By Zhou Xiongfei

The automotive industry has been discussing the 'soul' of intelligence for six years.

In 2021, Chen Hong, then Chairman of SAIC Motor, was asked whether SAIC would consider adopting third-party solutions in the field of intelligent driving. He responded, 'If we adopt third-party solutions, they become the soul. SAIC cannot accept such an outcome; we must keep the soul in our own hands.'

While this 'Soul Theory' quickly sparked widespread industry discussion at the time, with new-force automakers like Tesla demonstrating their intelligent driving capabilities, the entire industry rapidly entered a race to implement intelligent driving models and functions in 2023 and 2024. Traditional automakers have also made more pragmatic choices.

Currently, traditional or joint-venture automakers have three main options for intelligence deployment:

1. Abandon the 'soul' and pursue rapid functional implementation. Directly procure third-party supplier solutions to quickly integrate intelligent (driving/cockpit) functions into vehicles and capture market attention. However, core software and hardware capabilities remain out of their hands, and they are prone to falling into the 'vortex' of homogeneous competition.

2. Partial self-research, with the 'soul' trapped in a 'semi-black box.' The cockpit and in-vehicle app interfaces remain at UI customization and application-layer adaptation, while chips that determine on-device computing speed and intelligent driving algorithms that determine user experience are outsourced to third-party suppliers. Development timelines are constrained by solution adaptation progress.

3. Master the 'soul' through full-stack self-research + industrial collaboration. From vehicle EE architecture to intelligent driving chips, from intelligent driving algorithms to data loops, all R&D is kept in-house. Although this requires significant time and financial investment, collaboration across the industrial chain can accelerate overall progress.

Volkswagen Group, often seen as 'an elephant that is hard to turn,' embarked on the third and most challenging path in 2023. Three years later, I see them building a 'smart soul' capable of continuous thinking, interaction, and evolution.

Huang Chang, CTO of Carizon, shares insights

Based on the CEA one Team, Volkswagen integrates businesses such as VCTC and Carizon to complete the construction of a full-chain R&D system spanning battery management, powertrain systems, CEA electronic-electrical architecture, and platform-based software for intelligent driving, cockpits, interactions, and the cloud.

It can be said that Volkswagen's 'smart soul' has rapidly propelled its products into 'delivery mode.'

From an intelligent driving perspective, the highway pilot assist solution based on the CEA 1.0 architecture has been implemented in the all-new Virtus 07 and facelifted Virtus 06 models. Volkswagen's self-developed all-scenario advanced driving assist solution, Hyper Sense 8 (HS8), will be progressively integrated into seven new products built on the CEA architecture starting in the third quarter of this year.

Volkswagen is also advancing toward L3 and L4 autonomous driving. According to Volkswagen's plans, the first batch of L3-level conditional autonomous driving technologies will be ready by the second half of 2027 at the earliest.

This accelerated delivery pace is driven by Volkswagen's intelligent driving software-hardware collaboration technology system.

Through the AI foundation model Hyper Sense, the simulation data platform GAIA 2.0 world model, and the 'trinity' collaboration achieved by developing chip platforms in partnership with Horizon Robotics, Volkswagen can leverage data loops to drive the large-scale implementation and rapid iteration of the HS8 intelligent driving model, thereby enhancing the user experience of intelligent driving functions.

The intelligent driving industry is now in an elimination race, and Volkswagen, which is reshaping its 'soul' through the philosophy of 'leveraging external resources and focusing on core self-research,' is charging ahead with accelerated momentum.

For Volkswagen China, one of this year's key themes is large-scale delivery.

As Han Sanchu, Executive Vice President of Volkswagen Group China and CEO of CARIAD China, told me, 'We are no longer in the stage of discussing technical routes or validating concepts; we have truly entered the phase of large-scale delivery. What we need to deliver is not just a single vehicle but a systematic, large-scale delivery across multiple models and projects.'

Han Sanchu, Executive Vice President of Volkswagen Group China and CEO of CARIAD China, shares insights

This delivery rhythm began at the end of last year when Volkswagen introduced the highway pilot assist solution based on the CEA 1.0 architecture, a pure vision solution powered by Horizon's Journey 6M chip. This solution was implemented in the all-new Virtus 07 and facelifted Virtus 06 models from Volkswagen Anhui in May this year.

Subsequently, Volkswagen's self-developed all-scenario advanced driving assist solution, HS8, is ready for implementation and will be progressively integrated into seven new new energy vehicle models from three joint ventures—SAIC Volkswagen, FAW-Volkswagen, and Volkswagen Anhui—starting in the third quarter of this year.

