Smart driving and cockpit features are packed into one chip—elimination round has begun.
```
Author | Zhou Zhiyu
China's new energy vehicles have been developing rapidly for five years. Range, fast charging, and batteries have become homogenized. Intelligence is the decisive factor in the second half, and no one disputes this anymore.
But the cost structure of intelligent hardware is hitting a wall.
Putting a smart driving SoC and a cockpit SoC together in a 200,000 RMB car—two chips, two systems, two sets of cooling—this bill is increasingly impossible to settle at mainstream price points.
The industry response is “cockpit-driving integration,” combining two chips into one. The concept has been discussed for two years, but most solutions have stayed at the PPT stage. The 2026 Beijing Auto Show is a turning point; companies like Qualcomm, Horizon Robotics, and Black Sesame Technologies all showcased the latest progress in cockpit-driving integration at the show, which had a dedicated exhibition area for the first time. Around Qualcomm’s 8775 chip, at least five Tier 1s presented their own domain controller products.
The density of solutions is already high enough. But if you look closely, you'll find that everyone is fighting over different things—some vie for chip share, some for software definition rights, and some for the foundation of whole-vehicle intelligence.
This concentrated burst is no coincidence. Smart driving is moving from highway NOA to urban NOA, increasing in complexity while car prices go down, and rising memory chip prices push the urgency of integration even further. The architecture where each chip has its own set of memory is becoming increasingly extravagant. Improved processes make single-chip computing power adequate, and the window is opening right at this time.
The number of solutions is already surplus, but mass-production delivery is just beginning. The upcoming competition won't stay at the chip level, and the window for knockout rounds may be much shorter than the industry anticipates.
The Old Architecture Doesn’t Add Up
The problem with dual-chip split architecture isn't just cost. The computing power utilization rate of smart driving chips has long been below 30%; most everyday scenarios don’t require high-end compute, whereas the cockpit GPU is often fully loaded. Each chip reserves its own redundancy, resulting in serious overall waste.
Horizon’s Chief Architect Su Qing told Wallstreetcn on April 23, 2026, that resource utilization of computing systems should not exceed 70%—going over the limit easily leads to avalanche failures. Dividing the system into two naturally results in more redundancy than one. Static partition alone allows merged memory usage to drop to two-thirds of dual systems.
The day before, Horizon had just released China’s first cockpit-driving integrated vehicle intelligence chip.
“The more fragmented you make it, the more memory you need,” he said. “So it must be integrated together.”
Rising DDR prices make the numbers even harder to balance. Automotive-grade storage is experiencing structural shortages; since September 2025, DDR4 prices have risen by more than 150%, DDR5 by 300%. Ideal's VP of Supply Chain Meng Qingpeng warned as early as last December that in 2026 supply fulfillment rates may be less than 50%. Cockpit-driving integration merges two sets of memory into one, and in a storage-scarce environment, saving materials is itself a competitive advantage.
Cost reduction figures have already been validated in mass production. Desay SV said at the auto show that the 8775 solution can reduce costs by 20–30%, Horizon reports saving 50% space, 30% components, and cutting costs by 1,500 to 4,000 RMB.
Where does the saved money go? It goes to user experience.
Arcfox Alpha T5 is the first model mass-produced with the 8775 cockpit-driving solution. One 144 TOPS chip runs both cockpit and smart driving; urban NOA, highway NOA, and cross-floor memory parking are all included, with a starting price of 105,800 RMB. It's the industry's first time offering urban NOA on an SUV at the 150,000 RMB level. If using traditional dual-chip split solutions, this price range couldn’t support these features; the saved BOM space directly converts to benefits in smart driving capabilities.
Sohu Automotive Research forecasts China’s cockpit-driving market CAGR at 36% from 2026 to 2030, with 3.6x growth potential by 2030. The growth driver is not only cost reduction but also the space for function expansion released by the lower costs.
Beyond the economic calculations, there’s an architecture calculation. E/E architecture evolves from distributed to domain controllers to central computing, and cockpit-driving consolidation is a prerequisite for central computing. L3 human-machine co-driving is more straightforward: handover and takeover require cockpit and driving functions to run on the same chip for microsecond-level communication. If divided into two domains, latency can’t be compressed.
Su Qing elevates this matter to the level of physical law. Using the history of computing as an example: ENIAC had more than 8,000 vacuum tubes and failed every few hours, but today chips have many times higher complexity and computers rarely fail. “The essence of computing industry? Integration, integration, and more integration. Don’t ask whether it’s necessary, ask if the integration is high enough, and if the number of components has been reduced.”
