Author: Fang Yuan
Past the halfway mark of 2026, the automotive industry has delivered a disconcerting report card. In the first quarter, the industry's profit margin dropped to 3.2%, hitting a ten-year low. Almost concurrently, NIO announced that the shipment volume of its in-house Shenji chip has exceeded 550,000 units, reducing the per-vehicle intelligent driving hardware cost by approximately 10,000 RMB; Li Auto's Mach M100 achieved mass production and vehicle integration in May, becoming the world's first mass-produced dynamic dataflow AI chip. Between this decrease and increase, a clear industrial signal has been sent—when profit margins are compressed to the extreme, no technology capable of reducing BOM costs will be overlooked. Cabin-driving integration and central computing happen to stand exactly at this intersection.
In the past, a smart vehicle required at least two independent computing systems—one for the intelligent cabin and one for intelligent driving. Each had its own chips, memory, thermal management, and wiring harness layout. The two "brains" operated independently without interconnection, which not only caused hardware redundancy and space waste but also resulted in extremely rigid coordination between the systems. Today, however, cabin-driving integration is merging these two "brains" into one. The core of this transformation is using a single chip to simultaneously handle the computing tasks for both the cabin and intelligent driving, sharing the same memory pool—the ultimate form known in the industry as "single-chip shared memory." From "one box, two boards" to "single-chip sharing," this evolution is rewriting the rules of the automotive semiconductor game at an unexpected speed.
01. From "One Box, Two Boards" to "Single-Chip Sharing"
The hardware integration of cabin-driving integration did not happen overnight but followed a clear evolutionary path. The initial form was "One Box": one box containing two boards. The intelligent driving domain controller and the intelligent cabin domain controller were housed in the same physical chassis, living under the same roof but operating independently. Its value primarily lies in saving independent housings and simplifying wiring harness layouts. Tesla's HW3.0 and HW4.0, as well as Xiaomi's four-domain integrated controller, are typical representatives of this form.
The next step is "One Board": one board with dual chips. The computing chips for intelligent driving and the cabin are placed on the same printed circuit board (PCB), transforming inter-chip communication from a "long trek" across boards to "neighborly visits" via on-board routing. NIO's 2024 central computing platform ADAM adopts this hardware form: the intelligent driving and cabin chips are on the same PCB but hardware-isolated. By 2026, GAC officially released its Star Spirit (Xingling) E/E Architecture 4.0, equipped with a 3nm flagship chip, achieving the integration of six domains for the first time: intelligent driving, intelligent cabin, chassis, powertrain, body, and vehicle-to-everything (V2X) connectivity. XPeng launched its new-generation EEA 4.0 electronic and electrical architecture, and Xiaomi's YU7 achieved four-domain integration, reducing the volume of the electronic and electrical architecture by 57%. Leading domestic OEMs have built an EEA architecture system of "one central computing platform plus two to four zone controllers." Chips are no longer mere hardware procurement items but core components of the vehicle's overall intelligent architecture.
The ultimate final form is "One Chip": single-chip shared memory. A single system on chip (SoC) simultaneously handles the computing tasks for both the cabin and intelligent driving. Data flows freely within the chip with nanosecond-level latency and hundreds of GB/s bandwidth, turning computing resources into a uniformly schedulable pool. Solutions like Horizon Robotics's Star 6P and Black Sesame Technologies's Wudang C1296 are targeting exactly this level.
02. Who Is Building "One Chip to Support the Whole Vehicle"?
In this race, domestic chipmakers have become a formidable force. In April 2026, Horizon Robotics released the "Star" 6 series, China's first cabin-driving integrated vehicle intelligent agent chip. The flagship Star 6P adopts a 5nm automotive-grade process, featuring a BPU computing power of 650 TOPS and a memory bandwidth of 273 GB/s. More crucially, Horizon Robotics pioneered the "Castle" physical isolation architecture within the chip—the cabin and intelligent driving run on independent computing units without interfering with each other, yet they can share memory. Star 6P has secured cooperation intentions from over ten automakers including BAIC, BYD, Changan, and Volkswagen, as well as domestic and international Tier 1 suppliers like Bosch and Denso. Chery iCAR has announced becoming one of the first partners for the Star chip. Horizon Robotics founder Yu Kai stated that this chip will simultaneously make it a reality to "reduce space occupation by 50%, lower the comprehensive per-vehicle cost by 1,500 to 4,000 RMB, and shorten the R&D and delivery cycle from 18 months to 8 months."
