China Intelligent Driving (ADAS/AD) Supply Chain: Market Scale and Competitive Landscape 2026 Deep Research Report

Abstract

In 2025, the front-fitted penetration rate of intelligent driving in Chinese passenger vehicles exceeded 60%, with high-level Navigate on Autopilot (NOA) reaching approximately 22% — and the leap from highway NOA to urban NOA is accelerating. Three landmark events converged: end-to-end large models displacing rule-based code, the first L3 conditional-automation pilot approvals, and Tesla FSD Supervised officially launching in China. Together they have pushed the sector into a dual inflection point of technology and commercialization.

Starting from a full supply-chain panorama covering perception, decision, and actuation subsystems, this report deep-dives into five core sub-tracks — compute chips (Horizon Robotics, Black Sesame Intelligent, Huawei Ascend, NVIDIA Drive Thor), domain controllers (Desay SV, JOYSON Electronics), algorithm software (Haomo.AI, Momenta, Huawei ADS 4.0, XPENG XNGP, NIO NAD, Li Auto AD Max), HD maps (AutoNavi, Baidu, NavInfo), and L4 Robotaxi (Baidu Apollo, Pony.ai, WeRide) — alongside the competitive landscape and financial performance of each. Seven mid-stream industrial clusters are mapped: Beijing, Shanghai Jiading, Suzhou, Wuhan, Shenzhen, Guangzhou, and Chongqing.

On the forecast side, the report quantifies the outlook: NOA penetration surpassing 50% by 2030, L3 commercialization expanding across multiple cities, L4 Robotaxi scaling to mass operation, and the overall intelligent driving supply chain market targeting RMB 400 billion. Primary risks — Tesla FSD market impact, US export controls, algorithmic safety incidents, and uncertain regulatory cadence — are systematically assessed.

I. Definitions, Classification and Full Supply-Chain Panorama

1.1 The SAE/GB Automation Level Framework

Understanding the intelligent driving supply chain begins with the grading framework. The international benchmark is SAE J3016, defining L0 through L5; China's counterpart GB/T 40429—2021 aligns in core logic while adapting terminology and certain boundary definitions to local context.

L0 — No Automation: The driver executes all dynamic driving tasks; the vehicle may offer warnings or brief interventions (e.g., AEB).

L1 — Driver Assistance: Sustained support in either longitudinal (ACC) or lateral (LKA) control, but the driver must be ready to take over at any time.

L2 — Partial Automation: Simultaneous longitudinal and lateral control; the driver must continuously monitor and be prepared to take over. China's L2 front-fitted rate exceeded 60% in 2025, making it the de-facto standard.

L2+ (Enhanced L2): Industry shorthand for highway NOA and urban NOA. Technically still L2, but perception and decision complexity is far higher than basic L2. NOA front-fitted penetration reached approximately 22% in 2025.

L3 — Conditional Automation: Within a defined Operational Design Domain (ODD), the system fully manages driving; the driver must resume control when requested — not necessarily monitor continuously. This is the first level where a driver may legally look away. In December 2025, China issued its first L3 approvals for two models in Beijing and Chongqing, marking China's formal entry into L3 commercial exploration.

L4 — High Automation: Within a specific ODD, the vehicle drives fully autonomously with no expectation of the driver intervening. Robotaxi is the flagship L4 commercial scenario; Baidu Apollo (Luobo Kuaipao), Pony.ai, and WeRide all operate at this level.

L5 — Full Automation: Any scenario, any condition; still in research, far from mass production.

1.2 The ADAS Function Spectrum

Beyond SAE/GB levels, the market more commonly distinguishes products by function name:

• ACC (Adaptive Cruise Control): Longitudinal following via millimeter-wave radar or camera; the flagship L1 function.
• LCC (Lane Centering Control): Adds lateral control to ACC, keeping the vehicle centered; together with ACC = basic L2.
• NOA (Navigate on Autopilot): Full-journey highway or urban driving assistance from point A to B, covering lane changes, on/off ramps, and complex intersections. Highway NOA and urban NOA differ enormously in technical difficulty.
• AEB (Autonomous Emergency Braking): Passive safety; automatically brakes when a collision is unavoidable. Now mandatory in Euro NCAP and scored in China's C-NCAP and C-IASI.
• AVP (Automated Valet Parking): Full hands-off parking in car-park scenarios; an early L4 commercial form in enclosed environments. Several OEMs have shipped it.

1.3 Three Subsystems: Perception, Decision, Actuation

Regardless of grading or function names, the physical architecture always revolves around three layers:

Perception (Sensors): The vehicle's "eyes and ears." Key sensors include cameras, 77 GHz millimeter-wave radar (3D/4D), LiDAR, and ultrasonic sensors.

Decision (Compute): The vehicle's "brain." Includes the domain controller (SoC compute platform) and algorithm software: perception fusion → prediction → planning → control. End-to-end models are progressively merging these four stages into a single neural network.

Actuation (Controls): The vehicle's "hands and feet." Steer-by-wire, brake-by-wire, throttle-by-wire. The degree of electronic actuation directly determines whether high-level autonomy is achievable.

1.4 Full Supply-Chain Map

Along the perception → decision → actuation vertical, the principal participant tiers are:

Upstream (core components): Compute SoCs (Horizon, Black Sesame, Huawei Ascend, NVIDIA, Qualcomm, Mobileye); CIS image sensors (Sony, Samsung, OmniVision); LiDAR (Hesai, RoboSense, Innovusion, Luminar); mmWave radar chips (TI, Infineon, NXP, Calterah); HD maps (AutoNavi, Baidu, NavInfo).

Midstream (integration and software): Domain controllers (Desay SV, JOYSON Electronics, HiRain Technologies, Neusoft Reach, Hinge); algorithm software (Haomo.AI, Momenta, Yuanzhi Qixing, Zhuoyu Technology, Huawei Vehicle BU); data & simulation (51World, Huawei Cloud HUMO).

Downstream (OEMs and mobility): Passenger vehicle brands (BYD, Huawei AITO/Luxeed/Stelato, XPENG, NIO, Li Auto, Changan, Geely, GWM, GAC, FAW); L4 operators (Baidu Apollo, Pony.ai, WeRide, DiDi Autonomous Driving, Waymo).

1.5 Core Contradiction and Evolution Logic

The core contradiction of China's intelligent driving supply chain today: technology routes have not converged, yet capital and resources are accelerating toward a handful of players.

This manifests differently at each supply chain layer. At the chip layer, the contest between "pure vision + large compute" (Tesla FSD) and "multi-sensor fusion + moderate compute" (Huawei, XPENG, NIO) determines the LiDAR market ceiling. At the algorithm layer, the replacement of modular systems by end-to-end large models is accelerating. At the OEM layer, the "full-stack self-developed" (XPENG, Li Auto, NIO, Huawei's AITO model) vs. "integrated procurement" (traditional OEMs) competition is shifting from a pure technology contest to a data-flywheel contest.

1.6 Domestication Progress Across Three Layers

In 2025, domestication rates differ dramatically across the three layers:

Perception layer: Highest domestication. LiDAR (Hesai, RoboSense), mmWave chips (Calterah), camera modules (Sunny Optical, OFILM), domestication already exceeds 50% in most vehicle models. The bottleneck is high-end CIS — Sony and Samsung still dominate front-camera high-end sensors; OmniVision (now under Goodix) is catching up.

Decision layer: Fastest progress. Horizon Robotics leads in compute chips (largest front-installed shipments in China); Desay SV and JOYSON Electronics dominate domain controllers. GPU for training — NVIDIA H100 restricted, Huawei Ascend 910B/910C is the main domestic substitute; software toolchain maturity still trails CUDA.

Actuation layer: Relatively lagging but improving. Brake-by-wire (Bethel Automotive, NASN Automotive Electronics) gaining traction; steer-by-wire still largely Bosch/ZF due to higher safety criticality.

1.7 Supply Chain Security and Dual-Sourcing Strategy

Since 2023, a structural shift in OEM procurement: from single-supplier to dual-supplier strategies. Triggered by the 2021 Mobileye EyeQ shortage that halted production lines, the principle "at least two suppliers per function" is now standard practice. For intelligent driving chips, this means OEMs maintain parallel design-wins with NVIDIA Orin and Horizon Journey 6 — a structural tailwind for domestic chip makers.

