iteration efficiency.

However, modularity's limitations were overlooked in autonomous driving's early stages. Despite information loss in subtasks, local optimal solutions led to cumulative errors, an unsolvable systemic issue within existing architectures.

The AI Agent revolution marks a turning point.

AI researchers realize that data volume and computational power enhance systems qualitatively, with large models exceeding expectations. In autonomous driving's industrial transformation, module integration fosters small 'end-to-end' tasks, outperforming pure modular solutions. For instance, replacing perception modules with BEV combined with Transformer introduces more data, reduces engineering effort, and achieves superior performance.

What is the charm of "end-to-end"?

To achieve Level 4 autonomous driving, it is unrealistic to reason and process all possible scenarios. However, if perception-prediction-decision-control is considered as a complete module, using sensors to obtain raw data and directly outputting vehicle driving actions, the so-called step-by-step reasoning is not necessary. With a large number of data samples, the entire neural network is data-driven, with operators calculating various situations, and it also has better generalization without data loss.

Specifically, tens of thousands of lines of code in the past only require a few hundred lines with better results under the support of a large model.

This product architecture has reached a consensus within the autonomous driving industry and has become a template for many AI products. For example, industries such as robotics and human-computer interaction currently recognize the end-to-end processing model, which is the most popular technical direction at present.

Building such an autonomous driving solution also frees autonomous driving companies from the constraints of delivery forms. Whether it's Robotaxi or autonomous driving solutions for civilian vehicles, under the "end-to-end" technical model, as long as sufficient real data is collected, the entire link can be run through, enabling delivery in various scenarios.

This is also the core step for domestic automakers this year - XPeng released end-to-end large models for autonomous driving in 2024 - XNet perception neural network, XPlanner planning control large model, and XBrain large prophecy model. They are respectively responsible for perception, planning, and decision-making, working together to handle complex scenarios in autonomous driving. In addition to XPeng, Huawei, SenseTime, and DeepRoute.ai have also submitted their answers to the end-to-end autonomous driving solutions.

However, reviewing the current end-to-end solutions, there are still traces of modularization lurking - XPeng's tripartite collaboration does not realize the original intention of end-to-end integration, and SenseTime's theoretical model lacks the support of real data. Such solutions are closer to domestic manufacturers' stance on end-to-end applications, and there is still a gap from services that can be implemented.

On the other hand, the biggest pain point of end-to-end is the amount of data.

Essentially, the end-to-end large model replaces engineering difficulty with data and computing power. If computing power can be solved with funds, the amount of data requires actual accumulation. Building a large model for various autonomous driving scenarios requires an enormous amount of data. Accumulating sufficient data will become the next challenge for autonomous driving manufacturers exploring end-to-end solutions.

Reviewing the growth of the entire autonomous driving industry, after the intersection, a mature commercial monetization model needs to be combined with long-termism to demonstrate the universal applicability of autonomous driving solutions and achieve continuous progress in autonomous driving technology.

In 2023, Hesai Technology and Black Sesame Technology, both upstream and midstream players in autonomous driving, went public consecutively. Now, the baton of going public has come to the midstream and downstream players in the industry, Horizon Robotics, Pony.ai, and WeRide, marking another important turning point in the autonomous driving industry.

Five years ago, capital players, new forces, and established manufacturers competed fiercely. Five years later, end-to-end solutions are booming, and autonomous driving and cockpits are accelerating their reconstruction, with enthusiasm still intact.

Looking back on the past and looking forward to the future, this is a grand industrial story unique to China's autonomous driving industry.