The 2026 China OEM Data Capability Analysis Report has been released.
A single L4 autonomous driving R&D test vehicle can generate approximately 10TB of raw data per day—enough to fill two or three hundred Blu-ray HD movies in just one day; whereas the daily data volume for ordinary mass-produced intelligent connected vehicles mostly ranges from the GB level to 1–2TB. Faced with TB-level data per vehicle per day, automakers are no longer grappling with "how to store it," but rather "how to monetize it." The explosive growth in data volume is forcing automakers to shift from "collection" to "operation": data is no longer just an auxiliary tool for R&D and marketing, but has become a core asset defining automakers' competitiveness, just like engines and chassis.
The 2026 China OEM Data Capability Analysis Report uses the DCMM (Data Management Capability Maturity Model) as its framework. Starting from nine capability domains and 36 sub-domains, it systematically evaluates 24 mainstream automakers, overlaying a Technology-Value (T-V) analysis to present a panoramic map of data capabilities among China's mainstream automakers. The evaluation shows that FAW Group is the only automaker reaching the L5 Optimized level with a comprehensive score of 91.2. Geely, XPeng, NIO, Li Auto, Avatr, and BYD form the leading echelon, presenting an overall pattern of "traditional giants rushing to the top, new forces leading the way, and joint ventures under pressure." It also analyzes how data foundations and governance systems, such as the CDO system and the inclusion of data assets in financial statements, are driving data to shift from a factor of production to a quantifiable asset. On the compliance front, it reviews the latest regulatory frameworks such as the Guidelines for the Security of Automobile Data Export (2026 Edition). Furthermore, it summarizes 17 evolutionary trends in automotive data capabilities over the next 2 to 3 years, such as lakehouse integration and millisecond-level architecture, DaaS and AI Agents, trusted data spaces, and cross-border data ecosystems.
Two signals are worth noting: first, the L3 Steady level is the largest segment in the industry, meaning most automakers have built basic systems but have not yet formed quantitative management capabilities; second, leading new forces and traditional transforming automakers have already created a generational gap in data architecture and data applications. Taking the top five as an example, FAW Group is balanced across all nine domains with no shortcomings, XPeng leads in data architecture and data applications, while data standards are a common weakness for multiple new forces.
Source: ZoS Auto Research
I. Data Awakening: From "Building Cars" to "Nurturing Data", the Winning Hand for Automakers Has Changed
Automakers' competitiveness is shifting from "car manufacturing capability" to "data operation capability"—continuously generating real-scenario data through massive connected vehicles, feeding back into the iteration of autonomous driving algorithms, proactive after-sales services, and production supply chain optimization, thus forming a data flywheel.
Autonomous Driving Data Foundation: Nurturing Data with Scale. BYD has built an industry-leading data closed-loop with a fleet of over 3 million intelligent driving vehicles. As of May 2026, the fleet size of advanced driver-assistance system (ADAS) vehicles has exceeded 3.15 million. The "Eye of God" platform generates over 200 million kilometers of new ADAS training data daily. A 5,000-strong autonomous driving R&D team is driving the iteration of cloud-based world models, generative AI, and end-to-end algorithms. On May 28, 2026, BYD released the Xuanji A3 4nm automotive-grade autonomous driving chip, forming a positive cycle of "vehicle scale expansion—data accumulation—algorithm iteration—experience enhancement—usage rate increase." This is a typical model of using vehicle sales to "nurture" autonomous driving data.
In addition, XPeng's second-generation VLA is trained on a cloud-based computing cluster of approximately 30,000 GPUs, with cumulative R&D investment reaching over CNY 2 billion. It utilizes 50PB of training data to enhance spatial reasoning capabilities—relying on this massive data, the model can achieve self-adaptation in complex scenarios such as bypassing accident sites and proactively reducing speed on bumpy roads. More crucially, the full-link iteration cycle from the cloud to the vehicle end has been compressed to about once every 5 days, enabling the algorithm to quickly respond to real driving feedback and form an agile closed-loop of "data collection—training iteration—deployment verification."
Manufacturing and Supply Chain: Reshaping Production with Data. FAW-Volkswagen (the "Cloud-Based" Distributed Automotive Factory), in partnership with Huawei Cloud, has built the nation's first "cloud-based" distributed automotive factory in Foshan. Through the architecture of "Changchun Headquarters Private Cloud + Branch Factory CloudPond," it achieves comprehensive integration and unified scheduling of data between remote factories and the headquarters, promoting efficient collaboration in R&D, production, and supply chains. Taking the smart factory as an example, FAW-Volkswagen has promoted over 100 intelligent scenarios, generating an annual revenue of CNY 46 million.