Now there’s a chance to put the computing in the car onto a single chip. The window for dual-chip architecture is closing.
Same Mountain, Different Slopes
Qualcomm started in automotive from the cockpit. The 8775 chip, with virtualization isolation and ASIL-D safety island, now includes ADAS capability. Qualcomm doesn’t touch application layer software, selling chips and platform only. Desay SV landed Chery and Tata, Zhuoyu mass-produced the industry’s first 8775 single-chip cockpit-driving solution, Cheliantianxia teamed up with BAIC for their solution, and Huayang and Hangsheng each have their products.
More than five Tier 1s compete around the same chip. Qualcomm makes money on chips, not souls. Advantage: widest coverage. Disadvantage: differentiation depends entirely on solution providers’ software layers above the chip. Zhuoyu goes for native multimodal models, Bosch for three-domain integration (cockpit-driving-chassis), and paths are diverging.
The lineup of 8775 models at the auto show expanded. Beyond the earliest Arcfox Alpha T5 and Dongfeng Nissan N6, Buick E7, BAIC Wenda V9, Arcfox Alpha S5 also announced 8775 adoption. Buick and Dongfeng Nissan are JV brands—JV brands reverse-procuring Chinese intelligent solutions is an extra industrial signal in cockpit-driving integration.
Qualcomm isn’t stopping either. Next-gen Snapdragon 8797 debuts at the show, boasting 700 TOPS single-chip, 12x NPU improvement over last-gen, also supporting cockpit-driving fusion. Leapmotor D19 globally debuts dual 8797, and Qualcomm and Ideal took only 14 months to realize 8797 mass production. Wallstreetcn learned from insiders that more than 15 automakers plan to launch new cars with the Supreme platform this year.
8797 computing not only powers smart driving. Now, this chip supports local operation of 30B-parameter MoE mixed-expert large models; one chip runs smart driving and cockpit models, with requirements far exceeding traditional split architecture for computing and power management.
Just as 8775’s mass rollout begins, 8797 arrives to take over.
This rapid iteration is itself a barrier.
Horizon takes a different route. Founder and CEO Yu Kai made an analogy at a media mixer on April 23, 2026: moving from cockpit to driving, or vice versa, is like climbing the same mountain from two slopes: “Which slope is steeper? Obviously, going from cockpit to driving is harder.”
Horizon’s base is in autonomous driving. The Starry 6P chip launched before the auto show is China’s first native cockpit-driving fusion chip, 5nm, 650 TOPS. Unlike Qualcomm’s virtualization isolation, Starry uses a “Fortress” physical isolation design at the hardware layer, separating the domains. Cockpit and smart driving each have independent compute resource pools; cockpit crashes won’t drag down smart driving. Starry also includes an “Adaptive Compute Engine,” solving dynamic compute allocation in cockpit-driving fusion—more compute is given to cockpit rendering during regular cruising, but automatically shifts to smart driving in complex road conditions.
From chip to smart driving software, cockpit chip to cockpit OS ("KaKaXia"), Horizon does all four quadrants itself. iCAR will debut it. Horizon claims 50% space savings, 30% component savings, 1,500–4,000 RMB cost reduction, and cut R&D cycle from 18 to 8 months.
Every additional layer means more to prove to automakers. Cockpit graphics rendering and OS ecosystem richness are the barriers Qualcomm spent a decade building.
Black Sesame Technologies has its own idea. Wudang C1200 was designed from scratch as integrated, not as an extension of an existing domain. Pure architecture is an advantage, but the narrow client base is a risk.
However, automakers now need vendors to solve not only architecture- and chip-level issues, but competition has moved up the stack.
Above the Chip
At the ongoing 2026 Beijing Auto Show, Horizon launched more than just a chip. “KaKaXia OS” is its formal entry into OS, as Horizon aspires to be the foundation platform for whole-vehicle intelligence, not just chip sales. This means carving out a piece in the cockpit software ecosystem Qualcomm has built over a decade. Whether that happens depends on mass production performance.
Zhuoyu takes another route. It doesn’t make chips, but does the model layer above the chip. At the show, it unveiled its native multimodal foundational model, which, during pretraining, learns the universal laws of the physical world, targeting universal reuse across scenarios and countries. Zhuoyu is transforming from a solution provider to a model company.