Among them, the pure vision solution, based on the Journey 6M chip and featuring 11 cameras as perception hardware, will debut in the SAIC Volkswagen ID.ERA 5S and FAW-Volkswagen Jetta M6 models. The pre-sale price of the ID.ERA 5S was announced last week at RMB 115,900-145,900.

Although the Jetta M6 will be launched later this year, I saw its engineering vehicle at Volkswagen's self-developed intelligent driving technology communication event last week, and its completion level was already very high.

Judging from the intelligent driving road test videos of the ID.ERA 5S and Jetta M6, the HS8 pure vision solution is already mature. It not only performs well in basic scenarios such as unprotected left/right turns, lane changes, U-turns, and avoiding large vehicles but also excels in unmarked rural roads, where it can smoothly yield to oncoming traffic and navigate around other road users. Its overall driving capability is virtually indistinguishable from that of an experienced human driver.

HS8 pure vision solution: self-vehicle yielding to the right

In addition to the pure vision solution, HS8 also offers a LiDAR-based solution, built on Horizon's Journey 6H chip, 11 cameras, and one LiDAR, with on-device computing power increased to 420 TOPS.

This LiDAR-based solution will debut in the FAW-Volkswagen ID.AURA T6, which is also the first pure electric model based on the CEA 1.3 architecture. Subsequently, models such as the SAIC Volkswagen ID.ERA 5X based on the CEA 1.3 architecture and the LiDAR-equipped Virtus 06 and 07 models will also adopt this solution.

According to Volkswagen's plans, starting in 2027, the HS8 LiDAR solution will be further extended to more models based on the CEA architecture.

From the above intelligent driving solution implementations, it is evident that Volkswagen has taken a counterintuitive approach—deploying intelligent driving solutions first in entry-level products like Jetta and then extending them upward.

In contrast, the industry's common practice is to debut solutions in high-end products to signal technological superiority to the industry and consumer market before rolling them out to verified technologies downward.

This counterintuitive move hides Volkswagen's strategic thinking.

First is the consideration of unified platformization. In Huang Chang's view, excelling in high-end intelligent driving solutions does not necessarily guarantee success in mid-tier solutions. Conversely, if mid-tier solutions are well-executed, high-end solutions should not be overly challenging.

Currently, some players in the industry prepare two technology stacks for intelligent driving solutions, one for high-end and one for mid-to-low-end products, with separate algorithms and data. However, Volkswagen considered platformization and scalability from the outset, using a unified technology stack that requires only trimming and adaptation for different computing platforms, eliminating the need for rebuilding and significantly reducing migration and adaptation costs.

Second is the consideration of data loop growth. Across the industry, we have entered a data-driven era. Han Hongming, CEO of Carizon, believes that intelligent driving is a data-driven technology, and the volume of data directly determines iteration speed.

Han Hongming, CEO of Carizon, shares insights

After all, the initial steps in building an intelligent driving system are often the most challenging. Starting with entry-level products allows for collecting more high-quality data as more vehicles are sold, thereby driving faster iteration of the intelligent driving model and enabling efficient R&D at a lower cost.

It is evident that Volkswagen has quickly translated this strategic thinking into phased strategic achievements, supported by its intelligent technology foundation.

Is full-stack self-research of intelligent driving capabilities difficult?

The facts prove—yes, because it is a long-term endeavor built on capital and time, with every link from chips and algorithms to data loops being a bottomless pit for funding. From project initiation to true mass production and implementation, it often endures years of silence, and the final outcome may not even keep pace with market competition.

In the current industry landscape of price wars and elimination races, Volkswagen also faces challenges. As Han Sanchu put it, 'For the sake of time, we won't start from scratch.' Their solution is to fully leverage China's efficient supply chain and compress R&D cycles through external collaboration.

As early as 2022, Volkswagen's software company CARIAD officially announced its collaboration with Horizon Robotics and jointly established the joint venture Carizon (CARIZON) a year later. Subsequently, their collaboration accelerated, with mass production cooperation based on the Journey 6 in 2024 and deepened high-level cooperation based on HSD in April 2025.

By November 2025, they entered the chip co-definition phase, and at Horizon's technology ecosystem conference in December of the same year, Han Hongming further disclosed details of Volkswagen's self-developed chip.

This chip, the C7H, is built on Horizon's latest Riemann architecture, with a 3-5 nanometer process, delivering 500-700 TOPS of single-chip AI computing power, suitable for L2 all-scenario assist driving and L3+ autonomous driving.

