After the democratization of smart driving, why can't cockpit compute be democratized?

In early 2025, BYD's "God's Eye" launched intelligent driving for all, with high-level intelligent driving rapidly extending down to models in the 100,000 yuan class and even below, officially kicking off the era of intelligent driving for all.

But making intelligent cockpits accessible to all has run into a new dilemma. First, the question of whether smart cockpits exist at all has long been settled. Data from Gasgoo Automotive Research Institute shows that in August 2026, the penetration rate of smart cockpits in domestic passenger vehicles reached 87.5%, approaching full popularization. The smart cockpit configuration rate for models under 100,000 yuan rose from 52% to 70%. Large screens, voice control, and internet connectivity are no longer exclusive to high-end cars; basic cockpit hardware has already completed its price descent.

But hardware popularization does not equal equalized experience. At present, high-end chips such as Qualcomm 8295 are concentrated in models above 200,000 yuan, while mainstream models in the 100,000-plus yuan range and below still mainly use Qualcomm 8155.

Not long ago, the intelligent driving track offered a new direction for solving this problem: Momenta proposed "large models defining chips," achieving joint definition and synchronized iteration of algorithm models and chip architecture. This also confirms that how to make computing power keep pace with model iteration has become a core question that intelligent vehicles must answer.

More critically, cockpits are hard to replicate the path of intelligent driving for all. The functional boundaries of intelligent driving have converged, and computing power requirements can be calculated in advance, whereas the capabilities of cockpit large models are still expanding at high speed, with new features landing every quarter, making it difficult to predict future computing power consumption during vehicle development.

Against this backdrop, simply pre-installing more computing power at the factory is no longer a reliable solution. What the industry truly needs to build is a computing power upgrade system driven by intelligent upgrades. The main line of competition in smart cockpits has thus shifted from a contest of fixed computing power to a game of evolvable computing power.

Device-cloud collaboration has long been an industry consensus. The cloud handles complex reasoning and cross-platform ecosystem calls, while the device side handles always-on perception and millisecond-level response, while also handling interactions in weak-network and no-network scenarios, keeping sensitive data such as voice and occupant status in local closed-loop computation to meet privacy protection and data compliance requirements.

Cockpit intelligence has traced a rising curve: from the early 1.0 era of intelligence that relied on fixed commands and one-question-one-answer interactions, to today's 4.0 era of proactive service based on memory and decision-making. Every segment of this curve can only hold if device-side models truly run in the vehicle.

▲Schematic diagram of cockpit intelligence evolution from 1.0 to 4.0

Device-side computing power has become the core foundation for high-quality intelligent cockpit experiences, and the high-frequency, personal interaction experiences in the cockpit must be rooted in local device-side computing power. If device-side computing power iteration cannot keep up with the pace of model updates, it will ultimately fall flat in real-world vehicle use scenarios.

As more and more high-frequency interactions and cockpit control tasks move to the local side, the inherent limitations of device-side hardware become an unavoidable problem.

Unlike cloud computing power, automotive-grade cockpit chips are strictly constrained by power consumption, heat dissipation, hardware cost, and vehicle power supply conditions. Chip computing power ceilings are fixed and available memory resources are limited, making it impossible to load larger-scale models without limit, while high-load inference brings power consumption and heat pressure that affect cockpit system stability.

In addition, for device-side intelligence, the smart cockpit industry faces a set of contradictions: the iteration cycle of large models is 3-4 months, while automotive cockpit SoCs require at least six months to a year from selection to mass production.

Cockpit models are evolving rapidly, but once the in-vehicle hardware is finalized, computing power is essentially locked in, and subsequent software OTA can hardly break through the hardware boundary, creating a mismatch dilemma where "computing power supply cannot keep up with model evolution."

In the intelligent driving track, the boundaries of scenarios such as driving, parking, and active safety are becoming increasingly clear, and automakers can configure hardware as needed and reserve an appropriate amount of computing power headroom.

But cockpit large models follow an entirely different logic. Model capability boundaries continue to expand every quarter, and when planning models this year, it is hard to predict how much computing power the cockpit interactions and Agents will need to carry two years from now.

Such problems turn automakers' computing power pre-installation into a "gamble" on the future. If too little computing power is reserved, new cars will face lagging capabilities upon delivery, but if too much is reserved, the vehicle's cost rises out of thin air.

This contradiction is further amplified by model price positioning. High-end models generally have ample profit margins and can pre-install more generous computing power, but affordable models are constrained by cost and have no extra profit margin to pay for uncertain future computing power needs.