Almost simultaneously, Black Sesame Technologies reached a platform-level partnership with Dongfeng Motor. The jointly built Dongfeng Tianyuan Smart Cabin Plus cabin-driving integrated mass production platform is equipped with the Wudang C1296 chip. The Wudang C1296 adopts a 7nm automotive-grade process, achieving hardware-level integration of four major domains—cabin, intelligent driving, gateway, and MCU vehicle control—on a single chip, and has passed the highest ASIL-D certification for ISO 26262 functional safety. This platform will be first deployed on the Dongfeng Yipai 007 and is planned for scaled application across multiple mass-produced models from 2026 to 2027. This is also the first domestic cabin-driving integrated mass production platform. Ding Ding, Vice President of Products at Black Sesame Technologies, revealed that the company's next-generation chip is already in planning, expected to be launched around 2027, and will adopt a more advanced process node.
SiEngine, on the other hand, released the 5nm automotive-grade AI cabin-driving integrated chip "Longying No. 2" in April 2026. Its AI computing power reaches 200 TOPS, integrating a 12-core CPU and a 10-core GPU, with a bandwidth of up to 518 GB/s. It natively supports multimodal large models with over 7 billion parameters and plans to start adaptation in the first quarter of 2027. Wang Kai, founder of SiEngine, revealed that Longying No. 2 can simultaneously guarantee the parallel execution of three complex tasks: AI, intelligent cabin, and intelligent driving. After selling a million units of "Longying No. 1" in the cabin market, SiEngine seized the momentum to launch "Longying No. 2" in April 2026: 5nm, 200 TOPS, starting adaptation in Q1 2027. According to SiEngine's founder Wang Kai, this chip can handle the three tasks of AI, cabin, and intelligent driving simultaneously. It is worth noting that its architecture covers all price segments from entry-level to flagship—SiEngine clearly does not want to limit its business to just the cabin market.
SemiDrive Technology chose a different path—launching the AMU secure real-time computing foundation for the central supercomputing architecture. Its flagship product, the E3800, integrates over ten cores on a single chip, introduces aerospace-grade embedded memory with performance ten to twenty times that of traditional eFlash, and integrates a high-performance NPU to achieve deep coupling and collaboration between the CPU and NPU. SemiDrive's positioning is the "central intelligent control cerebellum," providing a real-time computing foundation for the vehicle's central computing. Unlike Horizon Robotics and Black Sesame, which focus on high-computing-power AI SoCs, SemiDrive focuses on the secure computing power that determines real-time tasks such as vehicle chassis control and body response—these two constitute the two sides of the same coin for the central computing platform.
International giants are also accelerating their layouts. Qualcomm's Snapdragon Ride Flex SoC (Snapdragon 8775) has achieved mass production deployment, securing design wins for nine vehicle models. The higher-performance Snapdragon 8797 is on the eve of scaled implementation. BAIC Arcfox Alpha T5 has become the first domestic vehicle model to achieve mass production of cabin-driving integration based on Qualcomm's 8775, earning the title of the "first cabin-driving integrated vehicle." Chelian Tianxia's AL-A1 domain controller, built on the Snapdragon 8775, is the world's first mass-produced single-chip cabin-driving integrated domain control solution. Leapmotor's new flagship SUV D19 debuts with dual Qualcomm 8797 central domain control chips, building a cabin-driving integrated architecture.
NVIDIA DRIVE Thor has been in mass production since 2025, with partner automakers including BYD, Li Auto, Zeekr, and Xiaomi. At the Beijing Auto Show in April 2026, Lenovo Car Computing launched the Auto AI Box, one of the first domestic cabin-driving intelligent computing platforms based on NVIDIA DRIVE AGX Thor. Jiyue announced that its mass-produced models will be equipped with NVIDIA DRIVE Thor starting in 2026. Aptiv announced in June 2026 that it is expanding its partnership with NVIDIA to build a "mass-production-grade foundation" for Jetson Thor. NVIDIA also partnered with MediaTek to make up for its cabin shortcomings, attempting to use Thor to "sweep" the entire market. Danny Shapiro, Vice President of Automotive at NVIDIA, stated that the company's automotive business in China is growing significantly, and 2026 will become a crucial year for revenue contribution in the automotive business.
The wave of in-house development by automakers is equally surging. NIO's Shenji NX9031 has been mass-produced on a large scale across its full range of high-end models such as the ET9, ES9, and ES8, with shipments exceeding 550,000 units by early 2026. Li Auto achieved mass production and vehicle integration of the Mach M100 in May 2026, making it the world's first mass-produced dynamic dataflow AI chip, with a computing power of 1,280 TOPS. Utilizing a dataflow architecture, it claims a utilization rate of over 82%. XPeng built a supercomputing platform based on three in-house Turing AI chips, with a total computing power of 2,250 TOPS. BYD's Xuanji A3 features three chips in parallel, exceeding 2,100 TOPS.
03. Why Exactly 2026?
If one must ask why cabin-driving integration exploded exactly in 2026, the answer might not be romantic at all—it's just about money.