This year, Volkswagen's collaboration with Horizon continues to deepen. Last month, Carizon announced that it had officially signed an agreement with Horizon to deepen technical cooperation in the field of AI foundation models.

Based on these collaborations, Volkswagen has built its core intelligent capabilities in China layer by layer on Horizon's platform foundation, making Carizon the intelligent capability center for Volkswagen China through platform-level deep integration.

As Han Hongming introduced at the communication event, Carizon currently has 800 full-time employees, with 92% in R&D and 60% hailing from Chinese high-tech enterprises. More importantly, from the GAIA 2.0 data platform to the HyperSense AI foundation model and the C7H chip, Volkswagen has established an autonomous and controllable technology system through Carizon, spanning data, models, and chips.

The underlying logic is that, despite close collaboration with Horizon, Volkswagen is not simply adopting a 'take-and-use' approach but firmly holding technological initiative in its own hands.

Take their technical cooperation agreement in the AI foundation model field as an example. Carizon will develop and accelerate the construction of a unified AI driving solution for Volkswagen Group China based on a 'white-box' licensing model, leveraging Horizon's advanced AI foundation model capabilities.

To elaborate, their collaboration primarily focuses on foundational model cloud pre-training, with Carizon responsible for trimming, distilling, quantizing, and adapting the foundational model to automotive-grade on-device requirements, autonomously completing subsequent foundation model development and management.

In other words, Volkswagen leverages Horizon's decade-plus of leading technology to build foundational models, with subsequent model training, functional development, and adaptation to chips and the CEA platform architecture fully controlled by Volkswagen itself.

At the chip level, although the C7H shares a similar underlying technology architecture with Horizon's Journey 7 series, Huang Chang noted that many exclusive designs have been made to ensure the C7H is a customized chip with Volkswagen's distinct characteristics.

A unified chip architecture means that many upper-layer software and applications can be fully reused, reducing software and hardware adaptation costs and ensuring long-term compatibility after chip integration, avoiding adverse effects from chip switches.

For Volkswagen China, they do not pursue the extreme goal of doing everything in-house but rather adopt a more pragmatic approach based on market development, considering which areas can be achieved through win-win collaboration and which must be self-mastered.

As Han Sanchu put it, 'The most core capabilities must be kept in our own hands, but that doesn't mean we do everything ourselves.' Based on this technological development philosophy, Volkswagen is accelerating its intelligent technology R&D progress, rapidly turning technological shortcomings into strengths.

In high-quality data simulation generation, the GAIA 2.0 data platform can achieve daily active safety simulations of 300,000-5 million kilometers and daily urban NOA simulation mileage of 10,000 kilometers, with simulation data volume currently reaching nearly 10 million hours.

If GAIA 2.0 is seen as a 'gym,' then Volkswagen's self-developed AI foundation model, Hyper Sense, is the 'fitness coach,' continuously training and iterating the intelligent driving model based on high-quality data, enabling the 'athlete,' HS8, to achieve more human-like understanding and decision-making, delivering safer, more comfortable, and efficient intelligent driving function experiences to consumers.

More importantly, in terms of algorithm, computing power, and data construction, Volkswagen has tailored them to the CEA architecture from the outset. As the CEA architecture is progressively integrated into Volkswagen's pure electric, hybrid, extended-range, and even fuel-powered models, Carizon's intelligent driving solutions will also enter a phase of systematic and large-scale delivery, generating more data to drive model training and iteration.

With this 'smart soul,' Volkswagen has achieved its current acceleration, but it must also tread carefully every step of the way in the future.

L3 or L4 has become the new target for automakers.

Take L3 as an example, brands such as Changan, BAIC, GAC, SAIC, Geely, BYD, and Hongmeng Intelligent Driving have all announced their respective timelines for L3 implementation, primarily concentrated between this year and next year. Meanwhile, Xpeng Motors has explicitly stated that it will skip L3 and move directly to L4.

While advancing the scalability of L2 intelligent assisted driving, exploration into L3 and L4 is also underway.

According to the solution proposed by Coretex, for future L3/L4 autonomous driving solutions, they plan to adopt a fused perception route based on two SOCs, with a computing power exceeding 1000TOPS.

Regarding chip selection, I was told that Volkswagen will choose both Horizon Journey series and their self-developed C7H, depending on the readiness time of the L3 solution and the adaptation progress of the C7H chip. However, there will be no arrangement for Series connection (tandem connection) of the two chips onboard.

According to Huang Chang, Volkswagen plans to be ready for L3 functionality onboard by 2027, with plans to begin implementing L3 on highways in early 2028 and introducing L3 functionality in urban areas by 2029. In the context of the entire industry, this represents a relatively rapid implementation pace.