The "Industrial Brain" at SERES' super factory connects over 300 partners and directly links to more than 50 core suppliers, improving R&D efficiency by 20% and on-time order delivery rate by 13%. After data instructions are issued, the production line can be adjusted in as fast as 20 minutes.
After-Sales Service: From "Passive Maintenance" to "Proactive Data Service". NIO Service uses data for "proactive service." The Service Guardian system remotely monitors and captures abnormal signals through vehicle networking, predicts vehicle issues, and provides remote warnings, transforming "passive maintenance" into "proactive service." The system features second-level remote diagnostic capabilities. When necessary, diagnostic experts follow up immediately. The comprehensive diagnostic accuracy rate exceeds 90%. It has provided proactive services for approximately 120,000 insured vehicles, reducing the average repair time for similar faults by about 30%.
Marketing and R&D Are Also Transforming. Changan Automobile has built a group-level lead center, officially termed "Lead Digitalized Operation" (Data×AI), also reported as a "digital funnel" full-process operation. It digitalizes the entire process of gathering, cleaning, rating, nurturing, distributing, and reusing leads, increasing the lead conversion rate by about 50% and bringing a cumulative net sales increase of 82,000 units. IM Motors has equipped its R&D with over 30 "digital engineer" intelligent agents, completing tasks such as solution selection, requirement distribution, and auxiliary consulting through multi-agent collaboration. Among them, the body sheet metal structure modeling intelligent agent compresses the 3D design time for a single reinforcing plate from about 3 hours to the minute level. Updating 50 reinforcing plates for the four doors and two covers 10 times can save 1,500 man-hours.
II. Comprehensive Assessment of Data Capabilities of 24 Automakers: FAW Tops, New Forces Lead, Who Is Falling Behind
Among the 24 representative automakers, why does FAW Group rank first by a wide margin at the L5 Optimized level in data capability? The answer is "institution-driven." In 2021, FAW issued the automotive industry's first enterprise-level Basic Data Law, establishing that "data sharing is an obligation, and data-generating departments must not refuse reasonable cross-domain sharing requests." It sorted out over 38,000 business units to achieve "single-point entry, multi-point invocation." Coupled with the "1174" all-employee digital and intelligent transformation strategy and the Seven-Star Cloud Workbench, FAW compressed its R&D cycle by 50% and reduced manufacturing costs by 40%. In addition, mainstream new forces (such as Voyah, SERES, etc.) have all built self-developed or deeply customized data platform architectures, generally adopting technical routes like lakehouse integration, separation of storage and computing, and real-time data warehouses. The latency of the data collection-processing-application link is low (millisecond to second level), and the data scale reaches the PB level.
In terms of data governance path selection, the approaches of different camps are distinctly different. New forces have established clear data/AI-related organizational departments, but their form differs from the "independent data governance departments" of traditional automakers, emphasizing more on the embedded nature of data functions and AI nativeness—they elevate data from "a function/department" to the "underlying operating system of the entire organization," reconstructing the organization in the form of an AI-native, flat, and data-driven middle platform. In early 2026, XPeng merged its Autonomous Driving Center and Intelligent Cockpit Center into the "General Intelligence Center," setting up secondary organizations around foundation models and the Infra foundation. A unified AI middle platform simultaneously supports autonomous driving, cockpit, and robotics businesses. Li Auto dismantled its autonomous driving R&D department, reorganizing it according to AI company structures into major sectors such as Infra (infrastructure, responsible for computing power and data), foundation models, software entities, and hardware entities, and established an AI-native middle platform similar to tech companies.
Traditional independent brands monetize through scale, such as BYD. As of May 28, 2026, BYD's fleet of vehicles equipped with advanced driver-assistance systems has exceeded 3.15 million. Its self-developed "Eye of God" intelligent driving system generates over 200 million kilometers of data daily, building a foundation with real-scenario data. A 5,000-strong autonomous driving team turns strategy into deliverable promises. International brand Tesla China is the only foreign automaker that has achieved a full-link domestic closed-loop of "collection—storage—training—deployment—iteration." The Shanghai Data Center and the Lingang AI Training Center form a dual-center structure, accumulating over 3 billion kilometers of Chinese road data. Joint venture automakers, constrained by global architectures and lagging in platform localization, are relatively behind in data governance.
It is worth mentioning that data assets are becoming a new watershed. Avatr Technology successfully passed the national DCMM Quantitatively Managed Level (Level 4) certification in May 2026—this is the highest-level certification in China's only national standard system in the field of data management, marking that the enterprise's data governance has achieved refined operations that are "quantifiable, measurable, and predictable." Relying on this standardized control system, Avatr has implemented the industry's first closed-loop of data asset inclusion in financial statements for automakers. It has transformed the massive vehicle-end data, previously considered "unable to be counted as enterprise assets and difficult to quantify in value," into real assets that are measurable and can be recorded on the company's financial statements. Furthermore, it has broken down cross-departmental silos with a six-dimension data quality scoring mechanism, compressing the cross-domain data retrieval time for R&D personnel from days to the minute level.
