Author | Liu Jiayi AI Coding is turning an extremely "sci-fi" concept into the daily routine of smart vehicle R&D. Imagine this: an idea thrown out during a product brainstorming session the previous day is already running on a car owner's infotainment system by the next morning. This almost "crazy" iteration efficiency once existed only in the imagination of engineers. But now, as AI Coding begins to be systematically embedded into the underlying R&D pipeline, this scenario is about to become a reality. Over the past few years, the practical application of "AI defining vehicles" in the automotive industry has focused more on the functional level—autonomous driving, smart cabins, and chips—all essentially aimed at making cars smarter. Now, AI is directly entering the production layer. AI Coding has taken over core tasks in the development pipeline. What is being restructured is no longer just the product competitiveness of the car, but the very way cars are built. Iteration speed, R&D efficiency, and organizational responsiveness are evolving from management priorities into a new battleground for deep AI intervention. The most definitive outcome brought by the deep application of AI Coding is precisely: improving quality and efficiency. In a sense, this is similar to the transformation of manufacturing by the Ford assembly line a century ago. The assembly line reshaped industrial production efficiency, while AI Coding is reshaping automotive software R&D efficiency. The former determines how many cars an automaker can build, while the latter determines how quickly an automaker can iterate and roll out the next-generation smart vehicles. Some leading players have already taken the lead. Tesla has fully integrated its AI coding toolchain; domestically, whether it is automakers like Geely and SAIC, or autonomous driving companies like Horizon Robotics, they have all begun to leverage AI coding tools such as Alibaba Cloud Qoder to embed AI Coding into their core R&D pipelines. This most hidden yet most potentially game-changing war in the smart vehicle industry has already begun. Whoever first restructures the R&D paradigm through AI Coding will gain the initiative in defining the next-generation smart vehicles faster.
01 AI Productivity Must Enter the Arena In the smart vehicle industry, at least three forces are simultaneously pushing "AI engineers" to the forefront. First is the exponential explosion in code volume. Smart vehicles are inherently complex systems, spanning autonomous driving, smart cabins, vehicle control, mapping, cloud services, and data closed loops... involving more than a dozen technical fields behind the scenes. Automakers are continuously raising the stakes: autonomous driving is competing on foundation models, smart cabins are witnessing a wave of AI Agents being integrated into vehicles, and full-scenario intelligence integrating cabin and driving is gradually becoming an industry consensus. Every innovation in smart vehicles essentially continues to push up software complexity. Today, the code volume of a single smart vehicle has ballooned to 600 million lines, equivalent to ten times the code volume of the Windows operating system, and it continues to grow rapidly. When software complexity exceeds the upper limit of organizational management capabilities, adding more manpower alone instead yields diminishing marginal returns. This means that the traditional automakers' tactic of relying on a massive army of R&D engineers has reached its end. The second force comes from the increasingly frequent update cadence. In the traditional automotive industry, a major version upgrade for a car after its launch might take years. But for smart vehicles, OTA (Over-the-Air) updates have been compressed from "monthly" to "weekly". According to incomplete statistics, Tesla has maintained almost one OTA update per month over the past three years, and during the FSD Beta phase, it once achieved version iterations every 5 to 10 days. Autonomous driving strategies, energy consumption control, cabin Agents, chassis coordination... more and more core capabilities are continuously evolving through OTA. In a sense, today's smart vehicle R&D is increasingly resembling the "high-intensity operations" of internet companies. However, smart vehicles demand "both and"—they require the iteration speed of the internet, yet cannot tolerate the errors typical of internet software. If internet software has a bug, a simple debug can fix it. But a defect in automotive R&D code could be linked to functional failure or even safety risks. Third, there is the bottleneck of traditional software engineering itself. At the Alibaba Cloud Summit, Qiang Li, General Manager of the AI Automotive Industry Department of the Public Cloud Business Unit of Alibaba Cloud Intelligence Group, proposed a metaphor: under the traditional programming model, code is solid. Engineers write it line by line and change it version by version, just like concrete—once poured, it sets, and making changes after it sets requires "massive reconstruction". Consequently, reinventing the wheel, difficulties in cross-language collaboration, low debugging efficiency, and the challenge of unifying code standards—these common chronic issues are further amplified in the automotive industry. Take autonomous driving companies as an example: on a single data production line, multiple languages such as Python, C++, Go, and Shell run in parallel. Complex pipelines overlapping with multi-person collaboration cause engineering challenges to climb exponentially. But changes are happening. Qiang Li observed that code is showing a trend of "liquefaction". Developers describe their intentions in natural language, and AI generates, runs, and adjusts in real time. Code becomes a dynamic process "flowing in a conversation". AI Coding is precisely the key breakthrough to break the deadlock. It is moving from "assisting in writing code" to "Agentic Coding". Boris Cherny, founder of Claude Code, once stated that AI is already capable of completing 100% of code writing, and "the problem of programming itself is being solved". Agentic Coding is not just icing on the cake; it is the necessary infrastructure to support the explosive growth of software scale. First-mover players are already deploying at scale—Tesla has fully integrated its AI coding toolchain; Geely Research Institute has over 2,300 developers applying AI Coding; and internally at Horizon Robotics, all 1,200 engineers are deeply integrated into the AI R&D system.