1.8 The Software-Defined Vehicle Trend

SDV architecture consolidates dozens of ECUs into 1–3 high-compute domain controllers connected via Ethernet backbone, enabling OTA updates across all software functions. Key supply chain implications: AUTOSAR framework importance rises; ISO 26262 functional safety becomes a true market barrier; "software as a profit center" (subscription-based smart driving features) emerges as a new business model.

II. Global Landscape and Key Players FY2025

2.1 Global Intelligent Driving Industry Structure

The global landscape: the US leads L4 technology narrative and capital story; Europe controls traditional ADAS supply chains; China is most aggressive in mass-production L2+/NOA and domestic supply-chain substitution.

2.2 Mobileye FY2025

Mobileye (NASDAQ: MBLY) remains the most important single participant in global front-installed ADAS supply chains. EyeQ SoCs represent approximately 92–93% of revenue. In Q3 2025, revenue was $504 million (+4% YoY), with EyeQ volume up ~8%. Key challenges: market share erosion by domestic Chinese chips (Horizon's mid/high-end shipments grew ~5x in 2025), SuperVision competition in China, and EyeQ-to-EyeQ Ultra transition timing.

2.3 NVIDIA Drive Thor

Drive Thor integrates Blackwell architecture, delivering 2,000+ TOPS with fused AD and cockpit computing. 2025 saw a transition from delayed mass production (pushed from mid-2024 to mid-2025) to acceleration, as BYD, SAIC Zhiji, Zeekr, and GAC confirmed design wins. Primary Chinese competitors: Horizon Journey 6P (560 TOPS) and Huawei Ascend MDC. MediaTek's planned pairing with Drive Thor for cockpit chips targets 2026 production.

2.4 Tesla FSD Entry into China

• Early 2025: Tesla quietly pushed limited FSD-like capabilities to Chinese buyers who had pre-purchased FSD;
• May 21, 2026: Tesla officially confirmed FSD Supervised is available in China — the 10th global market;
• Current status: One-time purchase price approximately RMB 64,000; monthly subscription not yet available; full regulatory approval (no-supervision full autonomy) targeted for Q3 2026.

FSD's impact on China's supply chain is bidirectional: technological pressure on domestic algorithm companies on one hand; data-localization regulations (Chinese driving data must remain in Shanghai data centers) creating a natural moat for domestic players on the other.

2.5 Waymo and Global L4 Commercialization

Waymo in 2025: daily rides exceeding 20,000 across San Francisco, Los Angeles, and Phoenix; cumulative mileage over 50 million miles; entering Washington D.C. and Austin. Core advantage: deepest data accumulation, most mature commercial operation, lowest accident rate. Core gap vs. China: hardware cost (~USD 100,000–150,000 per vehicle) and scale.

2.6 China–US–Europe Regulatory Comparison

SAE J3016 (US): Most widely cited but no direct legal force; no unified federal L3/L4 legislation; state-by-state variance.

EU UN-R157 (ALKS): L3 at highway speeds ≤60 km/h (2022); expanded to 130 km/h in several EU countries by 2025. Mercedes Drive Pilot: first commercially certified L3 production system in Europe.

China GB/T 40429—2021 + Permit Management: December 2025 — first two L3 production vehicles (Changan Deepal SL03, ARCFOX Alpha S6) approved; trial plates "渝AD0001Z" and "京AA0001Z" issued. China became the third major auto market after Germany and Japan to commercially pilot L3.

2.7 Key OEM Self-Developed Smart Driving Strategy Comparison

Tesla: Most vertically integrated globally — self-developed FSD chip, FSD algorithm (full stack), and Dojo supercomputer for model training.

BYD: "Integrated procurement + selective self-development" — core compute from Horizon Journey, self-developed 4D mmWave radar (via BYD Electronics), with HUAWEI ADS for premium sub-brands.

Huawei AITO ecosystem: Full solution supply (Ascend MDC + ADS algorithms + Qiankun sensor suite + HD Map); integrated package creates very high switching costs for partner OEMs.

Volkswagen in China: Strategic investment in XPENG (~5% stake, 2023); joint development agreement to import XNGP technology for China-market models from 2026.

Toyota + Momenta: Strategic investment aligning Toyota's global OEM scale with Momenta's data-flywheel algorithm; most significant China-algorithm-company × Japan-OEM collaboration.

2.8 From "Aid Recipient" to "Exporter": China's Positioning Shift

Twenty years ago, China was an "aid recipient" from Bosch, Continental, Denso. Today China is a systematic exporter: Hesai LiDAR ships to US/European front-installed markets; Desay SV domain controllers win European OEM design-ins; Momenta licenses algorithms to European OEMs via Bosch; BYD exports complete vehicles with domestic smart driving systems to Europe and Southeast Asia.

2.9 HD Map's International Barriers and China Specifics

Only licensed Chinese entities (NavInfo, AutoNavi, Baidu Map, etc.) can operate HD map collection in China — a compliance moat for domestic HD map companies and a barrier for foreign smart driving solutions. Meanwhile, the US, EU, and Japan are tightening restrictions on Chinese-supplied vehicle components collecting road data on their territory, creating complex non-technical export barriers for Chinese intelligent driving supply chain companies.

III. PEST Analysis: Political, Economic, Social, Technological Environment

3.1 Policy Environment

Key Policy Timeline:

• 2015: "Made in China 2025" lists intelligent connected vehicles as a top-10 priority;
• 2020: NDRC + 10 ministries release "Innovation and Development Strategy for Intelligent Vehicles";
• 2021: MIIT regulates smart driving function type-approvals; GB/T 40429 and PIPL/DSL enacted;
• 2023: "Four-ministry pilot permit document" allows L3/L4 vehicles on public roads in designated areas;
• September 2025: "Automotive Industry Steady Growth Plan 2025–2026" includes L3 permits as a stimulus tool;
• December 2025: First two L3 models approved — Changan Deepal SL03 (Chongqing) and ARCFOX Alpha S6 (Beijing).

By H1 2026, over 50 cities have issued autonomous driving pilot demonstration policies or local regulations.

3.2 Economic Environment

New energy vehicle sales in China exceeded 12 million in 2025 (>45% penetration), creating the carrier for intelligent driving "price-for-volume" dynamics. Highway NOA solution cost fell from RMB 5,000–8,000 in 2021 to RMB 2,000–3,500 in 2025, with some platform solutions below RMB 1,500. LiDAR unit prices declined from ~5,000 to 500–1,500 RMB. Robotaxi UE (unit economics) turned positive at Pony.ai (Guangzhou) with daily revenue of RMB 299 covering all direct operating costs; Baidu Luobo Kuaipao completed ~3.1 million orders in Q3 2025 (+212% YoY).

3.3 Social Environment

Consumer acceptance shows a clear generational divide: over 60% of users under 30 are willing to cede driving control to the system; under 30% of users over 50. Several smart driving incidents since 2021 (especially Tesla Autopilot cases) temporarily dampened active NOA activation rates. Urban/rural infrastructure gap constrains NOA coverage below tier-3 cities and rural roads.

3.4 Technological Environment

End-to-end large models replacing rule code is the defining paradigm shift of 2023–2025. Tesla FSD v12 was the first mass-produced end-to-end system; Chinese players XPENG, Li Auto, Huawei, Haomo.AI, Momenta all transitioned or rapidly advancing to E2E solutions in 2024–2025. VLA (Vision-Language-Action) is moving from research to engineering prototypes in 2025, expected to enter production testing in 2026–2027.

US export controls (EAR) on high-end GPU restricted H100/H200/A100 direct export; A800/H800 subsequently restricted too. Training cluster impact: forced migration to Huawei Ascend 910B/910C; vehicle inference chips (Orin/Drive Thor) not yet directly restricted but supply chain risk awareness sharply elevated.

3.5 China's Intelligent Connected Vehicle Policy System

Three phases: Exploration (2015–2020) → Framework building (2020–2023) → Pilot deployment (2023–present). Municipal-level "policy competition" — with Beijing Yizhuang, Shanghai Jiading, Guangzhou Nansha, Wuhan Economic Zone, Chongqing Liangjiang New Area as five leaders — accelerated Robotaxi commercial experience accumulation.

3.6 Deep Impact of US-China Trade Friction

Beyond chip restrictions, the friction creates: dual supply-chain systems (US-supply vs. domestic-substitute tracks); strategic dilemmas for foreign OEMs (use domestic systems for competitiveness vs. compliance complexity); listing challenges for US-listed Chinese L4 companies; and accelerated domestic AI chip development under "forced domestication" pressure.