The same technical challenge ultimately evolves into market divergence, with continuously iterating smart cockpit experiences often confined to the high-end price range.

For the mandatory question of "how computing power can keep up with model iteration," an open-route solution has emerged in the industry: no longer binding to a single fixed chip, but making computing power a pluggable, independently upgradeable module.

The AI Box created by ModelBest in partnership with Wutong Technology, Intel, and other partners is precisely a mass-production sample of the open route.

▲AI Box device-side cockpit intelligence solution

As a low-intrusion, configurable modular computing power node, it provides users with more powerful device-side intelligence while offering the in-vehicle computing platform an independent, pluggable computing power and intelligence solution, providing a mass-producible answer to the mismatch between computing power and models.

In addition, it supports OTA upgrades, allowing OEMs to continuously provide users with intelligent applications and services based on cutting-edge device-side large models, extending the vehicle's intelligent lifecycle.

▲The continuously upgradeable capabilities brought by AI Box

Currently, AI Box is officially delivered as an option on the Changan Deepal L06, equipped with a 4B VLM model plus multiple device-side models. The project went from its July debut to front-installation mass production in October, with an overall cycle of only three months.

▲Changan Deepal L06 equipped with AI Box

The open route is better suited to OEMs that lack in-house chip development capabilities but still want to preserve their competitiveness in cockpit intelligence, as well as to export models, where stringent overseas data compliance requirements further amplify the value of on-device local computing.

The open route is a bet that, in a cycle of rapid model iteration, flexibility is more "valuable" than certainty.

The open route of modular computing power upgrades has already produced its first mass-production sample, but turning a single case into an industry-wide paradigm still requires confronting three industry variables.

In the automotive market, the intelligent optional-equipment model carries uncertainty: whether consumers are willing to pay extra for subsequent intelligent upgrades still needs long-term market validation.

At the same time, third-party external computing power modules must clear the automotive-grade reliability threshold and the definition of fault liability, which also raises the question of whether OEMs are willing to open up cockpit hardware resources.

Whether to hand incremental computing power to external modular solutions tests automakers' openness.

The dilemma brought by fixed computing power is not unique to the automotive industry; robots, smartphones, and IoT devices all face the same bottleneck.

And the parameter race that simply compares TOPS computing power has already hit a bottleneck under the constraints of smart cockpit conditions. The large models with tens of billions of parameters touted at major manufacturers' launch events are far smaller in real scale when it comes to cockpit models actually achieving front-installation mass production and hitting the road; most on-device models actually running on vehicle head units are concentrated in the 0.9B–3B range, and paper specifications are hard to translate into real vehicles.

The density law proposed by researchers from Tsinghua University and ModelBest provides a new competitive yardstick: the capability density of large models roughly doubles every 3.5 months, and the computing power required for the same level of intelligence keeps declining.

▲The density law proposed by researchers from Tsinghua University and ModelBest

Once modular upgrade mechanisms break the fixed shackles of hardware computing power, the industry's competitive focus will naturally shift from simply comparing paper computing power scale to a contest over computing power utilization efficiency. Parameter comparisons remain only on launch-event slides, while model density is the core that determines the gap in mass-production experience.

The density race does not require believing anyone's slogans; leaderboards and mass-production models will vote for this path.

04.

Conclusion: On-device intelligence makes cars intelligent partners

In the smart cockpit field, the equality of "whether it exists" has basically been achieved, yet AI cockpit experience and continuous evolution capability remain locked in price tiers by chip grades and pre-embedded computing power. What cockpit intelligence needs is not greater pre-embedded computing power, but a "computing power upgrade mechanism."

Smart driving equality solved "whether it exists"; the upgradable computing power mechanism needs to solve bringing "continuously evolving cockpit intelligence" down to more models—letting computing power keep up with intelligence, and not belonging only to high-end cars. This is the next ticket to cockpit equality.

At the same time, automotive intelligence is evolving along two main lines in both directions: toward the vehicle side, it is becoming more deeply integrated as the cockpit AI computing power hub, driving changes in the vehicle's electrical and electronic architecture. Toward the user, it is becoming a cross-device personal intelligent carrier, with future support for plug-and-play, cross-vehicle memory migration, and connectivity across vehicles, PCs, phones, home devices, and other terminals.

When intelligence happens on the device side closest to people, the car will no longer be just transportation hardware, but will form a personalized intelligent system with perception, memory capabilities, and implemented privacy protection.

This article comes from the WeChat public account "CheDongXi"; author: R; editor: Zhihao; published with authorization from 36Kr.