Series: TrueView

Intelligence remains an enduring growth proposition in the tech industry, and firms wielding general-purpose large models often view scenario penetration as the ultimate solution. From booking flights, hotels, train tickets, or milk tea with a single voice command to end-to-end travel connectivity, AI agents attempt to break through the App ecosystem barriers solidified by over a decade of mobile internet development using natural language interaction. Yet, as the battleground shifts from palms to wheels, the smart cockpit—once seen as a blue ocean of incremental growth—faces thicker, more entrenched industrial power barriers than those in the smartphone realm. While algorithms define the upper limits of user experience, profit-sharing mechanisms determine the depth of commercialization.

This article focuses on analyzing:

1. Why do automakers and AI firms collaborate eagerly yet remain at odds?

2. How do diverging routes for in-vehicle large models determine cockpit discourse power?

3. What anxieties lie behind automakers' tightening control boundaries?

4. Why does the ultimate success of large models in vehicles hinge on industrial rules?

Content/Liu Ping

Editor/Yong E

Proofreader/Mang Fu

Huawei's recent automotive controversies have been nothing short of dramatic.

In contrast to the post-holiday "brake pedal fracture" incident involving the Zunjie model, a pre-holiday split-and-reunion drama unfolded between Huawei and Seres' Aito over half a month.

On September 15, Huawei and Seres' Aito announced an adjustment to their cooperation model, with Seres taking the lead in product definition, design, brand marketing, channel retail, and service systems, while Huawei's consumer business shifted to an enabling role. The outside world quickly labeled this a "breakup," but this independence dream lasted only fifteen days.

On September 30, Yu Chengdong and Zhang Xinghai reunited in Shenzhen to sign a new five-year cooperation agreement, establishing a dedicated Aito business team to continue upgrading the brand under an exclusive operation model.

Just three days after Huawei and Seres' Aito "parted ways," SAIC Roewe announced the deep integration of Doubao Assistant with its Roewe Marvel R, claiming it as the first AI-native concept car in the Marvel series developed in collaboration with Volcano Engine and the world's first mass-produced model equipped with the Doubao cockpit assistant.

Not just Doubao—mainstream large model providers like DeepSeek, Alibaba's QianWen, and Baidu's ERNIE are accelerating their vehicle integrations, sparking a collective race around in-vehicle cockpits.

In reality, Doubao and its peers face the same challenge: as AI large models increasingly become the core competitiveness (competitive edge) of smart cockpits, how much power and industrial status are automakers willing to cede to enable deep large model integration? How should responsibilities and benefits be divided and bound in their collaborative projects?

Huawei and Seres' "whirlwind reunion" and JAC Motors' two-day stock plunge merely highlight the core contradictions of large model vehicle integration. For large model firms flocking to the automotive sector, they face the same industrial power struggle.

The true barrier to large model vehicle integration lies not in model capability but in the redivision and contention of industrial discourse power.

Part.1 Fierce Competition in the Second Half

A widely circulated maxim in the automotive industry states: electrification is the first half, and intelligence is the second half. Multiple automakers' founders and executives have reiterated this viewpoint.

Electrification dismantled the century-old technological moats of the internal combustion engine era, forcing all automakers back to the same starting line. However, electrification merely reset the track; intelligence truly determines the rankings.

Even in the early days of electric vehicle (EV) adoption, intelligence became a core bargaining chip for automakers to seize definitional authority, albeit after a lengthy detour.

In 2012, Tesla's Model S debuted with a 17-inch vertical touchscreen, sparking cockpit evolution. This model drastically reduced physical buttons, integrating nearly all vehicle functions—navigation, multimedia, climate control, and driving modes—into the in-vehicle system. More importantly, it reconstructed the underlying logic of human-vehicle interaction, transforming cars from mechanical assemblies of independent sensors into intelligent terminals capable of continuous software evolution.

However, the industry soon fell into a trap of form over substance. Screen sizes ballooned from single displays to panoramic screens, even featuring Mercedes-Benz's EQS with a 56-inch "screen dominance" exceeding home TVs—yet the in-vehicle interaction experience remained lacking.

For too long, so-called intelligence remained stuck at visual-level luxury packaging.

Meanwhile, the advanced driver-assistance systems (ADAS) sector gradually hit a bottleneck. After multiple iterations, players' algorithms, chips, and LiDAR stackups grew increasingly similar, with no significant experience gaps emerging and a clear distance remaining from safe, reliable autonomous driving.

Automakers urgently needed a new path to establish differentiated competitiveness, and maturing AI large models exactly (happened to) fill the cockpit's "brain" gap. Automakers lacked an intelligent, interactive core, while AI lacked a physical carrier for implementation—a perfect match.