02 AI Coding: Rewriting the Automotive R&D Paradigm It is worth noting that there are currently quite a few choices for AI coding tool systems in the automotive industry. While these tools can boost individual developer productivity, they struggle to address the entire pipeline from requirement proposal to deployment and delivery, let alone meet the stringent engineering, safety, and compliance standards of the automotive industry. What is truly needed to run through the core R&D pipeline of automakers is a complete capability foundation spanning from chips to models, and from cloud to edge. At the Alibaba Cloud Summit, Alibaba Cloud announced a major reconstruction of its AI infrastructure—a comprehensive upgrade from underlying chips, models, and inference platforms to Agentic Cloud, building a true "super R&D factory" for the AI Agent era. Gu-rong Cui, Head of the Qoder Business Development Department, introduced that Alibaba Cloud's AI coding products provide the most complete global top-tier models, enterprise-level knowledge engines, and multi-agent capabilities to help automakers with their AI transformation. Currently, the Qoder family already owns six products: Qoder IDE, Qoder CLI, Qoder JetBrains plugin, Qoder Mobile, QoderWork, and QoderWake. More importantly, Qoder is not an isolated tool. What Alibaba Cloud provides is a complete set of runtime environments tailored for automotive R&D scenarios: a lightweight and efficient sandbox execution environment, multi-agent collaboration capabilities, cross-task memory capabilities, and an intelligent operations and maintenance system covering the entire R&D pipeline. This means that AI Coding is crossing the inflection point of single-point efficiency improvement and moving towards full-pipeline reconstruction. A typical smart vehicle R&D pipeline is often an interlocking assembly line, going through requirement proposal, product and technical planning, code development, and testing verification processes, before final deployment and delivery. Most people assume that AI Coding only plays a role in the "writing code" stage. On one hand, it covers code generation and completion: the tool rapidly generates code based on task requirements, automatically adds comments for complex logic, clarifies intent and improves readability, cutting maintenance costs at the source. On the other hand, it supports testing and troubleshooting: automatically intercepting defects online, providing repair suggestions, and even directly completing error fixes. Take Horizon Robotics as an example. From data production and governance, scenario mining, to R&D evaluation and closed-loop processes, Qoder is deeply embedded in the entire workflow. A typical scenario is performance optimization. In the past, reducing the CPU usage of a certain module relied on engineers manually analyzing bottlenecks, organizing multiple rounds of reviews, and repeatedly verifying, which resulted in long cycles and heavy pipelines. Today, Qoder can automatically identify bottlenecks and complete verification, freeing manpower from tedious troubleshooting. Zibing Zhao, Senior Expert of the Autonomous Driving Data Production Line at Horizon Robotics, stated that in the past, R&D engineers spent a lot of time on "from scratch" programming. Now that Qoder participates in most of the work, engineers can step back from repetitive building and focus on researching more cutting-edge algorithms and more advanced engineering practices. Data also confirms this: after the deep application of AI Coding, the daily per capita AI call volume on Horizon Robotics' autonomous driving data production line has significantly increased, with automatic code generation and adoption rates exceeding 40%, and comprehensive R&D efficiency improving by over 30%. Another typical case comes from Geely. Currently, Alibaba Cloud's AI coding products have been applied in Geely's core technical fields such as smart cabins and autonomous driving. Among them, the penetration rate in smart cabins is the highest. Weiwu Yang, AI Product Expert at Geely AI Center, revealed that in the first phase of applying AI Coding to Geely's 2,300 R&D personnel, they have already achieved an automatic code generation rate of 60% and a 30% improvement in code development efficiency. Automakers such as SAIC-GM and NIO have also applied Alibaba's intelligent coding products to their full-pipeline R&D. Multiple automakers have reported to reporters that there is a strong positive correlation between user activity and efficiency improvement results. This means that the value release of AI programming tools is not linear, but exhibits a "flywheel effect" where "the more you use it, the better it gets". However, as tools are used more deeply, the yardstick for measuring efficiency improvements may not keep pace. Jian Huo, General Manager of the AI Automotive Industry Solutions Department of the Public Cloud Business Unit of Alibaba Cloud Intelligence Group, reminded that currently many automakers use the "AI code generation rate" as a core assessment metric. However, since the coding stage only accounts for 20%-30% of the end-to-end software development process, and the commented code, unit tests, and glue code that AI is best at generating are not strongly positively correlated with business value, simply pursuing the code generation rate cannot truly reflect efficiency improvement results. Jian Huo suggests weakening this metric and instead focusing on three core dimensions: first, the per capita volume of effective code that enters the production system, is weighted by complexity, and ultimately delivers business value; second, the defect rate per thousand lines of code; and third, the overall compression margin of the end-to-end requirement delivery cycle. But a more urgent issue than assessment metrics is: can code written by AI be deployed in vehicles? Weiwu Yang, AI Product Expert at the AI Center of Geely Research Institute, and Jian Huo, General Manager of the AI Automotive Industry Solutions Department of the Public Cloud Business Unit of Alibaba Cloud Intelligence Group, both agree that automakers should not wait for a perfect solution, but rather "seize the speed dividend first, and then tackle the safety barriers", which means implementing it in layers:
Phase 1: Start with the intelligence domain. Businesses such as connected vehicles, smart cabins, and the perception layer of autonomous driving iterate quickly, use application-layer code, and teams are naturally receptive to AI collaboration. This is the entry point for teams to quickly form muscle memory for AI collaboration.
