On the evening of August 2, He Xiaopeng, CEO of Xiaopeng Motors, issued a document: Due to family and various reasons, Wu Xinzhou, vice president of autonomous driving of Xiaopeng Motors, will leave the company . At the same time , Wu Xinzhou's next stop was also determined : the highest-level Chinese executive of a well-known company (Nvidia).
In fact, in the past five years, Wu Xinzhou has led Xiaopeng Motors' smart driving research and development team to deliver HNGP (high-speed assistance), CNGP (city assistance) and XNGP (full-scenario intelligent driving assistance) in mass production. In addition to the hardware part, Most of the software is self-developed.
At the same time, Xpeng Motors is also one of the earliest partners of Nvidia in the Chinese market. From the previous generation of Xavier to the Orin platform, it can be said to be one of the successful practitioners of the Nvidia platform smart driving solution.
In addition, in terms of system development, compared with the early visual perception architecture, Xpeng Motors' latest XNet (based on Transformer network and BEV technology) new generation perception architecture has also achieved end-to-end data-driven algorithm iteration.
This means that unless there is a revolutionary technological breakthrough, the existing technical architecture can already support Xpeng Motors to move towards the L3 and L4 stages. Behind this, the requirements for high-quality data (model sales and mileage accumulation) have exceeded the technology itself.
And starting from this year, whether it is Xiaopeng, Ideal or Huawei (Qingjie), the biggest task is how to quickly expand the scale of urban NOA. "Minimize the dependence on high-precision maps, and the high-level intelligent assisted driving capabilities can quickly expand the city."
As Andrej Karpathy, former director of artificial intelligence at Tesla, said after leaving last year: In the past five years, autonomous driving (Autopilot, FSD) has "graduated", from high-speed to urban areas. In his view, scale may make problems easier to solve.
So far, the key to the continuous upgrading of high-end intelligent driving is nothing more than deep learning, computing clusters and high-quality data collection, labeling, and closed-loop training. In Andrej Karpathy's view, the methodology has been clear, so he himself hopes to refocus on AGI (General Artificial Intelligence).
The same is true for Wu Xinzhou.
At present, Xiaopeng Motors is no longer facing technical problems (whether it is R&D investment, technical team size or internal attention, it is already one of the best), but cost control (profit) and sales.
In 2022, because the sales of G9 are lower than expected, the performance of Xiaopeng Motors is not satisfactory. Previously, the company announced its sales target for 2022 to "guarantee 250,000 vehicles and hit 300,000 vehicles." In the end, the annual sales volume was 121,000 vehicles, and the target completion rate was only 48%.
In 2022, Nezha, Ideal, and Weilai will take the top three annual sales of new forces, and Xiaopeng will be squeezed out of the top three. In addition, BYD is even more outstanding in the new energy market, although it is only a standard entry-level L2 in terms of assisted driving.
In this regard, He Xiaopeng once responded, "Price involution is inevitable, and Xiaopeng's strategy is to do a good job of scale, and cost is a very important point." To a large extent, the continuous iteration of smart driving technology, It is precisely one of the largest R&D costs for car companies.
Previously, Xiaopeng Motors had tried to "highlight" the output effect of technical investment through software fees. But soon, the income from this business “disappeared” from the financial report.
However, compared with the current Xiaopeng Motors, for Wu Xinzhou, who already has enough mass production and development experience, its significance to Nvidia may be several orders of magnitude.
Nvidia’s automotive business (chips and software) sales in fiscal year 2021 will be US$566 million, accounting for only about 2% of the company’s total sales; in fiscal year 2022, this figure will increase to US$903 million, but still accounting for only about 3% , There is still a huge gap with the game, data center and other sectors.
But as the helm of the company, Huang Renxun still firmly believes that "automotive will be our next billion-dollar business." And behind this goal, we can't just rely on chip hardware.
Like Qualcomm, Nvidia's automotive business revenue accounts for a small proportion compared with data center and gaming revenue. As the Orin platform begins mass production and delivery in 2022, the automotive business has grown, but the scale effect has not yet appeared.