Baidu, an early mover, was the first in China to push large models into mass-produced vehicles. In February 2023, when announcing ERNIE Bot, Chang'an, Geely, Great Wall, VOYAH, and Dongfeng Nissan collectively announced their integration.

However, this resembled a concept-first positioning game. ERNIE Bot had not yet launched, but automakers were already lining up to endorse it, unwilling to miss the narrative dividends of being first.

In October 2023, the Zeekr 01 debuted as the "world's first intelligent vehicle with a large model," but its market performance fell far short of expectations. Likely dragged down by Zeekr 01's poor sales, the aforementioned collaborations faded into obscurity.

Due to strategic realignments, Baidu shifted focus to autonomous driving (Apollo Go and Robotaxi), while Zeekr collapsed amid funding shortages and shareholder capital freezes.

Other large model providers adopted a more cautious stance, awaiting a clearer product form (form).

In early 2025, DeepSeek sparked a wave of automaker integrations, with over 20 brands—including VOYAH, Geely, Zeekr, IM Motors, and Baojun—announcing fusion with DeepSeek.

These collaborations mostly leveraged DeepSeek's open APIs to enhance voice interaction capabilities for weather queries, restaurant recommendations, and long-distance planning, falling short of true customization and deep integration.

Overall, this surge represented a collective move with greater marketing significance than substantive experience upgrades. Automakers capitalized on DeepSeek's hype to promote AI concepts without granting vehicles full intelligent agent capabilities.

The real turning point came in July 2025, when Tesla formally integrated Grok into its vehicles, redefining in-car interaction with natural dialogue, deep retrieval, and emotional expression capabilities. That year became known as the "first year of large models in vehicles."

By 2026, automakers began abandoning post-production add-ons, shifting toward native deep integration of large models. Giants rapidly deployed solutions, and the sector officially entered a substantive racing phase.

Part.2 Routing Divergence in the Cockpit

Data from Gaogong Intelligent Auto Research Institute shows that in 2025, 9.4475 million passenger vehicles (excluding imports/exports) in China's market shipped with built-in voice interaction large models (excluding App-based solutions), up 118.90% YoY. Among them, DeepSeek, Doubao, and iFLYTEK Spark ranked as the top three third-party suppliers. In Q1 2026, penetration of new vehicles shipped with large models neared 50%.

Behind this data surge lies fierce competition among large model providers for vehicle integrations. However, varying deployment speeds reflect not just technical capability gaps but differing strategic logics.

DeepSeek pursues a lightweight, cost-effective, and rapid integration route, offering automakers a quickly deployable, affordable dialogue engine to bridge intelligence gaps within short cycles. This model's strong adaptability and low barriers made it a popular entry-level choice.

Doubao, backed by ByteDance's ecosystem, aims for the Vehicle Control Center (vehicle control hub) track (sector), positioning itself as the underlying brain for vehicle intelligence. Unlike traditional models where automakers build vehicles and then integrate large models, Doubao intervenes during vehicle definition stages, using AI capabilities to retroactively shape hardware configurations and interaction logics.

This positioning difference is evident in Tesla's China models: DeepSeek handles open-ended interactions like chatting, weather, and news, while Doubao manages vehicle control commands such as navigation, media playback, and climate adjustment.

However, Doubao does not solely bet on deep customization. To pursue scalability, it also offers standardized AI cockpit suites, allowing automakers to directly procure Doubao cockpit assistant solutions.

Volcano Engine data from April 2026 shows that smart vehicles equipped with Doubao have surpassed 7 million units, covering over 50 brands and 145 models, with daily cockpit interactions exceeding 30 million.

BAT firms follow distinct paths.

Baidu centers on ERNIE Bot and Apollo Go, supplemented by strategic synergy with Horizon Robotics (Yu Kai, a former Baidu executive, and Baidu as a major Horizon shareholder) in a classic investment-plus-business alliance.

Alibaba emphasizes ecosystem synergy. In September 2026, BYD launched the "Super Intelligent Agent Didi Xia," integrating QianWen into its cockpit cloud hub for intent recognition, task judgment, and service distribution while connecting to Alibaba ecosystems like Fliggy and Taobao Snap Shop. Additionally, Banma Intelligent, Alibaba's cockpit ambitions carrier, is preparing for an independent IPO.

Tencent adopts the lightest approach, eschewing model sales for ecosystem exports. Partnering with Jueyue Xingchen to tackle in-vehicle scenarios, Tencent provides music, mapping, and mobility service interfaces along with cloud computing power, while Jueyue Xingchen contributes end-to-end voice large models and multimodal understanding capabilities.

