<h2>Tesla's AI Moment Seen from RoboTaxi</h2>Tesla's AI Moment. As the top player in AI, Tesla has significantly increased its investment in AI this year to avoid computational limitations, with an estimated total investment in autonomous driving exceeding $10 billion. End-to-end reshaping of FSD, with FSDV12 experiencing exponential growth in mileage, rapidly surpassing 1 billion miles. V12.4 marks even steeper mileage growth, now exceeding 1.3 billion miles. Additionally, Optimus Gen 2 made its debut in Shanghai, equipped with AI derived from the vehicle itself, utilizing visual neural networks and FSD chips to mimic human training. Tesla is ushering in an AI moment.AI Accelerating Tesla's RoboTaxi. Tesla's RoboTaxi is expected to be released in October, with Musk stating, &#34;I have requested significant design changes to the front of the RoboTaxi, and the additional time gained from the delayed release date will allow the company to showcase some other content.&#34; FSD V12.5 will be the most important update before the RoboTaxi release. At Tesla's 2024 shareholder meeting, Musk mentioned that Tesla is currently developing HW5.0, which is expected to be 10 times more powerful than HW4.0 and is projected to be launched in about 18 months, with HW5.0's computational power estimated at 5000 TOPS. AI is accelerating RoboTaxi, with Baidu ApolloADFM, the world's first large-scale autonomous driving model supporting fully driverless operation, enabling operation in a 1500 square kilometer area within 3 months, with a 6-fold increase in area expansion speed.RoboTaxi - Speculation on Tesla FSD's Entry Path into China. In the first quarter of 2024, RoboTaxi provided 826,000 autonomous travel services, a 25% year-on-year increase, with cumulative services exceeding 6 million. Analyzing the timeline of RoboTaxi's entry into cities, a clear pattern emerges: &#34;Intelligent connected network policy -&gt; RoboTaxi enters pilot operation/commercialization -&gt; Application for vehicle-road-cloud integration pilot.&#34; RoboTaxi, from obtaining a license for road testing in autonomous driving demonstration areas to commercial operation, also provides a policy path speculation for Tesla's FSD entry into China.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/b3fccd78-dbc1-421c-b3fc-f16ed7e7ed6a.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/b3fccd78-dbc1-421c-b3fc-f16ed7e7ed6a.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw"/><h2>Tesla's AI Moment: The Top Player in AI</h2>End-to-End Reshaping of FSD. In March 2024, Tesla introduced FSD (Supervised) version 12.3.3. Compared to FSD 11, FSD v12 innovatively adopts end-to-end neural network technology (end-to-end NN), which can better understand and process complex driving environments, reduce driver intervention, and improve the accuracy and automation of autonomous driving. On June 5th, FSD v12.4 began to be pushed to Tesla's employee fleet, and on July 7th, Tesla pushed FSD v12.4.3 to some users. Next,FSD V12.5 will be the most important update before the Robotaxi release.In FSD v12, end-to-end NN is adopted (previously AI perception + software control), where large models only need to input videos for learning without the need for separate coding, making it more flexible and lighter on code. Perception - Decision - Execution is the basic process of FSD. In v11, the perception stage required obtaining surrounding object information through visual/radar methods and identifying/classifying them, while the decision stage relied on pre-written control rules. However, in v12, end-to-end NN technology is used, eliminating the need for identification and classification in the perception stage, as well as the need for pre-written control rules in the decision stage. By inputting a large amount of videos for NN learning, Tesla can discern what needs to be done in different situations. This has allowed Tesla to reduce its code by 100 times in FSD V12, making it more lightweight and flexible, and capable of operating in unfamiliar terrains even without a network connection.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/dc01041c-8cfe-42be-945d-5edc4d482fa7.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/dc01041c-8cfe-42be-945d-5edc4d482fa7.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw"/>Tesla's AI Moment: Autonomous driving investment this year is expected to exceed $10 billion.Supercomputing requires massive capital investment, and Musk expects Tesla to invest over $10 billion in related fields this year. On January 27th this year, Musk tweeted that Dojo's computing power is equivalent to 10,000 Nvidia H100s. To remain competitive in the AI field, Musk believes that billions of dollars need to be invested annually. Musk stated that Tesla's cumulative investment in training computing, data pipelines, and video storage will exceed $10 billion this year.The data gap and computing power requirements form the core barriers of supercomputing. Firstly, the training of any model relies on a massive amount of controllable training data. Since the launch of Tesla FSD, it has accumulated over 1 billion miles of driving data, providing ample training data for Tesla's models. Secondly, in terms of computing power, training large models requires powerful computing power. Tesla may currently have 30,000-35,000 Nvidia H100 chips, and the next generation Gork may require 100,000 H100 chips. If each H100 chip costs $45,000, then the total value of 100,000 H100 chips amounts to $4.5 billion, equivalent to over 30 billion RMB.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/3287f07b-eecf-45b9-8c9e-a4620520bafe.