Multiple Automakers Advance In-House AI Driving Chips to Seize Core Tech and Cost Control

As vehicles become increasingly intelligent, core user experience and competitive differentiation are growing more dependent on chips, algorithms, and operating systems—prompting a shift in the traditional division of labor between automakers and suppliers. NIO, XPeng, and Li Auto have successively launched and mass-produced their self-developed intelligent driving chips. BYD has also unveiled its first 4-nanometer automotive-grade intelligent driving chip, codenamed “Xuanji A3,” with a single-chip computing power of approximately 700 TOPS and a combined total exceeding 2,100 TOPS when three chips are used together. Manufactured on a 4nm process node, the Xuanji A3 is more advanced than NVIDIA’s Orin (7nm) and comparable to NVIDIA’s upcoming Thor (4nm). Previously, BYD primarily focused on developing power semiconductors and control-oriented chips such as IGBTs, SiC devices, and MCUs for use in electric drive systems, motor controllers, and battery management systems. The Xuanji A3 marks BYD’s first dedicated chip for intelligent driving applications. While the chip has completed preliminary validation, BYD has not yet disclosed specific vehicle models that will feature it or shared real-world performance data, indicating that large-scale production and deployment remain some distance away. Industry experts note that automakers are pursuing in-house intelligent driving chip development primarily to achieve technological autonomy, product differentiation, and supply chain security. If AI is viewed as a core competitive advantage, self-developed chips enable deeper co-optimization across hardware, software, and algorithms. Moreover, high sales volumes can help amortize substantial upfront R&D investments, creating cost advantages. For example, NIO claims that its self-developed Shengji chip reduces per-vehicle costs by roughly RMB 10,000 after mass production. Based on projected 2025 sales volumes, this could translate into annual savings exceeding RMB 1.8 billion. However, chip self-development isn’t suitable for all automakers. From project initiation to vehicle integration typically takes over two years—a period during which algorithms and requirements may evolve significantly. Additionally, reliance on external foundries for advanced process nodes presents ongoing challenges in manufacturing and stable mass production. For companies with limited sales scale or insufficient technical expertise, adopting mature third-party solutions remains the more practical choice. Currently, most automakers adopt a hybrid strategy combining external sourcing with in-house development. BYD continues collaborating with Horizon Robotics and NVIDIA to provide computational support for its “Tianshen Eye” intelligent driving system. Similarly, NIO, XPeng, and Li Auto relied extensively on NVIDIA’s Orin platform to validate and mass-produce their advanced driver-assistance systems before launching their own chips. Market data shows that third-party suppliers still hold significant scale advantages. Horizon Robotics forecasts shipments of over 4 million automotive-grade chips in 2025, including approximately 1.8 million mid-to-high-end units supporting urban navigation pilot functions. The industry may soon diverge: a few leading players could pursue full-stack in-house development akin to Apple or Huawei, while the majority will likely continue procuring standardized solutions and focusing on integration and optimization. Ultimately, consumers care more about real-world performance than chip origin. Some XPeng owners have already stated that their decision to purchase higher-trim models hinges on tangible improvements in intelligent driving capabilities—not on the number of chips or whether they are self-developed. Experts emphasize that while high-compute chips are a necessary condition for advanced intelligent driving, they are not sufficient on their own. True value must be validated through holistic vehicle development, mass-production readiness, and user feedback.