3.7 International Investment Ecosystem

Three layers: Strategic OEM investors (BYD, Changan, Geely, GAC holding minority stakes in smart driving companies); foreign participation (Toyota → Momenta, SoftBank → Pony.ai early rounds, Goldman Sachs → WeRide); state capital (National Manufacturing Transformation Fund, Beijing/Shanghai state investment funds filling the gap left by retreating foreign capital since 2023).

IV. China Market Scale and Penetration Rate Overview

4.1 Passenger Vehicle Intelligent Driving Overall Penetration

In 2024, China's L1+ front-installed rate exceeded 65%. In 2025, the estimated rate exceeded 70% — 7 out of every 10 new passenger vehicles ship with some form of driver assistance. Driven by: (1) OEMs treating L2 packages as standard fit across price tiers; (2) NEV expansion, as EV architecture naturally suits smart driving integration.

4.2 NOA High-Level Smart Driving Penetration

2025 NOA front-installed penetration rate: ~22%, nearly double 2024's ~11.7%. At ~18 million annual passenger vehicle sales, this translates to approximately 4 million vehicles shipped with NOA. Highway NOA is 2×+ urban NOA in shipment volume.

4.3 L3 Pilot Cities and Expansion Pace

• First batch (December 2025 approvals): Beijing (ARCFOX Alpha S6, highway segments) and Chongqing (Changan Deepal SL03, congested urban segments)
• Expected next batch: BYD, Huawei AITO, Changan variants targeting 2026 approvals; estimated 5–8 models by end-2026
• City expansion: "Mature one, approve one" principle; Beijing, Chongqing → Shanghai, Wuhan, Shenzhen expected by 2026

Core constraint: not vehicle-side technology (most high-level L2 vehicles already meet L3 hardware requirements) but legal liability assignment, insurance product availability, and accident forensic standards.

4.4 ADAS Total Market Scale

2025 China ADAS & intelligent driving solution market: approximately RMB 200–220 billion, growing 25–30% YoY (20 percentage points above total auto industry growth rate).

Sub-track breakdown: Smart driving chips (RMB 30–40B); domain controllers and system integration (RMB 60–70B); algorithm software and solution licensing (RMB 20–30B); LiDAR (RMB 10–15B); HD maps and data services (RMB 10–15B); other perception sensors (RMB 40–50B).

2030 China market forecast: Multiple institutions project the market exceeding RMB 400 billion, driven by NOA penetration exceeding 50%, L3 commercialization expansion, and L4 Robotaxi scaling; CAGR ~20–25%.

4.5 Penetration Rate Forecast Curve (2025–2030)

Metric	2025 Actual/Est.	2028 Forecast	2030 Forecast
L2+ front-installed rate	70%+	90%+	99%+
NOA front-installed rate	~22%	~40%	~55%
Highway NOA rate	~30–35%	~60%	~90%+
Urban NOA rate	~8–10%	~25%	~40–50%
L3 production models	2 (pilot)	20+	50+
L4 Robotaxi daily orders (national)	~6M	~30M	~100M

4.6 Structural Characteristics of China's Smart Driving Market

First, OEM stratification is extremely pronounced: frontrunners (Huawei AITO, XPENG, Li Auto, NIO) lead technical narrative by 2–3 technology generations over mid-tier and rear OEMs. Second, "pure vision" vs. "multi-sensor fusion" route contest unresolved: Tesla FSD's route has Chinese followers (XPENG XNGP 2.0), but multi-sensor fusion remains the mainstream for most premium solutions. Third, "software as a service" business model still evolving: bundled smart driving in high-spec trims dominates over pure subscription models in China's consumer culture.

4.7 Price-Band Penetration Dynamics

• RMB 300K+: NOA penetration >80%, near saturation;
• RMB 200–300K: 55–65% in 2025, primary growth band;
• RMB 150–200K: The 2025–2026 "main battlefield"; highway NOA projected to exceed 50% by end-2026;
• RMB 100–150K: The 2026–2028 "mass-market descent" opportunity if NOA solution cost drops to RMB 1,000–1,500;
• Under RMB 100K: Penetration forecast to remain minimal through 2030; focus remains on passive safety (AEB).

4.8 NEV vs. ICE Smart Driving Penetration Scissor Gap

NEV smart driving penetration is systematically higher than same-price ICE vehicles due to: (1) naturally EV-compatible electrical architecture; (2) tech-oriented buyer demographics; (3) competitive pressure from EV startups forcing traditional brands to accelerate smart driving on their NEV transitions. Implication: every 1-percentage-point rise in NEV penetration amplifies smart driving supply chain TAM by more than 1-to-1.

4.9 L4 Robotaxi City Data (2025)

• Wuhan: Luobo Kuaipao ~400 vehicles (driverless), ~2,000–3,000 daily orders, world's largest single-city commercial Robotaxi market;
• Beijing Yizhuang: Luobo Kuaipao + Pony.ai, driverless commercial operation;
• Guangzhou: WeRide + Ruqi Mobility, Pony.ai + GAC; daily orders exceeding 12,000 across operators;
• Shanghai, Shenzhen, Chongqing: Various operators with expanding coverage.

National aggregate: ~2,500–3,000 commercial Robotaxi vehicles, ~5–6 million monthly rides, ~10 operating cities.

V. Supply Chain Deep Dive

5.1 Chip Layer: The TOPS Arms Race

Compute level requirements:

• Basic L2: 5–30 TOPS (Mobileye EyeQ4, Horizon Journey 2/3)
• Highway NOA: 50–150 TOPS (Horizon Journey 5, 128 TOPS)
• Urban NOA: 150–500 TOPS (NVIDIA Orin-X 508 TOPS, Horizon Journey 6P 560 TOPS, Huawei MDC 810)
• L4 Robotaxi dev platforms: 500–2000 TOPS (NVIDIA Drive Thor, Huawei Ascend 910B clusters)

Horizon Robotics (HK 9660): 2025 revenue RMB 3.758B (+57.7% YoY); gross margin 64.5%; total Journey series shipments 4.01M units (+38.8%); mid/high-end chip (≥100 TOPS) shipments ~1.8M (+5× YoY); cumulative shipments exceeded 10 million by August 2025 — first domestic company to do so.

Black Sesame Intelligent (HK 2533): H1 2025 revenue RMB 253M (+40.4% YoY); full-year estimate ~RMB 500–600M; A1000 (106 TOPS) in Geely Galaxy E8, Dongfeng AEOLUS 007/008 production vehicles; overseas design wins at record high.

Huawei Vehicle BU / Qiankun: ADS 3.0 deepened mass production; ADS 4.0 launched end-2025; 22+ cooperating production models; Audi A5L Qiankun version marks first luxury foreign-brand OEM partnership.

Calterah (加特兰微电子): China's leading 77 GHz mmWave radar chip supplier; Alps-XD series for 4D imaging radar now in BYD, Li Auto, Geely, Chery production vehicles.

5.2 Domain Controllers: The System Integration Battlefield

Desay SV (002920): 2025 revenue RMB 32.557B (+17.9%); intelligent driving product revenue RMB 9.7B (+116% YoY); China market share in smart driving domain controllers: 21.2% (global: 8.8%) — #1 in both; April 2026 Hong Kong IPO application; estimated valuation ~HKD 63B.

JOYSON Electronics (688326): 2025 revenue RMB 6.848B (+23.59%); net profit ~RMB 99.5M, turning profitable from RMB -550M loss in 2024; ZCU domain controller cumulative deliveries exceeded 2 million units.

HiRain Technologies (HK 1274): Huawei ecosystem's key domain controller integrator; scales with Qiankun ADS production volume.

Foreign Tier 1 (Bosch, Continental, Aptiv): Facing domestic competition pressure; Bosch still important for basic L2 and AEB; Continental strong in mmWave radar sensing; Aptiv in brake/steering actuation.

5.3 Algorithm Software: End-to-End Reconstruction Battle

Haomo.AI: DriveGPT "Xuhu Hairu" model in mass production on GWM WEY brand; launched end-to-end route in 2023, data flywheel exceeds 6B equivalent km.

Momenta: "Flywheel" data-driven route; architecture claimed to reduce training cost 10–100×; partners include Zhiji, Denza N7, Yangwang U8; strategic investment from Toyota, Mercedes-Benz, SAIC, Bangkok Bank.