The price wars of the past three years, coupled with high-intensity R&D investments, have squeezed the profits of Chinese automakers to the limit. In the first quarter of 2026, the industry's sales profit margin was left at only 3.2%, with the net profit per vehicle for the vast majority of Chinese automakers falling below 10,000 RMB. Under such thin margins, any technology that can save a few hundred RMB will immediately be brought to the table for discussion.
Cabin-driving integration is precisely the solution that saves money. Horizon Robotics says the Star chip can slash the comprehensive per-vehicle cost by 1,500 to 4,000 RMB. SiEngine's Longying No. 2 claims to save 2,000 to 3,000 RMB. SemiDrive's X10 solution can also save 1,500 to 3,000 RMB. The R&D cycle is shortened from 18 months to 8 months, and space occupation is halved.
NIO's Li Bin also stated: previously, NIO purchased NVIDIA Orin-X chips in large quantities, spending 300 million USD a year at the peak. The cost of four Orin chips per vehicle was between 1,900 and 2,500 USD. After the Shenji chip was deployed in vehicles, the per-vehicle intelligent driving hardware cost dropped by approximately 10,000 RMB. This figure, placed on the financial statements of any automaker, is enough for a CEO to make the final decision to push forward with in-house development.
The second driving force is that the distributed architecture has reached a dead end.
Over the past decade, the automotive industry did one thing—piling ECUs onto vehicles. For every added function, an ECU was stuffed in: one for seat heating, one for window lifting, one for rain sensing, one for automatic headlights. Ultimately, a single vehicle was packed with hundreds of ECUs, with a total wiring harness length exceeding two kilometers and a weight comparable to that of an adult. What are the consequences of this piling approach? Software upgrades become nearly impossible, cross-functional coordination is virtually unachievable, and the material and assembly costs of the wiring harness are outrageously high.
In 2026, this model reached its end. Continuing to pile them up would affect range due to weight, cause R&D cycles to spiral out of control due to complexity, and leave the vehicle's overall response lagging due to communication latency. A consensus formed in the industry is that the next-generation vehicle architecture must be central computing plus zone control. Cabin-driving integration has become a critical stepping stone for the electronic and electrical architecture to move towards central computing—first merging the cabin and intelligent driving, the two most computing-intensive domains, to set an example for the subsequent comprehensive consolidation of the chassis, powertrain, and body.
04. The Real Challenge Is Not on the Silicon Wafer
The single-chip solution sounds wonderful, but the reality engineers face is much more complex. The cabin chip itself is the epitome of heterogeneous computing—CPU, GPU, NPU, video codec, audio processing, and display output are all squeezed onto a single silicon die, with complexity already astonishingly high. Autonomous driving chips pursue a different logic: determinism. When each frame of image data comes in, how long it takes to complete inference and output control signals must be precisely predictable, without the slightest jitter. One is the ultimate in multi-functional integration, and the other is the ultimate in timing determinism. Placing these two design philosophies on the same chip is in itself an extreme engineering challenge.
Zhang Yuantao, Vice President of ADI China, corroborated this judgment from another perspective. In his view, there are only two reasons why high-end automakers are reluctant to pursue single-chip integration. On the cost side, "high-end vehicles do not need to consider cost issues too much, nor will they pursue the ultimate multi-in-one cost-reduction solution in terms of materials." On the risk side, "integrated solutions bring an increase in complexity; using dedicated chips in specialized fields will be better both in terms of safety and performance." High-end vehicles would rather use one top-tier chip for the cabin and two or even four for intelligent driving, maximizing each without interfering with one another. It is not that single-chip technology is inadequate, but for high-end products, the "determinism" of performance and safety is much more important than saving a few thousand RMB.
It is precisely the reverse of this logic that makes mid-to-low-end vehicles the most attractive battlefield for cabin-driving integration. For a car priced in the hundreds of thousands of RMB, using two separate chips for the cabin and intelligent driving will immediately drive up BOM costs; a single chip has excess performance, while two together are still not enough—this awkwardness is easily resolved by the single-chip solution. Zhang Yuantao's judgment is that in the mid-to-low-end market, cabin-driving integration is a definite major trend, "and it can even be said that ultimately one chip will cover both intelligent driving and the cabin, because this can truly achieve cost control and the trend of platformization." In the mid-to-high-end market, "it is highly likely to adopt discrete SoCs, meaning there will be an independent cabin SoC and an independent intelligent driving SoC, but it may form a One Box solution—two boards placed in one box, or two chips made on one board."
Therefore, the penetration path of cabin-driving integration is highly likely not "top-down," but "bottom-up"—first breaking into the mid-end market with the largest shipment volume, while the high-end market will follow the route of discrete architecture plus physical integration for the long term.