As autonomous driving technology accelerates its market entry, the full-stack in-house R&D model has emerged as a consensus in the advanced intelligent driving technology sector. This model entails OEMs developing comprehensively from underlying hardware to software algorithms and system integration, eschewing reliance on third-party suppliers or the Tier 0.5 model (a collaboration between automakers and suppliers). The core advantages of this model lie in its comprehensive technological control and deep data mining capabilities, facilitating rapid iteration, precise optimization, and robust market competitiveness. However, this approach also brings high R&D costs, complex technology integration demands, and significant resource investment pressures, necessitating strong R&D capabilities and efficient collaboration among OEMs.

Advantages of Full-Stack In-House R&D

One of the greatest advantages of the full-stack in-house R&D model is its comprehensive control over technology. Under this model, OEMs can oversee the entire technical process, from underlying chips to upper-level algorithms, enabling end-to-end deep optimization. This integration mitigates compatibility issues that arise from the separation of software and hardware in traditional modular development. For instance, Li Auto's full-stack in-house 4D One Model architecture integrates visual perception, decision-making planning, and control, making urban NOA functional without high-precision maps and capable of autonomous driving in complex road conditions using real-time perception data alone. This level of technological control not only enhances system response speed and decision-making efficiency but also allows products to adapt more flexibly to diverse driving scenarios.

Furthermore, the full-stack in-house R&D model allows OEMs to establish a complete data closed-loop system. Data is crucial for advanced intelligent driving technology, yet the third-party model often struggles to establish an efficient data cycle due to decentralized data ownership. In contrast, full-stack in-house R&D collects vast amounts of real-driving data through its own vehicle models and applies it directly to model training and iterative optimization. Tesla, leveraging its extensive fleet network, has amassed over 2 billion miles of real-driving data, significantly enhancing the performance and generalization ability of its end-to-end models through a fully autonomous data closed-loop system. Similarly, Huawei's ADS 3.0 system achieves rapid iteration with its in-house data closed loop, updating its model every five days on average, continually improving intelligent driving performance in complex scenarios.

Additionally, the full-stack in-house R&D model fosters deep collaborative optimization between algorithms and computing power. Under the traditional model, OEMs often rely on third-party suppliers' algorithm solutions, which are typically designed for general scenarios and difficult to fine-tune for specific vehicle models or user needs. The full-stack in-house R&D model, however, achieves more efficient computing power utilization and performance output through collaborative design of independently developed algorithms and proprietary hardware. Xpeng Motors' in-house XNGP system, for example, employs an end-to-end large model architecture based on BEV+Transformer, seamlessly integrating visual perception, dynamic planning, and decision-making control. Coupled with its dedicated domain controller, it significantly reduces redundant consumption of computing resources, enabling efficient nationwide deployment of the urban NOA function.

Disadvantages of Full-Stack In-House R&D

Despite its technical advantages, the full-stack in-house R&D model faces numerous technical challenges and resource pressures. This model necessitates OEMs' mastery of the entire technology chain, from chip design to advanced algorithm development, placing significant demands on the size and depth of the R&D team. Huawei's Intelligent Automotive Solutions BU currently boasts an R&D team of over 7,000 people and has invested over RMB 30 billion in intelligent driving R&D. Such substantial resource investment is unfeasible for automakers with limited R&D capabilities.

Moreover, the in-house R&D model demands efficient internal collaboration. The development of intelligent driving technology involves close cooperation among multiple departments, such as sensors, smart cabins, and chassis control, and the collaboration efficiency within OEMs often dictates the speed of technology integration and product launch.

Furthermore, the success of the in-house R&D model hinges on continuous computing power improvements. Training end-to-end models requires extensive cloud computing support. Tesla's supercomputing center boasts 100 EFLOPS of computing power, and Li Auto increased its cloud computing power to 8 EFLOPS by the end of 2024. Deploying and operating such large-scale computing power requires substantial capital investment, which is beyond the reach of small and medium-sized automakers.

Conclusion

The full-stack in-house R&D model offers significant technical advantages in the advanced intelligent driving technology field, including comprehensive control, data closed loop, and collaborative optimization of algorithms and hardware. This enables OEMs with full-stack in-house R&D to maintain a leading technological position in the fiercely competitive market. However, this model also poses severe challenges to automakers in terms of technology accumulation, team size, and resource allocation. In the future, as intelligent driving technology matures and the industrial ecosystem optimizes, OEMs that continuously break through technical bottlenecks under the full-stack in-house R&D model are poised to gain more initiative in the advanced intelligent driving market.

-- END --