XPENG XNGP: End-to-end model, 200+ cities for urban NOA, ranked among top domestic urban NOA experiences; deep partnership with Volkswagen group.

Li Auto AD Max 3.0: End-to-end large model OTA update; urban NOA covers all roads in all Chinese cities with road infrastructure; consistently top-ranked in user evaluations.

Huawei ADS: "GOD (General Obstacle Detection) + PDP (Predict, Decide, Plan)" single-stage E2E architecture; ADS 4.0 upgrades multi-modal perception and V2X fusion.

5.4 HD Maps: From "Required" to "Optional"

AutoNavi (Alibaba ecosystem): China's largest navigation map provider; HD map covers 80%+ of national expressways; expanding to urban coverage.

Baidu Map / Apollo Maps: HD map tied to Apollo's full platform; powers Luobo Kuaipao L4 operations with real-time map updates.

NavInfo (002405): Established HD map provider with national surveying license; transitioning to "Lite-HD Map" and "map-as-a-service" to counter mapless NOA disruption; HD Map-as-a-Service model targeting L2+/L3 tier.

Mapless NOA: XPENG XNGP and Huawei ADS 3.0/4.0 have achieved large-scale "mapless" urban NOA rollout, fundamentally disrupting traditional HD map business models.

5.5 Data Flywheel and Simulation Platforms

"Data flywheel" — vehicle fleet collection → cloud training → OTA update → fleet re-collection — is now more defensible a moat than any single algorithm breakthrough. Key simulation platforms: 51World, Huawei Cloud HUMO, CARLA (open source), NVIDIA DRIVE Sim. Auto-labeling tools and federated learning for data compliance are critical competitive dimensions.

5.6 Complete Perception Sensor Landscape

Cameras: Key IC is CIS (CMOS Image Sensor). Global CIS dominated by Sony, Samsung, OmniVision. Camera module manufacturing: Sunny Optical (2382.HK), OFILM transitioning to automotive. Standard configurations: forward mono/binocular main camera + 360° surround-view + DMS interior camera.

Ultrasonic sensors: APA/low-speed collision avoidance standard; ~8–12 per vehicle at RMB 30–60 each. Valeo (France) largest market share; domestic suppliers gaining ground.

LiDAR routes: Mechanical spinning (early development platforms) → Semi-solid MEMS/prism scanning (current mass production mainstream, Hesai AT128) → Solid-state Flash/OPA (next-generation, targets sub-RMB 300).

4D mmWave radar: BYD self-developed 4D mmWave radar ~RMB 100 cost, deployed in Yangwang U8, Denza; industry forecasts >10 million units in 2025 passenger cars alone vs. ~3M LiDAR units.

5.7 Compute Chip Process Node Race

• 7nm: Horizon Journey 6P (TSMC 7nm EUV), NVIDIA Orin (Samsung 8nm)
• 5nm/4nm: Qualcomm Snapdragon Ride Gen 2 (TSMC 5nm), Drive Thor (TSMC 4nm/5nm)
• 16nm and above: Black Sesame A1000 (TSMC 16nm), traditional ADAS chips

Key implication: Advanced process chips (7nm and below) require TSMC or Samsung; SMIC's most advanced node remains ~7nm (DUV multi-patterning) due to EUV equipment export controls, making manufacturing supply chain autonomy an unresolved vulnerability.

5.8 "Algorithm Moat" vs. "Data Moat": Business Logic Evolution

Pre-2022: "Algorithm = moat" — who writes better perception code wins. Post-2022: "Data = moat" — same Transformer architecture trained on 10B km outperforms one trained on 1B km. Structural implications: OEM "fleet data" becomes a strategic asset; data annotation services gain strategic value; GPU cluster infrastructure is critical (Huawei Cloud, Alibaba Cloud, Baidu AI Cloud Ascend clusters as key smart driving infrastructure).

5.9 HD Map "Crowdsourcing" Evolution

Traditional HD map: dedicated collection vehicles, monthly/quarterly update frequency. Crowdsourced HD map: every production vehicle as a mobile collection node; Mobileye's REM system pioneered this globally; AutoNavi, Baidu, NavInfo advancing similar China systems. For pure HD map companies, crowdsourcing is the core defensive strategy against "mapless NOA" disruption.

VI. Key Company Deep Profiles

6.1 Horizon Robotics (HK 9660)

2025 Key Financial Data:

• Revenue: RMB 3.758B (+57.7% YoY)
• Gross margin: 64.5%
• Mid/high-end chip (≥100 TOPS) shipments: 1.8M units (+5× YoY)
• Journey series total shipments: ~4.01M units
• Cumulative shipments exceeded 10M by August 2025 — China first

Product matrix: Journey 2 (2 TOPS), Journey 5 (128 TOPS), Journey 6M (256 TOPS), Journey 6P (560 TOPS). Journey 6 series entered full mass production in 2025; Desay SV IPU06, HiRain Technologies use it as their compute core.

Key customers: BYD (largest), Changan, Chery, Geely, GAC, FAW, plus NIO, Li Auto, XPENG (selected models), Leapmotor.

6.2 Black Sesame Intelligent (HK 2533)

H1 2025 data: Revenue RMB 253M (+40.4%); net loss RMB 762M (includes large equity compensation and R&D); gross margin ~24.8%; overseas design wins at all-time high.

A2000 series (>200 TOPS planned) in development, targeting Journey 6M segment. Global expansion via Bosch partnership to penetrate European OEM supply chains.

6.3 Huawei Vehicle BU / Qiankun

Three cooperation modes: HI Mode (OEM lead + Huawei full hardware/software supply); Smart Selection Mode (Huawei participates in design and sales channels); Components Mode (individual components/chips). "Yinwang Intelligence" (Vehicle BU independent entity) seeking external strategic investment; Chery Automobile invested ~15% stake.

Huawei's "bundled supply" strategy — from chip to sensor to algorithm to map, everything inseparable — creates maximum switching cost but also triggers OEM fears of losing their "soul."

6.4 Desay SV (002920)

2025 Key Data:

• Revenue: RMB 32.557B (+17.9%)
• Smart driving revenue: RMB 9.7B (+116%)
• Smart cockpit revenue: ~RMB 20.6B (63.2% of total)
• April 2026 Hong Kong IPO; estimated valuation ~HKD 63B
• China smart driving domain controller market share: 21.2%; Global: 8.8% — #1 in both

Competitive advantage: covers global top-15 OEM customers; not exclusively tied to any single chip platform (simultaneously collaborates with NVIDIA, Horizon, Huawei); global expansion into European OEM supply chains.

6.5 JOYSON Electronics (688326)

2025 Key Data: Revenue RMB 6.848B (+23.59%); net profit ~RMB 99.5M — turning profitable from RMB -550M loss in 2024; ZCU cumulative deliveries >2 million.

Joint ZCU 3.0 platform with Geely (using Geely SE1000 + Horizon Journey); second-generation integrated-parking-and-driving domain controller launched. Turnaround through product focus, cost reduction, and customer depth.

6.6 Haomo.AI

DriveGPT model iterating through multiple versions; HPilot 5.0 urban NOA rolled out across GWM WEY and Haval brands; data flywheel accumulated ~6B equivalent km; planning A-share or HK listing at ~RMB 10B valuation.

6.7 Momenta

D-round investment at ~USD 6B valuation; investors include Toyota, Mercedes-Benz, SAIC, Bangkok Bank; production partners: Zhiji L6/L7/LS7, Denza N7/Z9GT, Yangwang U8; single-stage E2E + data flywheel architecture; collaborating with Bosch for European market penetration.

6.8 Baidu Apollo / Luobo Kuaipao

Q3 2025: ~3.1M rides (+212% YoY); cumulative >11M rides by end-2025; full-driverless operation in Wuhan, Chongqing, Beijing, Shanghai; ANP (Apollo Navigation Pilot) commercial OEM design wins.

6.9 Pony.ai (PONY.AI)

Listed on NASDAQ in late 2024; fleet of 1,200 vehicles across 10 cities by June 2025; cumulative mileage 48.6M km; UE turned positive in Guangzhou; PonyTruck with China Merchants and BYD partnerships; driverless in Shenzhen.

6.10 WeRide (WRD)

Listed on NASDAQ October 2024; core operations in Guangzhou, Wuhan, Chongqing; >12,000 daily orders; H1 2025 gross margin ~23%; partnership with Ruqi Mobility (GAC-owned ride-hailing); expanded to Abu Dhabi, UAE — most internationally active Chinese L4 company.