By the end of 2025, the capitalization of automakers' data assets has moved from "inclusion in financial statements" to "capital operations," truly entering the value vision of the capital market. BAIC New Energy used data assets such as after-sales, quality claims, and power batteries as pledges to obtain CNY 200 million in financing from the Bank of Beijing, setting a record for the largest single data asset financing in the automotive industry. Wuhu Jinju Technology, under Chery, also completed the first data asset pledge financing in the Yangtze River Delta using an automotive aftermarket dataset. On the exchange front, the Shenzhen Data Exchange has completed the registration of 126 data assets from 32 automakers, with the transaction scale breaking through CNY 1.5 billion. Data centers in Wuhu (Yangtze River Delta) and the Western Data Trading Center are also intensively listing in the automotive industry data zone. Head automakers such as BYD, SAIC, and Geely are among the first batch to complete data asset-related layouts.
Whether they can "operate data" is beginning to affect the valuation of automakers in the capital market.
III. Policy Red Lines and Safety Bottom Lines: The "Compliance Compulsory Course" for Automakers
2026 is a "big year" for data element policies. The Guidelines for Data Property Rights Registration were issued in July, and the national registration system went online for trial operation in September, giving vehicle-end data a unified national "legal ID" for the first time. The National Data Administration launched the High-Quality Data Set Construction Action for the industry, with 106 chain-leading enterprises taking the lead in pilot programs, and the automotive sector is a key vertical. The Ministry of Industry and Information Technology's (MIIT) "Industrial Data Foundation Building Action" listed automotive and steel as the first batch of key industries. The automotive industry consortium jointly led by China Automotive Engineering Research Institute (CAERI) and China Mobile Shanghai has been implemented in Chongqing. The Implementation Plan for Digital Transformation of the Automotive Industry issued by four ministries set hard targets: the numerical control rate of key processes will exceed 70% by 2027.
Beyond policies, there are red lines. The Provisions on the Administration of Automobile Data Security advocates four principles: in-vehicle processing, default non-collection, applicability within accuracy ranges, and desensitization processing. It also stipulates that important data must be stored domestically, and providing it overseas requires a security assessment. In February 2026, the MIIT and seven other departments jointly issued the Guidelines for the Security of Automobile Data Export (2026 Edition), detailing important automobile data into 27 categories and 51 items, providing nine types of exemptions and three export paths. The determination that "remote access equals export" directly hits the global IT architecture of joint venture automakers. Meanwhile, the trusted data space for the automotive industry undertaken by CAERI covers the entire chain of R&D, production, supply, marketing, and service with a "four horizontal and four vertical" framework, supporting scenarios such as autonomous driving insurance, battery passports, and carbon footprints. BAIC New Energy obtained CNY 200 million in financing from the Bank of Beijing by pledging data assets, setting a record for the largest single transaction in the automotive industry—data has transformed from a "cost" into "collateral."
Regulation is also "opening the floodgates." In December 2025, the MIIT announced the first batch of L3 autonomous driving model access permits, with Changan and Arcfox models approved. In July 2026, Beijing, Chongqing, Shanghai, Guangzhou, Shenzhen, and Wuhan simultaneously launched L3 full-domain road trials. Coupled with the expansion of the "Vehicle-Road-Cloud Integration" pilot from 20 cities to 50 cities, automotive data is not only being regulated but also being put to use.
Safety is the bottom line, and compliance is the ticket. In January 2026, Qi An Xin released the Vulnerability Analysis Report on Cloud Platforms of Intelligent Connected Vehicles: among the cloud platforms of 30 mainstream manufacturers, 28 were found to have security vulnerabilities, a proportion as high as 93.3%; 23 had critical or high-risk vulnerabilities (76.7%); a total of 207 vulnerabilities were discovered, of which 135 (65.2%) originated from low-level errors such as unverified identities and unauthenticated interfaces; 22 manufacturers had identity authentication and access control vulnerabilities (73.3%), and 60% had failed access controls. Digital key management failures, over-exposed interfaces... Cloud security has become the biggest "hidden reef" for the large-scale implementation of intelligent driving.
The counterattack by automakers has also begun. XPeng and Alibaba Cloud signed a post-quantum encryption cooperation agreement, with the new P7 being the first to be equipped with post-quantum security technology. NIO independently developed the NPCC cloud-edge integrated AI agent framework, building a five-layer data privacy protection. Geely released the Global Security 2.0 system, introducing quantum security technology into vehicle networks. BYD built a defense line with its "i-Di Dog" intelligent security platform overlaid with UN R155/R156 and ISO 27001 certifications. SERES, on the other hand, shouted that "safety is the ultimate luxury." When it comes to data security, automakers are beginning to treat it as a "new luxury configuration" to compete on.