Phase 2: Pilot low-risk ASIL-A/B modules. Such as headlights, wipers, and windows, which have entered the ISO 26262 system but have relatively lower safety levels. This phase requires establishing a standardized process of "AI generation + static analysis + formal verification".
Phase 3: For ASIL-C/D control code that concerns driving safety, the red line of "human-led + AI-assisted + mandatory review" must be adhered to. This tiered strategy is about finding a practical balance point between efficiency and safety. To ensure the engineering safety of this tiered and scaled implementation, Qoder Enterprise Edition has built a five-layer defense-in-depth architecture: from network boundary security and identity and access control security, to runtime and AI security and data storage security, and finally to supply chain and ecosystem security, forming a complete security barrier. Each layer is designed for specific threat models in AI Agent scenarios, ensuring that a single-point breakthrough will not lead to systemic security failure. This is just the beginning. Geely is accelerating its transformation into an "AI Native organization". Weiwu Yang shared that in the past, a new requirement for a smart vehicle took at least half a year from planning to implementation—half a year is enough to miss a complete product cycle. In the future, this process can be compressed to the daily level. Requirement understanding, code generation, testing verification, and deployment go live almost seamlessly in an AI-driven pipeline. This also means that the work content and role definitions of developers are being systematically reconstructed. The core capability of product managers becomes the precise control of requirement descriptions; developers transform from the primary producers of code to supervisors and architects of AI; test engineers need to evolve from executors to makers of AI testing rules. Zooming out, the AI Coding paradigm will further reshape organizational structures. Jian Huo believes that the structure of R&D teams will evolve from the professional division of labor of dozens of people in the past to a super combat unit of "a few people + enough Agents".
03 Alibaba Cloud: Supporting Full-Scenario Intelligence for Automakers Why does intelligent coding in the automotive industry naturally need players like Alibaba Cloud? An inescapable fact is that Alibaba Cloud has ranked first in China's automotive public cloud IaaS market for five consecutive years, with its market share reaching 42% in the second half of 2025. This means that when automakers need elastic computing power to digest the computing power explosion brought by VLA (Vision-Language-Action), world models, and AI large models being deployed in vehicles, Alibaba Cloud itself is the ready infrastructure layer. Virtually all Chinese automakers, whether in domestic or overseas markets, have chosen to run on Alibaba Cloud. The result of this depth of industry penetration is: Alibaba Cloud has completed practical verification in real mass-production projects in the two core battlefields of smart cabins and autonomous driving. On the cabin side, Alibaba Cloud has deployed an "edge + cloud" collaborative architecture: the edge side deploys the Qwen-Omni omni-modal large model, which can independently perceive the physical world and respond quickly even in weak network environments, while maintaining privacy and security boundaries; the cloud side uses Qwen to connect to the digital world, providing full-scenario service capabilities. This product will be implemented in models from over a dozen brands including GAC and IM Motors. On the autonomous driving side, T-Head's self-developed PPU, Asia-Pacific leading cloud computing infrastructure, and the Qwen large model have their three layers of capabilities vertically integrated, allowing automakers to achieve better efficiency when training and invoking large models. Currently, more than 30 automakers and autonomous driving solution providers are conducting autonomous driving R&D on Alibaba Cloud. The actual usage of T-Head PPU has exceeded 130,000 cards, marking the automotive industry’s largest deployment of homegrown AI chips on public cloud. This full-stack capability has also been completely replicated in overseas markets. Automakers can quickly build core overseas systems, complete connected vehicle deployment within 30 days, and obtain full-pipeline support for technology, services, and ecosystems. Automakers including FAW, Chery, and Li Auto have all established deep cooperation with Alibaba Cloud in their overseas expansion businesses. Therefore, returning to the new era where AI determines the future landscape of smart vehicles, Alibaba Cloud still has the capability to continue serving as the most hardcore infrastructure. Looking back from today, smart vehicles have gone through three rounds of competition. The first round was the competition in computing power and sensors, with the goal of making cars "smarter". The second round was the competition in software-defined vehicles, competing on product experience and the ability to iterate continuously. Today, the third round of competition has quietly begun—AI defining vehicles. The core of this competition is the AI transformation of the R&D system itself. It is about who can build the next-generation smart vehicles in a more efficient way, with less manpower, and with safer quality. A century ago, Ford's assembly line revolution reshaped the production efficiency of the automotive industry, making it possible to "build more cars". A century later, AI Coding is reshaping the R&D efficiency of automotive software, making it a reality to "build smart cars faster and better".