6.11 ECARX / Zekr ecosystem chip

ECARX's "Dragonhawk One" (SE1000) in 2023 production (7nm, Geely ecosystem); primarily for smart cockpit; smart driving chip direction in development.

6.12 Black Sesame (L4 Compute)

Cambricon Travel (MLU370, 400+ TOPS) targeting L4 Robotaxi compute; deep partnership with Baidu Apollo; long-term logic: accumulate L4 technology to target passenger vehicle L3 market when regulations allow.

6.13 NavInfo (002405)

HD map coverage >160,000 km of expressways; transitioning to "light map" and crowdsourced updates; joint solutions with NVIDIA and Desay SV; "Map + Positioning + Perception" integrated product targeting "location intelligence data service provider" repositioning.

6.14 Calterah Semiconductor

China's #1 77 GHz mmWave radar chip supplier; Alps-XD series for 4D imaging in BYD, Li Auto, Geely, Chery production vehicles. Competitive vs. TI AWR and Infineon: better price and supply chain responsiveness.

6.15 Hesai Technology (HSAI)

~150–180M front-installed LiDAR units in 2025; AT128 in Li Auto L7/L8/L9, Lotus Eletre, GWM Tank; ET30 solid-state LiDAR for sub-RMB 300 production target.

6.16 Neusoft Reach

Eastsoft Group subsidiary; AUTOSAR software platform + domain controller software + E/E architecture solutions; primary customers: DFAC, FAW Group, Dongfeng-PSA; key "software middleware" enabler for traditional OEM SDV transitions.

VII. Mid-Stream Industrial Clusters: Seven Core Hubs

7.1 Beijing: Algorithm and L4 Brain Center

Highest concentration of smart driving algorithm companies in China: Horizon Robotics, Haomo.AI, Baidu Apollo (Luobo Kuaipao operational HQ), DiDi Autonomous Driving, Momenta Beijing R&D, Pony.ai Beijing team, WeRide Beijing. Beijing's Yizhuang (Economic and Technological Development Zone) leads in "driverless" commercial operation. Top universities (Tsinghua, Peking, BUAA, BIT) supply AI and autonomous driving engineers.

7.2 Shanghai Jiading: OEM and System Integration Heart

SAIC Group HQ; Zhiji and R Autos in Jiading; Momenta HQ; HiRain Technologies key R&D center; Desay SV's Jiading R&D and JDP teams. Tesla's Gigafactory Shanghai: ~1 million vehicles/year, local data center for FSD data. Extensive PCB/PCBA, wire harness, precision structural component manufacturers in Jiading/Qingpu/Songjiang.

7.3 Suzhou: Sensor and Precision Manufacturing Base

Hesai and RoboSense mass production factories; Desay SV Suzhou manufacturing center; large clusters of PCBA manufacturers, precision structural component workshops (aluminum die-casting, LiDAR housings) in Kunshan; benefits from optical precision manufacturing heritage and Shanghai-Suzhou commuting corridor.

7.4 Wuhan: Dongfeng Ecosystem and Future Technology City

Dongfeng HQ (Dongfeng AEOLUS, Lantu, Menshi); Luobo Kuaipao's largest operation city (~400 driverless vehicles); Huawei Research Institute; "Future Technology City" (Guanggu) technology park for smart driving startups.

7.5 Shenzhen: Hardware Fast-Iteration Arsenal

BYD HQ (largest single Horizon customer; self-developed 4D mmWave ~RMB 100 via BYD Electronics); Calterah collab with Shenzhen manufacturing; camera module ODMs (OFILM, Sunny Optical vehicle division); PCB/FPC manufacturers (Shengyi Electronics, Zhen Ding) for domain controllers and sensors.

7.6 Guangzhou: Robotaxi First City and OEM Cluster

GAC (invested in Momenta; AION brand); WeRide and Pony.ai core Robotaxi cities; Guangzhou Nansha "smart city" demonstration zone for L4 testing and operation.

7.7 Chongqing: Changan Ecosystem and Western Smart Driving Center

First L3 pilot city (plate "渝AD0001Z"); Changan + Huawei (Deepal + Avatr) + Horizon deep cooperation; concentration of traditional automotive electronics manufacturers transitioning to intelligent driving.

7.8 Suzhou Kunshan: Hidden Champion PCB/Precision Cluster

Auto-grade PCB/PCBA manufacturers (IATF 16949 certified), aluminum die-casting for domain controller housings, high-speed Ethernet cable assemblies for vehicles; JIT supply proximity to Shanghai OEMs.

7.9 Chengdu-Xi'an: Emerging Inland R&D Centers

Chengdu: testing on Sichuan's expressway network; Yutong/Foton Chengdu plants; cost-effective R&D talent. Xi'an: Shaanxi Automotive Group commercial vehicle smart driving; BYD's largest production base; NWPU Unmanned Systems Institute-incubated startups.

7.10 Industrial Cluster Density Analysis — factory data platforms Data Insights

Based on data from factory data platforms (www.tianxiagongchang.com) covering 4.8 million verified Chinese factories, four key insights emerge:

Insight 1: Perception hardware manufacturing heavily overlaps with consumer electronics supply chains — camera modules, CIS, VCSEL — drawing from factories originally serving the smartphone industry with vehicle-grade certification upgrades.

Insight 2: Domain controller manufacturing is highly concentrated in the Shanghai–Suzhou corridor, driven by JIT delivery requirements to co-located OEMs.

Insight 3: Algorithm software companies cluster into a stable "dual-hub" pattern: Beijing (Haomo, Baidu Apollo, Pony.ai, Huawei Vehicle BU R&D) and Shanghai (Momenta, HiRain Technologies, Neusoft Reach Shanghai).

Insight 4: Robotaxi operating cities show a "deep cultivation effect" — Baidu's concentrated Wuhan operations have grown a local ecosystem of vehicle maintenance, sensor calibration, and fleet dispatch service companies that appear nowhere in listed-company databases but are indispensable to commercial L4 operation.

VIII. Sub-Track Deep Analysis

8.1 Highway NOA: Three-Year Leap from Niche to Standard

Three phases: Phase 1 (2020–2022): pioneers — Tesla NOA, NIO NOP, XPENG NGP. Phase 2 (2022–2024): cost descent — Haomo, Huawei driving highway NOA into RMB 150K+ price bands. Phase 3 (2025–): mass popularization — Journey 5/6M cost breakthroughs pushing highway NOA below RMB 200K vehicles; ~250–300K vehicles shipped with highway NOA in 2025.

8.2 Urban NOA: The Crown Jewel of Smart Driving

Urban NOA is technically ~3× harder than highway NOA. In 2025: Huawei ADS 3.0/4.0 leads in mapless coverage breadth; XPENG XNGP 200+ cities, 100-km no-takeover urban segments growing; Li Auto AD Max 3.0 end-to-end, covers all Chinese cities with road infrastructure; NIO NAD strong in complex scenarios, constrained by subscription conversion rate; BYD DiPilot advancing urban NOA via Journey 6M, expected scaled OTA rollout in 2026.

Takeover rate (interventions per X km) is the core performance metric. Premium solutions in 2025: one intervention per 15–30 km urban; budget solutions: one per 3–5 km.

8.3 Automated Valet Parking (AVP)

Huawei ADS+AITO: HAVP (Huawei Automated Valet Parking) for compatible car parks; XPENG "memory parking" for personal spots. AVP deployed by 10+ brands in premium trim levels; full nationwide scaling requires 3–5 years of car park infrastructure upgrades.

8.4 Robotaxi: Luobo Kuaipao Leads, Three-Player Structure

Baidu Apollo (Luobo Kuaipao): Largest scale — >11M cumulative rides, Wuhan ~400 driverless vehicles, Q3 2025 ~3.1M orders (+212%). UE positive in Wuhan/Chongqing; pricing ~60–80% of regular ride-hailing.

Pony.ai: Premium route — 1,200 vehicles, 10 cities, 48.6M km; UE positive in Guangzhou; simultaneous Robotruck development.

WeRide: Multi-format — Robotaxi + Robobus + cargo; H1 2025 gross margin ~23%; Abu Dhabi international expansion.

DiDi Autonomous Driving: Integrating L4 tech with DiDi platform; still in testing, more conservative timeline vs. peers.

Two scale-up barriers: (1) hardware cost — full L4 sensor suite RMB 50,000–100,000/vehicle today; target RMB 20,000–30,000 by 2027; (2) insurance/legal framework — commercial full-driverless liability assignment unresolved.

