CITIC Securities: FSD V12 pushes forward on a large scale, intelligent driving with ChatGPT is getting closer and closer

According to the financial news app Zhitong Finance, CITIC Securities released a research report stating that recently, Tesla has gradually started to push FSD V12 to North American users on a large scale and has removed the beta name. The new version has a higher performance limit, a more human-like driving style, and faster model convergence speed. The advanced autonomous driving chatGPT is approaching, and the bank expects the end-to-end model to undergo rapid evolution. The bank believes that the application of the end-to-end model will continue to improve the performance of intelligent driving, which is expected to boost consumer willingness to pay and increase Tesla's revenue and profit. At the same time, leading domestic intelligent driving OEMs, intelligent industry chains, and targets in the Tesla industry chain are also expected to benefit.

CITIC Securities' views are as follows:

Recent changes in Tesla: FSD V12 has started large-scale deployment in North America, with scene adaptability and rule control being the highlights.

Tesla's end-to-end FSD V12 was launched in December 2022 and has undergone multiple version iterations. According to IT Home, on March 26th, Musk stated in an internal email that all applicable vehicles in North America will receive a one-month FSD trial. According to Stats APP's statistics, as of April 7th, the proportion of FSD V12.3.3 in various versions has increased to 30.8%. In terms of performance, according to Tesla FSD Tracker data, the average miles between critical takeovers for FSD V12.3 has increased to 386.7 miles, significantly reducing driver takeover incidents.

From social media test drive experiences, user feedback on FSD V12.3.3 has been positive, specifically in: 1) Closer to the industry goal of universal full-scenario use: not relying on lane lines, able to drive steadily in snow, roundabouts, and other scenarios; 2) High degree of human-like vehicle control: able to proactively move forward during parking to let vehicles behind pass, and can also avoid standing water on the road, behaviors that are difficult to achieve through manual coding. According to the company's official website, from March 12th to the present, FSD has undergone 5 version upgrades (from V12.3 to V12.3.4), accelerating product iteration; the company's Robotaxi will also be launched on August 8th this year, and the intelligent driving ChatGPT is gradually approaching.

Potential financial impact: The improvement in FSD performance and large-scale deployment are expected to enhance user willingness to pay, thereby bringing substantial revenue and profit increments to Tesla.

Tesla continues to iterate and upgrade FSD software algorithms, with its outright purchase price increasing. According to the company's official website, as of now, the outright purchase price of FSD in North America is $12,000. In addition to the outright purchase model, Tesla has also been offering a monthly subscription model since 2021, with prices of $99/month (for those who have purchased EAP) or $199/month. On April 13th, Tesla announced on the X platform that the subscription price for FSD has been uniformly adjusted to $99/monthIn terms of accounting treatment, Tesla previously recognized a portion of the FSD price as current revenue, while the remaining portion will be gradually recognized with software updates.

Taking the 2023 financial report as an example, Tesla had $2.91 billion in deferred revenue on the liability side that year, with an increase of $1.20 billion in deferred revenue for the period and recognized revenue of $595 million for the period. In terms of profitability, according to Tesla's 2022 performance data, the gross profit margin corresponding to the deferred revenue realized by FSD can reach 90%. The bank believes that with the improvement of FSD performance and the large-scale promotion and trial use of this feature in North America, the willingness of North American users to pay is expected to increase significantly, thereby bringing substantial revenue and profit growth to Tesla.

Establishment of industry trends: End-to-end models are more holistic and have higher performance limits.

The end-to-end model synthesizes modules such as perception, planning, and control into a large neural network. During operation, the intelligent driving system inputs signals obtained by sensors into the neural network, which can directly output vehicle control commands such as steering, acceleration, and braking. The advantages of the end-to-end model are: 1) Unified training for the entire model, which is more conducive to finding the overall optimal solution; 2) Full information transmission eliminates cumulative errors between modules; 3) Transition from manual coding to data-driven, enhancing the ability to deal with long-tail problems through parameter adjustment. At the same time, this approach does not rely on specific rules, which can enhance the human-like nature of intelligent driving systems. Taking the best paper at CVPR 2023, "Planning-oriented Autonomous Driving" (authors: Yihan Hu, Jiazhi Yang, Li Chen, etc.) as an example, the end-to-end model UniAD mentioned in the paper has improved the multi-object tracking accuracy by 20% compared to the previous best solution, increased lane prediction accuracy by 30%, and reduced prediction errors for motion displacement and planning by 38% and 28% respectively. On the other hand, there are clear boundaries and distinctions between the modules of UniAD, and researchers can conduct white-box analysis of the intermediate results of module outputs. Therefore, this model still has considerable interpretability.

Domestic companies are also expected to actively promote data collection and training algorithm construction, pushing forward the training and mass production deployment of end-to-end models.

In end-to-end models, the data is less targeted, requiring a larger amount of data and training power to improve performance. Currently, Tesla is significantly ahead in these areas. In terms of data reserves, according to the company's official website, as of April 6 this year, Tesla's FSD Beta has accumulated 1 billion miles of driving, an increase of over 25% from the nearly 800 million miles disclosed at the company's performance meeting in January. In terms of training power, according to Tesla's official website, as of August 2023, Tesla has a computing power scale of 10EFLOPS; Tesla plans to increase training power to 100EFLOPS by October this yearOn the Chinese automaker front, XPeng Motors announced at the national autonomous driving launch event in January 2024 that the company will strive to achieve the goal of passive takeover frequency <1 time per thousand kilometers in core areas. The end-to-end model will be fully implemented in the next step. According to various company official websites, Nio plans to launch end-to-end active safety features in the first half of this year, while Li Auto's end-to-end intelligent driving model will also be launched in 2024. In terms of third-party suppliers, according to the Yuanrong Qixing official website, Yuanrong Qixing has successfully adapted the end-to-end model to mass-produced vehicles, and this batch of mass-produced vehicles will be launched in the consumer market this year. Overall, Chinese companies are also continuously following up on the end-to-end model, with the opportunity to follow in Tesla's footsteps and achieve a leap in autonomous driving capabilities.

Investment Strategy:

Tesla has started mass pushing the end-to-end FSD V12 to North American users, which has a higher performance limit, a more human-like driving style, and faster model convergence speed. The high-level autonomous driving chatGPT is gradually approaching, and the industry expects the end-to-end model to start rapid evolution. Chinese companies such as XPeng, Huawei, and Li Auto are actively following up on the end-to-end model and are committed to advancing data collection and training algorithm construction. The application of the end-to-end model promotes the continuous improvement of intelligent driving performance, which is expected to drive consumer willingness to pay and increase Tesla's revenue and profit. At the same time, domestic intelligent industry chains and Tesla industry chain targets are also expected to benefit.

In terms of the intelligent industry chain, the industry recommends: 1) Complete vehicle: High-level autonomous driving is expected to become an additional source of income for car companies; the end-to-end large model has high barriers, and full-stack self-developed models from host factories iterate quickly, potentially pulling ahead of competitors in intelligence; 2) Domain controller: Computing power is the core foundation of intelligent driving, closely cooperating with leading chip manufacturers, and products with leading tier-1 industry chain positioning effects; 3) Intelligent chassis: High functional safety requirements and high barriers.

Risk Factors: Risks of technological and product iteration; product promotion speed lower than expected; risks of declining sales in the automotive industry; tightening policies in the intelligent driving industry; occurrence of serious accidents in intelligent driving