Similar to the implementation of L2 intelligent driving functions, the land (implementation) of L3 conditional autonomous driving functions also requires high-quality data to drive the iteration of algorithm models. However, the scale of data required for these two functions is not on the same level.

Generally speaking, the model training data for L2 intelligent driving initially comes from real-world road testing data. As model training iterates, simulated data generated based on real-world data becomes the majority of training data, as obtaining high-quality Corner Cases data from real-world road testing is 'elusive.'

For Coretex, 60%-70% of the data collected three years ago was effective, but now this proportion has dropped to 3%-5% and will further decrease to 1%-2% in the future. In other words, the effectiveness of real-world data collection is decreasing, and simulated data is becoming more important.

This is even more so for L3 model training.

'The scale of testing mileage for L3 is about 1-2 orders of magnitude larger than L2, roughly 100 times. This 100-fold data expansion places higher demands on the system's safety baseline, making it impossible to replicate the data approach used in L2 technology solutions, which rely entirely on real-world data,' Huang Chang told me.

Their solution is that over 90% of the training data is simulated, with the remaining 10% real-world data used to verify the reliability of models trained on simulated data. However, the reality is that many automakers' L3 road testing is still restricted by policies and regulations to small-scale, simple, and repetitive scenarios.

So where should this 10% of high-quality real-world data come from?

A research and development leader from a leading intelligent driving company suggested a possible path: just as when L2 intelligent driving first began to develop, road testing data collection was also subject to regulation, but automakers would find ways to obtain the training data they needed. The same situation may arise with current L3 training.

This means that the key to determining automakers' L3 technological advantages in the future may not just be evaluating the quality of the 90% simulated data but also the quality of the 10% real-world data.

In addition to L3, cockpit-driving integration has also become a focus for automakers and suppliers this year.

For example, at last week's pre-sale launch event for the Xpeng G9L, CEO He Xiaopeng revealed Xpeng's progress in cockpit-driving integration, which will involve the implementation of the VLM model working in tandem with the second-generation VLA model. At this year's WAIC conference, Geely also jointly launched a cockpit-driving integrated super intelligent agent product with Qianli Technology and Step.ai, which will first be implemented in the Zeekr 8X.

Additionally, Horizon officially released a chip named 'Xingkong' (Starry Sky) in April this year, calling it China's first cockpit-driving integrated vehicle intelligent agent chip, capable of covering both cockpit and intelligent driving cross-domain computing with a single chip.

After in-depth discussions with Volkswagen China, I believe they are also actively exploring and laying out strategies in cockpit-driving integration, especially since Huang Chang believes there is significant room for improvement in this area for Volkswagen.

However, truly achieving cockpit-driving integration places higher demands on collaboration among different departments within an organization.

As a result, automakers such as Xpeng and Li Auto have adjusted their organizational structures. Xpeng merged its intelligent driving and intelligent cockpit divisions into a General Intelligence Center, led by Liu Xianming.

Li Auto's autonomous driving team was integrated into the software ontology (entity) team led by Gou Xiaofei, Vice President of Li Auto Intelligent Space, to coordinate the research and development of Li Auto's intelligent cockpit and intelligent driving.

In comparison, Volkswagen is larger in scale and naturally has more work to do. It can be seen that Volkswagen already established the aforementioned CEA one Team in 2024 to achieve cross-departmental and cross-domain research and development collaboration within the Volkswagen system.

Among these, the CEA architecture serves as the 'main heartbeat' of Volkswagen's vehicle technology delivery in China, effectively 'shaking hands' (i.e., integrating technology and products) with different research and development aspects of the vehicle, architecture, and systems to ensure everyone advances research and development according to the same 'heartbeat' rhythm.

Han Sanchu shared a vivid example with me: after joining Volkswagen, he established a dedicated Product and Engineering team where CEA architecture product experts work with Coretex's intelligent driving experts to study, from a user perspective, how cockpit and intelligent driving joint debugging and interaction should be.

In summary, it can be said that Volkswagen China has made relatively comprehensive preparations in terms of technology and organizational structure for future development. However, in Huang Chang's view, Coretex's intelligent driving is not yet perfect and will continue to accelerate optimization, including in subsequent OTA versions and new vehicle models.

In my opinion, Huang Chang's attitude should represent Volkswagen China's current mindset: correct goals, calmness, and pragmatic progress. Therefore, I believe Volkswagen China should continue to accelerate and gain a greater advantage in the increasingly fierce industry elimination race.

(The header image and photos in the article, as well as PPT illustrations, are all from Volkswagen China's official sources.)