ThunderSoft and Zhida Chengyuan also showed their cross-domain OSes at the auto show. The software layer above the chip is as crowded as the chip layer.
Some are already ahead. Banma Intelligence, based on Qualcomm 8397 and 8797, deploys the Tongyi Qianwen large model in cars, doing inference, decision, and planning all on-device, with 90% of scenarios usable without network and user data staying in-car. This is microcosm of software-layer differentiation within the Qualcomm camp: with the same chip, whose model runs best and whose experience is most complete can break out of the Tier 1 melee.
Su Qing says everyone is chasing a “golden bullet.”
But it’s been nearly 20 years since the iPhone, and no Android phone has matched its smooth interaction. Automotive safety concerns human life and death, so simply deploying a large model can’t solve cockpit-driving integration.
No silver bullets, so what matters? Yu Kai’s answer is ecosystem. He says open ecosystems are for strong players—empowering others on chips, supply chain, and software is essential.
Horizon President Zhu Wei has a more plain view. He says the key is to make every partner who’s worked with Horizon once willing to do it again. He cites the “Apple supply chain”: the iPhone opened a new world and supply chain partners naturally grew with it, thriving because value creation was big enough. Ecosystem isn’t a goal—it forms naturally through value creation.
The industry’s direction already has precedents. Yu Kai says the smartphone era had only two chip suppliers plus Apple’s self-development. Zhu Wei adds from a business logic: For big business (To B), ultimate players are few—“Once there’s a better choice, no one wants second best.”
With five Tier 1s in the Qualcomm camp differentiating above the same chip, it’s clear that the gap isn’t in the chip itself, but in the OS and model layers above. As the chips go from two to one, software becomes more valuable.
Who Stays at the Table
Chip companies and Tier 1s are fighting for definition rights, but automakers aren’t just waiting.
NIO’s central computing platform Cedar ADAM has upgraded from cockpit-driving split architecture to multi-domain fusion. The Shenji chip runs not only smart driving but also some cockpit functions, powering cross-domain compute scheduling. Strictly speaking, it’s not single-chip cockpit-driving integration, but the direction is the same: break domain boundaries and let compute flow freely. Self-developed Shenji’s single-chip mass-production cost is about 45% lower than NVIDIA’s Orin-X, saving about 10,000 RMB per car; the cost reduction from in-house chips surpasses that from supplier solutions.
Xpeng last year merged its Autopilot Center and Smart Cockpit Center into a new “General Intelligence Center.” Organizational integration precedes product integration, typically heralding product convergence.
Suppliers are being squeezed from both sides. Above, automakers are moving down towards the chip layer; below, Ouye Semiconductor teamed with Freetech, UNISOC to launch the “Fuxin No. 1” using domestic chips to bring cockpit-driving integration below 100,000 RMB, while ECARX used its own “Longying No. 1” for a single-chip three-domain fusion platform (cockpit-driving-parking). Cockpit-driving integration may become a standard faster than expected.
Volume growth itself has bottlenecks. Global storage giants are shifting capacity toward HBM, and automotive-grade storage faces structural shortages. Su Qing says cockpit-driving integration can reduce DDR usage, but only if chips are integrated enough—if not, storage bottlenecks will constrain volume. Safety certification is another slow variable; ASIL-D level end-to-end validation may take longer than chip R&D itself.
Ultimately, cockpit-driving integration competition isn’t about chip specs, but organization ability and engineering culture.
Cockpit-driving integration was originally meant to simplify: two chips become one, BOM is reduced, wiring lessened, cooling simplified. But as hardware simplifies, software competition is actually intensifying. Five Tier 1s run in five directions with the same chip; each doesn't want to be just a hardware integrator, but whether they can truly become software companies is still uncertain. Integration simplifies the architecture but accelerates elimination.
Industry forecasts give cockpit-driving integration a five-year growth window. But the real window may be much shorter than five years. The few thousand RMB saved in BOM costs decides not just a few percentage points in profit margin, but whether features like urban NOA can penetrate from 300,000 RMB to 150,000 RMB, and then down to 100,000 RMB.
When universal solutions lower the threshold to below 100,000 RMB, companies without mass-production delivery are not just catching up—they’re already out.
Risk Alert and DisclaimerThe market has risks, and investment needs caution. This article does not constitute personal investment advice and has not taken into account individual users' specific investment objectives, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article suit their own situation. Investments based on this are at your own risk. ```