Compliance is entering the "field-level" era. The 2026 Guidelines for the Security of Automobile Data Export clearly defines three scenarios that constitute export. The traditional model of joint venture automakers opening databases to overseas headquarters via global IT systems must be transformed. "Chinese data center plus local AI training" has become the standard configuration for foreign automakers in China—the Tesla model is being replicated by Volkswagen, BMW, and Mercedes-Benz. Joint Venture 2.0 is also reconstructing data rights and responsibilities: SAIC-GM renewed its contract until 2047, transferring product definition rights to the Chinese team; FAW-Volkswagen is reconstructing the joint venture pattern with a Chinese-led digital and intelligent transformation. The CDO (Chief Data Officer) and data governance committees are expanding from standard configurations for leading OEMs to industry-wide requirements.
IV. The Battle of the Foundation: From 10,000-GPU Clusters to Trillion-Level Real-Time Analysis of IoV Signals
If the scoring is a "physical examination report," the data platform is the "hematopoietic system."
XPeng has turned its data foundation into a "cloud model factory"—starting from 10,000 GPUs to a 30,000-GPU cluster, XPeng's second-generation VLA is based on a 30,000-GPU computing cluster with a computing power of 10 EFLOPS. The cluster utilization rate is consistently over 90% (peaking at 98%). With a CNY 2 billion investment in training and utilizing 50PB of training data, the full-link iteration from the cloud to the vehicle end occurs about once every 5 days. Based on the vehicle-cloud integrated data closed-loop of "collection, encrypted upload, cleaning, storage, annotation, training, simulation, OTA, and attribution," it horizontally connects the seven major business domains: vehicles, users, sales, after-sales, manufacturing, supply chain, and R&D.
Li Auto, on the other hand, has pushed real-time analysis to the extreme. Based on Alibaba Cloud's Hologres + Flink, it has built a trillion-level IoV signal millisecond-response real-time analysis platform. With costs remaining the same as before, write performance has increased by 200%, achieving trillion-level data writes per second, reducing computing costs by 40%, and supporting the real-time processing of over 10,000 signals per second per vehicle for millions of connected vehicles.
Geely is driven by a dual-wheel approach of "organization plus computing power." In 2022, it established a data governance committee chaired by the CDO, building a tier of first-level and second-level data owners and data stewards, making it the most standard governance organizational model among domestic automakers. The computing power of Xingzhi Intelligent Computing Center 2.0 reaches 23.5 EFLOPS, ranking first among Chinese automakers. It uses the GDMP big data platform and Jingzhou Lakehouse to undertake the access of over 3 million IoV connections, and jointly established the "Intelligent Vehicle Computing Power Alliance" with Step Star and Afari Technology.
Source: ZoS Auto Research
Joint venture automakers are also catching up. SAIC Volkswagen and Alibaba Cloud Lingyang Dataphin jointly built a full-domain data asset system, systematically sorting out over 10,000 data objects (tables, fields, metrics, dimensions, tags). Through field-level lineage tracking, it connects R&D, production, supply, and marketing, building a standardized data asset inventory covering the entire company. The successful implementation of Dataphin at SAIC Volkswagen not only constructs a clear, trusted, and easy-to-use enterprise-level data asset system but also reshapes the work paradigm of data development and governance. It has become the core engine for SAIC Volkswagen to achieve full-lifecycle management of data "collection-construction-management-application," laying a solid foundation for digital transformation.
The gap in data foundations is shifting from "having or not having" to "using well or not."
Looking ahead to future trends, three certainties are worth focusing on:
Real-time processing: Data architecture will comprehensively leap from T+1 batch processing to second-level or even millisecond-level. Stream-batch integration (Flink + Paimon/Iceberg) will become the default choice for newly built platforms;
Training closed-loop: The competition in autonomous driving is moving from a "data closed-loop" to a "training closed-loop." World models and simulation are shifting training data from passive collection to active synthesis. Data quality and simulation capabilities are replacing pure data volume as the core barriers;
Compliance localization: The Guidelines for the Security of Automobile Data Export (2026 Edition) pushes regulation to field-level practical operations. The determination that "remote access equals export" forces joint venture automakers to complete the "logical isolation + physical separation" between their data platforms in China and global systems.
In the future, to judge the "data quality" of an automaker, one does not need to look only at how many technical terms it releases, but can look at four things: whether the organization has set up a CDO and a data governance committee, whether the platform is truly used at high frequency by the business, whether data has completed assetization confirmation and inclusion in financial statements, and whether data capabilities have been transformed into verifiable revenue or efficiency metrics. In short, data capability is the watershed for the next round of reshuffling.