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/3287f07b-eecf-45b9-8c9e-a4620520bafe.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw"/>Tesla initiates an AI computing power war, with domestic leading players following suit in the computing power race.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/a537f35c-8cef-47ea-bd7e-b443cf0c50d7.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/a537f35c-8cef-47ea-bd7e-b443cf0c50d7.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw"/>Tesla FSD: V12 sees exponential growth in mileage, rapidly surpassing 1 billion miles, aiming for tens of billions of kilometers.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/3eb15411-62a7-4a64-9ca7-0d0d7f68edec.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/3eb15411-62a7-4a64-9ca7-0d0d7f68edec.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw"/>Optimus Gen 2: Equipped with AI technology integrated with vehicles, leveraging visual neural networks and FSD chips.With AI technology integrated with vehicles, and leveraging visual neural networks and FSD chips, Optimus humanoid robot has made significant technological breakthroughs. It can now accurately assemble battery units in Tesla's battery factory, thanks to an end-to-end neural network running entirely on the robot's embedded FSD computer. This network achieves high precision operations with limited data inputs such as 2D cameras, hand touch sensors, and force sensors. Additionally, Optimus has demonstrated the ability to autonomously recover from failures, proving its stability and reliability. These enhanced skills are attributed to a large amount of training data collected through human remote operations, providing a solid foundation for Optimus to perform more complex tasks.Testing and Deployment: Optimus has undergone extensive testing in the real working environment of Tesla's factory, with remarkable performance. As testing progresses, the rate of human intervention continues to decrease, indicating that Optimus is becoming more independent in completing tasks. At the same time, Optimus's walking speed and distance are continuously improving, laying the foundation for its broader application scenarios. Tesla plans to first use Optimus to perform useful factory tasks in its own factory by the end of this year, with over 1000 units expected to be in use in factories next year, and considering market release by the end of 2025.The technical highlight of Optimus lies in its dexterous five-fingered hand, which has up to 11 degrees of freedom (planned to increase to 22 degrees of freedom in the future). This hand not only has tactile sensing capabilities but can also withstand a large amount of object interaction without the need for continuous maintenance. In addition, Optimus is equipped with advanced remote operation software, allowing human operators to wear VR glasses and gloves for low-latency real-time control. To achieve efficient data collection and training, Tesla has deployed a large-scale robot team and arranged well-trained human operators for multiple shifts to ensure that the robots are always busy<h2>Tesla Robotaxi: Expected to be Released in October, Continuously Catalyzing</h2>The Tesla Robotaxi is expected to be released in October. Musk previously stated at a shareholder meeting that the Tesla Robotaxi will be a combination of the Uber and Airbnb models. Recently, Tesla has postponed the Robotaxi launch event to October. Musk stated: &#34;I have requested significant design changes to the front of the Robotaxi, and the additional time gained from the delay will give the company the opportunity to showcase some other content.&#34;Tesla FSD: V12.5 will be the most important update before the Robotaxi release.HW4.0: 5x Computing Power Upgrade.The chip will use a more advanced process: The HW4.0 FSD chip uses Samsung's 7nm process, while the HW3.0 uses Samsung's 14nm process. The performance of the new autonomous driving chip will be around 5 times that of the current autonomous driving chip, whether in terms of overall energy consumption/computing power parameters or single-chip computing power (in which case the total computing power of the new platform is likely to reach 500TOPS).• Camera Upgrade: The number of cameras may be reduced from 9 to 8, but the clarity of the cameras will increase from 1.2 million pixels to 5 million pixels. This makes the HW4.0 have stronger perception capabilities than the 3.0, with a maximum detection distance of up to 424 meters, eliminating blind spots and dead angles.• Reserved Radar Interface, but Model Y Currently Not Equipped with Radar: HW 4.0 reserves space for radar equipment, but the Model Y does not come equipped with radar functionality from the factory. This may be related to Tesla's preference for visual solutions, reducing vehicle manufacturing costs, and providing users with some choice.• 66.67% Increase in CPU Core Count, FSD Computer 2 Quantity Increased to 3. In terms of CPU, the HW4.0 has increased from 12 cores in the 3.0 to 20 cores, a 66.67% increase in core count. In terms of computing platform, although the HW4.0 still uses the FSD Computer 2, the quantity has increased to 3.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/19d5eab6-d1cd-4fde-9b6a-16d9a801c3f5.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/19d5eab6-d1cd-4fde-9b6a-16d9a801c3f5.png?"/>HW5.0: AI5, 10x Computing Power Boost.<ul><li>At Tesla's 2024 shareholder meeting, Tesla CEO Elon Musk stated that Tesla is currently developing the fifth-generation autonomous driving hardware (HW5.0), which will be named AI5.0. The design of HW5.0 has been completed, and its performance will be 10 times that of HW4.0. It is expected to be launched in approximately 18 months.