The reason is that in the past, chip manufacturers provided more hardware development reference designs for downstream customers to help them quickly provide overall software and hardware solutions for automakers. But today, software is becoming the new barrier to entry.
Of course, Nvidia does the same.
For example, the NVIDIA DRIVE™ open source software stack launched by NVIDIA can help developers efficiently build and deploy various applications, including perception, positioning and mapping, planning and control, driver monitoring and natural language processing.
In addition, NVIDIA provides corresponding modular configurations, including NvMedia for sensor input processing, NVIDIA CUDA® libraries for efficient parallel computing, NVIDIA TensorRT™ for real-time AI inference, and other hardware-accessible engines. Developer tools and modules.
To this end, Nvidia even acquired DeepMap, a high-precision map start-up company, to strengthen the market competitiveness of Nvidia Drive by integrating the latter's technical solutions. "They will expand our global mapping business and expand the solutions capabilities of our fully self-driving technology stack."
The real turning point came with the Mercedes-Benz project.
Three years ago, Nvidia and Mercedes-Benz announced that they would cooperate to develop fully autonomous vehicles in the next ten years. The two parties used their respective experience in high-performance computing and high-end car manufacturing to jointly create a new software-defined car.
This cooperation is regarded by Nvidia as "the largest single business model transformation" in the company's development history. Both parties will be able to share revenue from future user purchases of features and subscription services. Behind it, Nvidia provides a full-stack software and hardware solution, and it is jointly developed with Mercedes-Benz.
For this reason, Huang Renxun also specifically emphasized, "In the next few years, it will no longer be a fantasy to sell new cars at cost prices, because profits will mainly come from software."
According to public data, for every $1 sold by Nvidia, hardware, software and other related system providers will get $8 in revenue. This is why the capital market previously gave Tesla such a high valuation.
In Huang Renxun's view, the most important purchase factor for users in the future is software that is constantly iteratively upgraded and continuously enhanced. Whether it is possible to look forward to new software and be satisfied after using it means that the automotive business model will fundamentally change.
And Qualcomm directly adopted the acquisition strategy. "The deep integration of Arriver means that Qualcomm will be able to provide customers with a full-stack, fully integrated hardware and software solution; at this point, it is similar to the solution provided by Mobileye." Industry insiders said.
Some people in the industry said that Huang Renxun recruited Wu Xinzhou this time, and the key reason was the latter's experience in full-stack systems, especially software mass production. For general-purpose chip manufacturers, this is the core competitiveness of a certain industry segment, and it is also a more effective way to develop next-generation products that meet customer needs.
It is reported that at present, due to the lack of full-stack software capabilities of the former in the strategic cooperation project between Nvidia and Mercedes-Benz, coupled with sudden factors such as the epidemic in recent years, the project has been delayed. In addition, Mercedes-Benz is even seeking help from other suppliers for its projects in China.
Another industry insider broke the news that Nvidia is looking for autonomous driving company targets (including in the Chinese market) that can be acquired to make up for the lack of capabilities of related software systems.
And Mercedes-Benz is also strengthening China's local software research and development capabilities. Previously, public data showed that by the end of this year, the size of the Mercedes-Benz China R&D team is expected to reach 2,000, which is nearly double that of 2020. Autonomous driving software development is a key part of it.
For Nvidia, whose market capitalization has exceeded one trillion U.S. dollars, it is obviously not in line with investors’ expectations to just do a little chip business in the automotive industry (competition in the future will also be fierce, and auto companies’ self-developed and even customized chips will become mainstream). Expectation of future value.
Previously, Huang Renxun also proposed the concept of "Software 3.0" (relying on data, algorithms and computing engines), combined with the company's strategic transformation, he believes that "Nvidia hopes to sell more software that only runs on its GPU."
Industry insiders believe that Nvidia saw the opportunity of artificial intelligence many years ago and invested heavily in the development of a complete artificial intelligence stack, including software, services and hardware.
The smart car track is no exception. With the technical cooperation between Xiaopeng Motors and Volkswagen Group, Nvidia has seen opportunities and felt risks.