8.5 Robotruck: Clearest Commercial Logic

Market: China road freight RMB 3T/year; trunk highway freight ~40% (RMB 1.2T); driver wages 35–45% of operating cost; one Beijing-Shanghai route saves ~RMB 150K/truck/year.

Inceptio: DFAC-backed; >1,000 commercial trucks operational; L2+-to-L4 evolution path.

PonyTruck: Partnerships with China Merchants Logistics, BYD; targeting full-driverless on select routes by 2027.

TuSimple / TuSimple China: Deep technical expertise; China operations adjusting due to geopolitical complexity.

8.6 End-to-End Large Models: Core Narrative of Autonomous Driving 2.0

E2E replaces modular systems: one neural network from raw sensor input to steering wheel angle. Tesla FSD v12 (2024 early, first mass-produced E2E system). China: Huawei ADS 4.0, XPENG XNGP 2.0, Li Auto AD Max 3.0, Haomo DriveGPT, Momenta single-stage E2E — all in 2024–2025 transition.

VLA (Vision-Language-Action): Adds language reasoning to E2E — handles semantic complexity like "follow the traffic officer's instructions." Engineering prototypes in 2025; production testing expected 2026–2027.

8.7 Robobus: Urban Public Transit Automation

WeRide's Robobus in Guangzhou Nansha and Singapore; Shenlan Science Alpha-Bus early trial in Suzhou Industrial Park; Yutong/Zhongtong OEM partnerships for L4 shuttle buses. Robobus expected to achieve large-scale commercial operation 1–2 years ahead of open-road urban Robotaxi.

8.8 Robotruck Trunk Logistics: Full Commercial Logic

Market: China road freight RMB 3+ trillion/year; trunk highway freight ~40%; driver wages 35–45% of operating cost. Inceptio (DFAC-backed): >1,000 trucks in commercial operation; PonyTruck: Q4 2027 target for full-driverless commercial trunk lines; TuSimple China: deep technical expertise, strategic adjustment ongoing.

8.9 Port, Mine, Campus: L4 "Enclosed Environment" Beachheads

Ports: Shanghai Yangshan Phase IV with ~300+ automated container trucks; Tianjin Port, Xiamen Port similar deployments; ZPMC (600320) provides equipment.

Open-pit mines: Inner Mongolia, Shanxi, Anhui coal/iron ore mines with L4 mining truck fleets; Huitu Technology, Yikong Zhijia, Tage Zhijia as main technology suppliers; transitioned from demo to routine commercial operation.

Campus shuttles: Low-speed (<30 km/h), enclosed; Baidu Apollo Moon, WeRide One in multiple campus deployments.

8.10 "Perception Safety Baseline" and Active Safety

L3 mandatory requirements: (1) Dual-redundant braking (BBW Redundancy); (2) Dual-redundant steering (SBW Redundancy); (3) Dual-power supply; (4) Driver monitoring system (DMS); (4) Fault Detection and Diagnosis (FD&D) triggering Minimum Risk Maneuver. These add RMB 5,000–10,000 to L3 vehicle BOM, a meaningful cost barrier to L3 mass production.

IX. Technology Roadmap

9.1 E2E and Neural-Network Driving Paradigm Established

"Data > Rules" is now the industry consensus. Tesla has ~5 million FSD-activated vehicles generating the world's largest autonomous driving dataset. China's defense: rapid mass production to build domestic data flywheels, plus data-localization regulations blocking Tesla data repatriation.

9.2 BEV + Transformer: Perception Dimension Uplift

BEV perception converts multi-camera images into a unified top-down feature space via Transformer network, greatly improving multi-camera fusion consistency. Popularized by Tesla (Occupancy Networks); adopted by Huawei, Li Auto, XPENG, Momenta. BEV adoption improved "pure vision" route viability while deepening the "pure vision vs. multi-sensor" debate.

9.3 4D mmWave Radar: Challenging LiDAR at Lower Cost

BYD self-developed 4D mmWave ~RMB 100 in Yangwang U8; Huawei 4D imaging radar in Qiankun suite; Maogan Technology (WeRide spin-off) MVRA188 chip completed tape-out; projected 2025 passenger car volume >10M units (vs. LiDAR ~3M). Resolution reaching 0.5° angular resolution in some designs. Consensus: "complementary, not replacement" — premium vehicles keep LiDAR; mid-tier uses "4D mmWave + high-compute vision."

9.4 LiDAR Domestication: Hesai and RoboSense

Hesai (HSAI): ~150–180M front-installed units in 2025; global front-installed LiDAR shipment leader; AT128 in Li Auto, Lotus, GWM Tank; ET30 solid-state targeting sub-RMB 300. RoboSense (2498.HK): MEMS LiDAR M2 in BYD, GWM, GAC; rapid capacity expansion. Cost reduction root: domestic VCSEL/APD manufacturing, semiconductor process scale-up.

9.5 V2X: China's Distinctive "Vehicle-Road-Cloud" Route

China's "Vehicle-Road-Cloud Integration" (V2X + roadside perception + cloud algorithms) is the most distinctive technology route difference from the rest of the world. C-V2X standard (5G-V2X or PC5 direct communication). In 2025: 20+ cities piloting large-scale Vehicle-Road-Cloud deployment; Beijing Yizhuang, Shanghai Jiading, Guangzhou Nansha, Wuhan Economic Zone, Chongqing Liangjiang New Area as five pioneer zones. V2X enables low-compute vehicles to achieve L3/L4 equivalent perception via roadside sensing + cloud compute — reducing in-vehicle BOM at the cost of infrastructure investment.

9.6 Software-Defined Vehicle and OTA Iteration

SDV consolidates functions from 70–100 ECUs into 1–3 high-compute domain controllers; replaces CAN bus with Ethernet backbone. OTA enables quarterly or monthly feature updates. Key supply chain implications: AUTOSAR (CP/AP) framework expertise rises in strategic importance; functional safety (ISO 26262) becomes a genuine barrier; "software as profit center" (smart driving subscription, OTA upgrade package) is an emerging business model.

9.7 Sensor Fusion Architecture: From "Stacking" to "Synergy"

Evolution from Late Fusion (each sensor processed separately, results merged) to Early/Mid Fusion (raw data or features fused at neural network early layers). Tesla's Occupancy Network (voxel-based 3D space occupancy prediction from camera images alone) pioneered a new paradigm subsequently followed by Huawei, XPENG, and others. Compute requirements for Early Fusion: 3–5× higher than Late Fusion, explaining genuine 150–500 TOPS requirements for urban NOA.

9.8 High-Precision Positioning: RTK + IMU + Visual Fusion

Production vehicle standard: RTK-GNSS + IMU + visual positioning (HD map feature matching) Kalman filter fusion; visual positioning takes over from GNSS in urban "canyon effect" (signal blocked by high-rise buildings). L4 Robotaxi additionally uses LiDAR point cloud map matching (NDT/ICP, 3–10 cm accuracy).

9.9 Hardware Redundancy and Functional Safety: L3's "Extra Cost"

L3 vs. L2 fundamental physical difference: L3 must execute Minimum Risk Maneuver (auto-pull to roadside) on system failure, not just "release hands." Requirements: dual-redundant braking/steering/power supply; DMS; FD&D. Additional BOM ~RMB 5,000–10,000 per L3 vehicle. Foreign Tier 1 (Bosch, ZF, Continental) still have an advantage in braking/steering redundancy, where domestic substitution maturity trails perception and compute layers.

9.10 LLMs and Driving: A New Fusion Direction

2024–2025: combining LLMs with autonomous driving perception-planning. Current practical form: "LLM-Agent + dedicated perception model" dual-layer architecture — fast perception model (<50ms), LLM-Agent for complex semantic judgment (500ms–1s). Chinese implementations: Huawei MDC+LLM joint inference, Li Auto AD Max VLM-assisted understanding, NIO NOMI AI with driving intent understanding. Expected milestone: "vehicles that understand context and adapt decisions" from demo to production rollout in 2026–2027.

X. Key Risks and Challenges

10.1 Regulatory Pace Uncertainty

China's L3 commercialization speed is ultimately determined by regulatory pace, not technology maturity. The "mature one, approve one" principle means expansion could be far slower than industry expects if legal liability framework, insurance product design, and accident forensic standards are not resolved. Risk: L3 hardware upgrade cycle (high-end domain controllers, dual-redundant sensors, DMS cameras) delayed demand if regulatory expansion stalls.