</li><li>The performance of HW4.0 is 5 times that of HW3.0, and the computing power of HW5.0 is expected to be 5000 TOPS.</li></ul>AI Accelerates Robotaxi.Baidu Apollo ADFM, the world's first large-scale automatic driving model that supports fully autonomous driving. It supports the operation of 1500 square kilometers in 3 months, with a 6-fold increase in the speed of area expansion.<ul><li>In the past, the research and development cycle of L4 level autonomous driving technology was exceptionally long. Many startups that were once dedicated to this field have adopted a &#34;strategic transformation&#34; strategy and turned to the fierce competition in the L2+ advanced driver assistance market. In the first quarter of 2024, Carrot Run provided 826,000 orders of autonomous driving travel services, a year-on-year increase of 25%, with a cumulative service of over 6 million orders.</li><li>Other Robotaxi players in China, such as Pony.ai, WeRide, AutoX, Didi, etc. Apart from Didi and Antu (AutoX), other companies have intersections with Baidu. It is worth noting that Pony.ai and Carrot Run have significant differences in the control strategy of L4 level autonomous driving vehicles, which may profoundly affect the efficiency of the ratio between remote safety personnel and unmanned vehicles. Specifically, Pony.ai adopts the &#34;remote assist&#34; mode, which places higher demands on the vehicle's autonomous decision-making ability. It does not directly control the vehicle's braking, steering, and acceleration systems, but provides suggestions for the vehicle to make its own judgments and execute operations. Under this mode, the current ratio of people to vehicles has reached 1:5 and is expected to increase to 1:10 within the year. In contrast, Baidu leans more towards the &#34;remote control&#34; approach, where the balance and trade-offs between human-vehicle coordination efficiency and vehicle intelligence level will be a key issue.</li></ul><h2>Analyzing the Landing Path of Tesla's FSD in China from Carrot Run Robotaxi</h2>Tesla's Robotaxi landing path in China can be predicted by referring to Luobo Kuaipao.• Luobo Kuaipao's landing in a city generally follows the following sequence: &#34;autonomous driving (safety driver in the driver's seat) - unmanned driving (safety driver in the front passenger seat) - fully unmanned driving (removing the safety driver from the vehicle)&#34;. Each stage is tested first, then services are provided, ultimately achieving commercial operation (charging fees).• In comparison, Luobo Kuaipao's Beijing business started the earliest and the project progressed most steadily. On August 8, 2022, just three months after starting operations in Wuhan, Luobo Kuaipao successfully obtained a license to charge fees and began providing fully unmanned driving services.• As the first stop for Tesla's FSD road testing in China, Shanghai is likely to become the first city for FSD to enter China. Analogous to Luobo Kuaipao's business timeline in Beijing, Tesla is likely to complete most of its testing in Shanghai before expanding to other cities.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/090acaed-9c94-4bbf-b1f8-a120be9c62da.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/090acaed-9c94-4bbf-b1f8-a120be9c62da.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw"/><h2>NOA, a city without maps, redefines smart cars</h2>Tesla takes the lead in defining and leading the layout of domestic new forces in NOA cities that do not rely on high-precision maps, with the current mass production plan fully shifting to mapless solutions.• Tesla leads the way with high-speed NOA: In October 2016, Tesla was the first to introduce high-speed NOA (Navigate on Autopilot), defining the advanced direction of autonomous driving. Since the end of 2020, Nio, XPeng, and Li Auto have successively pushed high-speed NOA to customers.• NOA redefines smart cars in cities: In October 2020, Tesla first released its FSD BETA test version to early access program testers, supporting the use of NOA on city roads, marking a redefinition of the intelligence level of cars. By the end of 2022, XPeng became the first domestic company to push city NGP to users in Guangzhou. In March 2023, Li Auto announced an all-in strategy for autonomous driving. In May 2023, Aito equipped with Huawei ADS2.0 achieved mapless NOA. In February 2024, XPeng launched the XNGP Unlimited. In July, at the 2024 Intelligent Driving Summer Conference, Li Auto officially announced that mapless NOA will be fully rolled out in 2024. At the same time, Huawei's Qiankun ADS3.0 is about to be implemented.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/15c37b24-1522-4489-96ef-041f0c07421b.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/15c37b24-1522-4489-96ef-041f0c07421b.png/query-aW1hZ2VWaWV3Mi8yL3cvNjQw"/>This article is edited and authored by analyst Liu Hongchen (S1190524010002), source: Pacific Securities, original title: Insights into Tesla's AI Moment from Luobo Kuaipao Robotaxi

IDRV

DRIV

Tesla has significantly increased its AI investment this year, with cumulative investments in autonomous driving expected to exceed $10 billion. The Robotaxi is expected to be launched in October, with Optimus Gen 2 equipped with AI that is derived from the vehicle itself, mimicking human training. Tesla is developing HW5.0, with performance expected to be 10 times that of HW4.0. RoBoTaxi provided 826,000 autonomous travel services in the first quarter of 2024, a year-on-year increase of 25%

Tesla welcomes the AI era