Their collaboration also enjoys clear capital and business support. Tencent Cloud and Jueyue Xingchen announced deep integration of Jueyue's foundational large models with Tencent's content and application ecosystems to co-create intelligent cockpit AI agents. Capitally, Jueyue Xingchen secured nearly $2.5 billion in financing on May 11, 2026, accelerating preparations for a Hong Kong IPO, with Tencent participating in three consecutive funding rounds.

Huawei adheres to a "no car manufacturing" strategic bottom line (bottom line), directly exchanging technological density for industrial discourse power. Leveraging its Qiankun Intelligent Driving and HarmonyOS Cockpit as core levers, it offers tiered cooperation models tailored to automakers of varying scales.

This approach demands intense R&D investment. In 2026 alone, Huawei will allocate over ¥18 billion to automotive intelligence solutions, including ¥7–8 billion for cloud computing power.

This explains the underlying conflict in Huawei's recent controversies: when a tech provider's capabilities grow strong enough to redefine a vehicle, the automaker's manufacturing sovereignty and brand agency inevitably clash with the tech firm's industrial discourse power.

For smart cockpit solution providers, in-vehicle cockpits represent not just high-value, high-stickiness traffic entrances but rare B2B paid markets. However, this is no open Incremental market (incremental market); switching clients does not resolve all issues.

Training a model is merely the first step.

Part.3 Inevitable Power Barriers

At 2:25 AM on February 25 this year, a man driving a Lynk & Co electric vehicle on a highway attempted to turn off the interior reading light via voice command. The system failed to recognize the instruction, shutting off all vehicle lights instead. A second voice attempt to restore the lights also failed, causing the vehicle to crash into a guardrail. Fortunately, no injuries occurred.

AI errors on smartphones typically remain at the experience level, with inaccurate responses or delays. But when large models control vehicle functions, every misjudgment escalates into a safety risk.

This uncertainty highlights a deeper contradiction: cockpit intelligence relies on continuous feeding of driving data, requiring automakers to cede some data sovereignty—a sensitive nerve for automakers.

Currently, automakers widely adopt tiered permission controls and capability isolation. For example, in Doubao's collaboration with Roewe, clear permission zones are set: AI-regulated operations like climate and lighting fall into the "color zone"; vehicle state-related functions like seat adjustments during driving enter the "gray zone" under bottom-layer rules; core driving controls like braking and steering belong to the "black zone," off-limits to cockpit large models.

Automakers' tightening permissions and strict boundaries appear safety-driven but stem from the prolonged "soul debate" in the intelligent driving era. Their deepest anxiety is not technology itself but the fate of becoming mere hardware assemblers. Safety is the baseline; control rights represent deeper power struggles.

Huawei and Seres' recent split-and-reunion drama amplifies this game theory (struggle). Aito has been a sales pillar for Huawei's HI Mode, contributing 420,000 units in 2025 (over 70% of HI Mode's total). The dramatic reversals were not about cooperation breaking down but both sides reassessing their accounts.

Channel data shows that of Aito's ~1,146 brand-authorized touchpoints, 637 (55.6%) are located in Huawei consumer electronics stores. A complete channel split would halve Aito's sales touchpoints, forcing Seres to build its own channels at enormous customer acquisition costs. With Huawei's cooperation expanding and its halo diluting, Aito—its flagship project—cannot afford setbacks.

Analysts note that under previous models, Seres held weaker discourse power, "risking reduction to a contract manufacturer, which is not its desired growth path." Clearly, automakers want control, but possessing the capability to uphold it is another matter.

This mirrors smartphone logic. The newly released Doubao Phone II still garners attention but falls short of expectations due to mainstream App ecosystem barriers. Mobile internet's profit structure has long solidified; even powerful algorithms cannot bypass ecosystem walls.

Today, large models face similar challenges in vehicles, confronting centuries of automotive industrial inertia. Automakers' core roles in defining products, controlling supply chains, and holding user data will not easily yield to a single large model.

Large models can transform interaction experiences, reshape product selling points, and even create new usage scenarios, but they struggle to directly rewrite industrial power distributions. Algorithms determine technological heights, but profit-sharing mechanisms determine commercialization depths.

From smartphones to vehicles, from Doubao's grand mobility ambitions to Huawei-Seres' cooperation tug-of-war, large models' every breakthrough challenges entrenched industrial power structures. Mobile ecosystem walls remain unbreached, and automotive permission barriers run equally deep.

The ultimate victory in this in-vehicle intelligence race hinges not on who launches a stronger model first but on who pioneers industrial rules balancing cost, experience, and multi-stakeholder interests.

This time, large models face not an empty entrance but a cockpit already dense with power dynamics.

END