10.2 Smart Driving Safety Incidents and Public Opinion

High-profile accidents could trigger: consumers actively disabling NOA; regulatory investigations pausing software updates; insurers raising smart-driving-related premiums. 2025 is a higher-risk year — more vehicles activating more complex urban NOA creates more edge-case exposure. Balancing "expanding coverage" with "managing edge-case failure rates" is the core operational challenge for every algorithm company.

10.3 US Export Controls: The Sword Over Compute Chips

Current restrictions: H100/H200/A100 direct export banned; A800/H800 subsequently restricted; AMD MI300 similarly constrained. Vehicle inference chips (Orin/Drive Thor) not yet directly restricted, but supply chain risk consciousness sharply elevated. Impact: training clusters migrate to Huawei Ascend 910B/910C; "dual-track supply" (imported + domestic backup) now standard OEM procurement strategy; export controls de-facto accelerating domestic chip substitution — a structural tailwind for Horizon and Black Sesame.

10.4 Tesla FSD Entry Market Impact

On user market: Direct experience comparison with Huawei ADS, XPENG XNGP, Li Auto AD Max in premium segment. If FSD performs well in China-specific urban conditions (dense cycling, unregulated intersections), pressure on domestic algorithm companies intensifies; if not, "domestic data advantage" narrative is reinforced.

On supply chain: Tesla's pure-vision approach doesn't directly harm Chinese LiDAR makers, but FSD's scale may accelerate "vision-first" design discussions and challenge LiDAR "necessity" arguments in mid-tier vehicles.

On data compliance: China requires local data storage, blocking FSD data repatriation to US training clusters — creating a genuine data moat for domestic players in China-specific scenarios.

10.5 Algorithmic Safety and the "Automation Paradox"

E2E black-box decisions complicate accident investigation: neural network decisions are not traceable to explicit rules. The "Automation Paradox": extended NOA use degrades manual driving readiness, increasing accident severity when the driver is actually needed for take-over. Europe's ALKS regulation and China's DMS requirements address this; engineering it correctly requires extensive validation in edge cases.

10.6 Compute Cost Curves and Industry Concentration

"Big model + big data + big compute" requirements are creating a "capital-intensive barrier" unfavorable to smaller players. Winner-takes-all concentration is increasing in algorithm software: small algorithm startups likely to be acquired or exit in 2026–2028, consolidating toward frontrunners. This is actually an opportunity for mid-stream Tier 1 (Desay SV, JOYSON Electronics): fewer algorithm companies means OEMs seek "one-stop" system integrators, strengthening strategic Tier 1 positioning.

10.7 Talent Competition: Global Race for Top AI Engineers

Compound-skilled AI × autonomous driving engineers are globally scarce. Top-tier Chinese CV PhD total compensation (RMB 2–3M) now approaches Silicon Valley parity, driving a return wave of overseas talent. In 2025, with some Robotaxi companies tightening hiring due to valuation adjustments, "talent concentration" is mirroring capital concentration toward frontrunners — self-reinforcing positive feedback of data + compute + talent at scale leaders.

10.8 Insurance and Legal Framework: The Shortest Stave

Traditional vehicle insurance assumes "human drives the car." L3 flips this: liability allocation between system provider and driver in ODD-active periods is unresolved. No specialized L3 commercial insurance products yet in China; OEM self-insurance/warranty promises are not formal insurance products. Timeline: with L3 pilot vehicle accumulated mileage expected to reach statistical significance by 2027, actuarial models may mature enough for formal L3 commercial insurance products to launch in 2027–2028.

10.9 "Ghost Braking" and False-Trigger User Experience Risk

2021–2022 Tesla Autopilot "ghost braking" incidents led to NHTSA investigation and ~2.2 million vehicle OTA recall. False-positive rate (spurious emergency braking without real obstacles) is as important as false-negative rate (missed obstacle detection). E2E models theoretically improve the balance via large-scale data training; the challenge is achieving this in all extreme scenarios (strong backlighting, nighttime, fog/rain/snow, mirror-reflective road surfaces). False-trigger rate is a key "production readiness" acceptance criterion for Tier 1 suppliers delivering to OEMs.

10.10 "Over-Reliance" and Take-Over Capability Degradation

Extended NOA use degrades manual driving readiness (reduced situational awareness). MIT AgeLab and Stanford REVS studies confirm increased take-over latency and quality degradation even with adequate warning time. Engineering response: progressive safe-stop maneuver as mandatory L3 system behavior; DMS in China L3 pilots monitors driver state continuously, preventing simple "hand-off without readiness confirmation."

XI. 2026–2030 Outlook and Investment Logic

11.1 NOA Penetration Charging Toward 50%

Core judgment: highway NOA will exceed 60% by ~2028, approaching "standard fit"; urban NOA will approach 40–50% by 2030 as standard on 30K+ RMB vehicles.

Three driving forces: (1) continued cost curve decline — Journey 6M/6P, AT128 solid-state, 4D mmWave national cost reductions make urban NOA viable below RMB 150K by 2026–2028; (2) "smart driving arms race" OEM competition effect forcing mid-tier OEMs to accelerate; (3) L3 pilot expansion raising consumer "the system can manage driving" awareness, boosting NOA active activation rates.

11.2 L3 Commercialization: 2026–2028 Critical Window

Three observable metrics:

• Models in L3 pilot: 2025 YE = 2; 2026 YE target = 10–15; 2027 YE target = 30+
• L3-specific insurance product launch: by Chinese insurers (PICC, Ping An, CPIC)
• First complete judicial precedent: first accident with L3 system active adjudicated with clear liability attribution

If all three align by 2027, L3 mass commercialization may enter rapid scaling phase in 2028, generating incremental demand for high-end domain controllers, dual-redundant sensors, and DMS cameras.

11.3 L4 Robotaxi Scaling: Three-Step Path

Step 1 (2025–2027): Core city volume expansion — national monthly rides from 5–6M toward 20M; UE positive in Wuhan, Guangzhou, Chongqing for Luobo Kuaipao and peers.

Step 2 (2027–2029): Full-driverless + cost inflection — sensor kit cost to RMB 20,000–30,000 (from RMB 50,000–100,000); full-driverless commercial licenses in multiple cities; complete vehicle cost down to RMB 250,000–300,000.

Step 3 (2029–2035): National network — 100+ city coverage, deep integration with public transit (last-mile connector), national daily orders exceeding 30M.

11.4 Four Investment Logics in Smart Driving Supply Chain

Logic 1: Domestic chip substitution "elasticity dividend" — Horizon, Black Sesame under "US export controls → forced substitution + own technology rapidly catching up" double accelerator; Journey 6P TAM expanding upward as algorithm companies accept domestic compute.

Logic 2: Tier 1 system integrators gaining bargaining power — as algorithm space consolidates, OEMs seek "one-stop" solution providers; Desay SV's 21% China domain controller share to hold and grow; domain controller market expected >RMB 150B by 2030.

Logic 3: Robotaxi operators' platform value re-rating — when Luobo Kuaipao, Pony.ai achieve full-fleet profitability (~2027–2028), valuation logic shifts from "burning cash on R&D" to "high-growth platform operations," analogous to early DiDi's valuation expansion. Robotaxi has no driver supply constraint — scale limited only by vehicle production speed and regulatory licensing.

Logic 4: HD map and data infrastructure "public utility" repositioning — HD map business model evolving from "high-fee OEM licensing" to "infrastructure-as-a-service"; crowdsourced map technology will define next-generation market leaders; NavInfo's "light map" transition and data service model is the key strategic pivot to watch.

11.5 Chinese Smart Driving Supply Chain's Global Opportunity

2026–2030 growth through export: LiDAR (Hesai, RoboSense) certified in European/US/SEA front-installed markets; domain controllers (Desay SV) winning European OEM design-ins; algorithm software (Momenta through Bosch, WeRide in UAE) as two pioneer international templates; complete vehicles (BYD, Zeekr) exporting with domestic smart driving systems. Export path faces Western protectionism, data sovereignty management, and local certification barriers — progress will be slow and steady.

11.6 Supply Chain Integration Trends: "Platform" or "Fragmentation"?

Integration Direction 1: "Super Tier 1" consolidation — Desay SV expanding from domain controllers to LiDAR integration and cockpit-drive fusion platforms; OEM "one-stop sourcing" demand.

Integration Direction 2: SDV software platform layer standardization — AUTOSAR/automotive Android/QNX battle for a unified software platform; Huawei HarmonyOS smart vehicle OS vs. Aptiv central compute platform vs. various ASPICE implementations.

Integration Direction 3: "Algorithm-hardware vertical decoupling" — leading OEMs (XPENG, Li Auto, NIO, Huawei) push "self-developed algorithm + market-procured hardware," reducing single-platform hardware lock-in risk.

11.7 International Competition Landscape Evolution Forecast

Tesla FSD global rollout: FSD performance in China-specific urban conditions (dense cycling, unregulated intersections) vs. domestic solution in 2026–2027 will be the industry's defining product comparison.

Waymo's limited China participation: Three barriers (regulation, data, partners); unlikely to operate commercially in China at scale before 2030.

Korean/Japanese OEM responses: Hyundai/Kia (IONIQ series), Toyota (Momenta partnership), Honda (Honda SENSING Elite L3) — foreign OEMs in China shifting from "use proven Western solutions" to "partner with Chinese local technology."

Chinese supply chain going global: The line of "China smart driving outbound" will become progressively thicker in 2026–2030, as China shifts from "defending against foreign competition" to "proactively competing in global markets."

11.8 China Smart Driving Industry's Long-Term Moats

First moat: Data scale and local road condition coverage — China's annual ~14M+ smart driving vehicle shipments generate the world's most voluminous driving dataset; China's urban road conditions (dense cycling, non-motorized lanes, construction zones, unmarked intersections) create a "data moat" that foreign datasets cannot replace.

Second moat: Complete domestic supply chain ecosystem — from chips (Horizon, Black Sesame) to sensors (Hesai, Calterah) to domain controllers (Desay SV, JOYSON) to algorithms (Huawei, Haomo, Momenta) to test certification (CATARC, Waikay), China has built a functionally complete domestic smart driving supply chain ecosystem.

Third moat: Policy support and regulatory pioneer zones — China's government has the most proactive smart driving industrial policy among major auto markets, with L3/L4 regulatory pilots, V2X infrastructure investment, and Robotaxi commercial licensing providing structural "home court advantage" for domestic players.

XII. Conclusions and Brand Anchor

The main thread of this report can be pulled in one line.

Intelligent driving is a track that braids five cables together: chips, sensors, algorithms, regulation, and data. Its complexity lies not in any one cable being exceptionally difficult, but in all five needing to advance simultaneously — if any one falls behind, the whole braid goes slack. This is why, sixteen years since Google's first autonomous vehicle test, L5 "full autonomy in any scenario" remains in research — technological frontiers have always been harder than prophets imagined.

But if you pull your gaze back from L5's horizon to 2025–2026, the scene is entirely different.

Right now, seven of every ten new passenger vehicles rolling off China's production lines ship with driver assistance; roughly two in ten carry the "full-journey assisted driving from A to B" capability called NOA. A company called Horizon Robotics shipped ten million smart driving chips into Chinese cars in eight months of 2025 — equivalent to approximately 40,000 cars every day acquiring smarter eyes and a smarter brain because of one domestic chip. A company called Desay SV used a 21% global market share to lock in the "domestic domain controller number one" position, while simultaneously sprinting for a Hong Kong IPO to fund the next round of software transformation. Baidu's Luobo Kuaipao has completed over eleven million rides in Wuhan — in a car with no driver, delivering passengers to their destinations, producing the world's largest-scale L4 commercial result.

These are not imagined futures. They are the present.

Looking forward from this present, two "order-of-magnitude jumps" deserve marking.

One is L3 pilot expansion. From December 2025's two models in two cities, to potentially dozens of models and twenty cities by 2028 — each approval means a collective supply-chain upgrade: high-compute domain controllers, dual-redundant sensors, DMS takeover-monitoring cameras are not L2 standard equipment; they are the "entry ticket" to L3. Who locks in the production supply chain in this window owns first-mover advantage in the L3 era.

The other is Robotaxi's cost inflection. LiDAR fell from RMB 5,000 to just over RMB 1,000, still declining; 4D mmWave radar's domestic chips have taped out, cost converging toward RMB 100; domestic compute chip substitution accelerating — each declining cost curve advances the "L4 commercialization profitability" inflection date. The 2027–2028 window is where Luobo Kuaipao and Pony.ai may achieve genuine full-fleet profitability in select cities, and where the entire Robotaxi industry narrative may shift from "burning R&D cash" to "platform growth."

Behind both the L3 regulatory breakthrough and the Robotaxi cost inflection lies not a single star company's lone advance, but the coordinated evolution of an enormous factory network. Behind one Horizon Journey 6P: a foundry, packaging-and-testing house, PCB manufacturer, module integrator. Behind one Hesai AT128 LiDAR: VCSEL chips, APD detectors, optical lens assemblies, structural die-castings. Behind one Luobo Kuaipao operating vehicle: sensor suppliers, domain controller factories, wiring harness plants, system calibration service providers — the full chain.

This factory network is mapped in full on the factory data platforms platform — from LiDAR sensing modules to domain controller structural components, from algorithm development services to production verification testing. In the 4.8 million verified Chinese factories tracked by factory data platforms, you can identify every category of factory and every supply-chain node behind the production of this industry.

Chips being domesticated, domain controllers being upgraded, algorithms going end-to-end, regulations being piloted, L4 being commercialized — these five cables, in the 2025–2030 stretch, will continue to be braided tighter at the scale and speed unique to China. This is not a question of "is there hope," but of "what is the pace." Our judgment: the pace is faster than most people expect, but more measured than a handful of optimists proclaim — fast, but not skipping steps.

From one chip, to a car that thinks. The road is still long. But in China, 4.8 million factories have begun supplying it with blood.

In the past three years, factory data platforms Industry Research Institute has continuously tracked the evolution of China's core manufacturing supply chains, remeasuring "how deep China's factory network really is" with each new research topic. The intelligent driving topic is the broadest supply chain yet encountered — from atomic-scale semiconductor fabrication to every NOA-equipped vehicle on urban roads, traversing over a dozen industrial categories, thousands of enterprise nodes, spanning the geography from Beijing algorithm labs to Suzhou precision manufacturing workshops.

A network woven from 4.8 million factories does not disappear in a quarter, and cannot be replicated overnight by any single competitor. This is the most authentic moat of China's intelligent driving supply chain, and the fundamental basis for our long-term conviction in this sector.

Data Sources and Key References

Tianxia Gongchang Industry Research Institute — based on Tianxia Gongchang platform factory and supply chain data, combined with public documents, official information, and authoritative media reports, synthesized and analyzed for this research report. Tianxia Gongchang Industry Research Institute focuses on deep research of manufacturing supply chains; its core data source is the database of 4.8 million verified Chinese factories on the Tianxia Gongchang platform (www.tianxiagongchang.com), covering factory geographic distribution, industry classification, supply chain positioning, and supplier relationships. Key references include:

• factory data platforms platform Chinese factory database and industrial cluster data (www.tianxiagongchang.com)
• Horizon Robotics (HK 9660) 2025 Annual Report (HKEX disclosure)
• Black Sesame Intelligent (HK 2533) 2025 Interim Financial Report
• Desay SV (002920) Hong Kong IPO Prospectus (April 2026)
• JOYSON Electronics (688326) 2025 Annual Report
• Mobileye Global Inc. SEC Form 10-Q (FY2025 Q3)
• MIIT et al.: "Automotive Industry Steady Growth Action Plan (2025–2026)"
• MIIT: First L3 Conditional Automation Vehicle Type Approval Announcement (December 2025)
• SAE International: J3016 Driving Automation Level Standard
• PRC National Standard GB/T 40429—2021 "Taxonomy of Driving Automation for Vehicles"
• China Data Security Law (2021), Personal Information Protection Law (2021)
• China "Regulations on Auto Data Security Management (Provisional)" (2021)
• Baidu Apollo (Luobo Kuaipao) 2025 operational data announcements
• Pony.ai (PONY) NASDAQ listing information and 2025 operational data
• WeRide (WRD) NASDAQ listing information and 2025 financial summary
• TrendForce: NVIDIA Drive Thor automotive chip market outlook (December 2024)
• Xinhua News Agency, People's Daily, Jiemian, First Financial, and other authoritative media reports
• 36Kr, Jiqizhixin (Synced), Autohome, and other industry media in-depth reporting