China autonomous driving: China leads autonomous-driving adoption and cost, but the path from L2+ to safe, scalable L4 remains unproven

Overview

This industry report examines whether China’s scale, cost advantages, adoption of advanced driver assistance, and AI progress can make it a global leader in autonomous driving. Bernstein sees substantial long-term potential, but contrasts the promise of the L2+-first pathway with the presently stronger demonstrated safety record of dedicated L4 operators.

Core views

Bernstein argues that China is already ahead in the adoption of advanced L2+ driver-assistance systems and is likely to be among the first markets to deploy L3 at scale. China’s L2+ penetration reached 22% of 2025 new-vehicle sales, compared with 6% in the U.S. and 1% in both Europe and Japan. The report attributes this lead to the world’s largest EV fleet, receptive consumers, rapid OEM launches, intense domestic competition, and a maturing local supply chain. Regulatory progress supports the thesis: MIIT granted initial L3 road-access permits to Changan Deepal SL03 and BAIC Arcfox Alpha S on December 15, 2025; issued mandatory automated-driving safety standard GB 44721-2026 on July 30, 2026; and set out a 2030 deployment roadmap in September 2026. With the safety standard effective July 1, 2027, Bernstein expects broader L3 approvals, although it remains cautious on the pace of consumer adoption and subscription monetization. Cost is a central Chinese advantage. Bernstein estimates Pony.ai’s Gen 7 robotaxi at roughly US$40k today and US$33–34k by end-2026, around US$50–60k below Waymo’s latest Ojai platform. Lower vehicle and hardware costs can support larger fleets, which generate more driving data and enable faster model iteration. This creates the report’s proposed competitive flywheel: lower costs support deployment scale; scale supplies data; data improves models; and better models reinforce adoption and cost competitiveness. The report also identifies overseas expansion as increasingly important because domestic competition could compress returns and leave long-term monetization uncertain. Pony.ai and WeRide are extending robotaxi operations and tests across the Middle East, Europe, and other markets, though volumes remain small. Momenta’s Germany-wide approval for L4 urban-road testing removes the need for city-by-city testing permits, but is not an operating license. XPeng has begun delivering ADAS-equipped vehicles in Europe and targets a broader global rollout of its driver-assistance capabilities by the fourth quarter of 2026. Regulatory approval and geopolitical constraints remain material limits on international scaling. Bernstein frames the strategic debate as bottom-up versus top-down autonomy. The bottom-up group—Tesla, XPeng, and Momenta—starts with L2/L2+ systems in mass-production cars, gathers real-world data, and seeks to progress toward L3 and L4. It benefits from millions of vehicles, lower-cost sensor configurations, shared perception/planning/control technology, and funding from vehicle sales. Tesla had nearly 200 supervised robotaxis in Austin, Dallas, and Houston by August 2026, with more than 380,000 autonomous miles without major accidents according to the company. XPeng had completed about 2,000 robotaxi passenger trips in Guangzhou and targets fully driverless operations by early 2027. Momenta reported 219 design wins across 26 OEMs, including 114 programs in production, and expects L3 launches in 2027 subject to approvals. The top-down group—Waymo, Pony.ai, and WeRide—was built for L4 driverless services from the outset, initially within defined operating domains. These operators emphasize richer sensor suites and redundancy to prioritize safety. Waymo completes more than 500,000 paid rides per week across 11 U.S. cities; Pony.ai had about 2,000 robotaxis and more than 1.5 million registered users in China by the second quarter of 2026; and WeRide operates a fleet exceeding 1,800 vehicles. The report views L4-dedicated systems as having the clearest demonstrated safety lead today and established driverless commercial operations, albeit with higher bill-of-materials costs. AI could narrow the divide because L2 and L4 share core technical building blocks, while end-to-end neural networks, world models, and vision-language-action frameworks may accelerate learning. However, Bernstein does not find sufficient evidence that L2+-first systems can currently achieve L4-comparable safety. Large data sets are not enough when rare, safety-critical edge cases must be identified, labelled, and learned. The report also highlights architectural differences: L2 systems rely on a supervising human and prioritize affordability and broad usability, while L4 systems must operate without human fallback and prioritize robustness within a defined operating domain. Whether vision-only systems can meet L4 redundancy requirements remains unsettled, as most L4 operators retain lidar, radar, and camera redundancy. Beyond technology, commercialization depends on regulation, liability, insurance, fleet operations, and consumer acceptance. Comparable public safety data are limited, and Bernstein finds NHTSA crash-report comparisons between Waymo and Tesla inconclusive because fleet scale, operating domains, autonomous miles, safety-driver presence, and accident attribution differ. Extreme-weather incidents and robotaxis’ inability to perform some community-safety functions further illustrate unresolved long-tail challenges. Nonetheless, Bernstein concludes that the industry has made meaningful advances in capabilities, operating reliability, and commercial scale, and may be entering the early stages of broader adoption.

Analysis framework

Bernstein compares adoption, regulation, deployment costs, fleet scale, international progress, hardware architectures, and business models. It then contrasts the data-led L2+-first pathway with the safety- and redundancy-led L4-dedicated pathway, using company operating metrics and available safety evidence to assess whether the two approaches can converge.

Methodology notes

• Industry AnalysisSupply-demand framework

  Adoption, regulation, deployment cost, and ecosystem-scale analysis

  The report assesses how consumer demand, regulatory permission, vehicle deployment, and supply-chain costs could determine the pace of autonomous-driving commercialization.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Autonomous-driving cost and data flywheel

  Lower hardware and vehicle costs are presented as enabling larger fleets, more data collection, faster model improvement, and potentially stronger commercial competitiveness.

• Competition & strategyValue chain analysis

  Bottom-up versus top-down autonomy strategies

  The report compares differing technology stacks, sensor choices, funding models, safety priorities, and routes to commercialization across L2+-first and L4-dedicated operators.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• BYD (1211.HK)

  Explicitly covered Chinese auto company rated Outperform.

• Geely (175.HK)

  Explicitly covered Chinese auto company rated Outperform.

• Xiaomi (1810.HK)

  Explicitly covered Chinese auto company rated Outperform.

• XPeng (XPEV; 9868.HK)

  Covered L2+-first player pursuing global ADAS rollout and robotaxi operations; rated Market-Perform.

  Strengths

  Vision-only strategy, in-house VLA and world models, and mass-market data collection.

  Weaknesses

  Has not yet scaled driverless robotaxi operations comparably with established L4 operators.

  Comparison

  Positioned with Tesla and Momenta in the bottom-up camp, versus Waymo, Pony.ai, and WeRide in the L4-focused camp.

  Risks

  Safety, regulatory approval, and whether vision-only hardware can support L4 autonomy.

• Li Auto (LI; 2015.HK)

  Explicitly covered Chinese auto company rated Market-Perform.

• NIO (NIO; 9866.HK)

  Explicitly covered Chinese auto company rated Market-Perform.

• Pony.ai (PONY.US; 2026.HK)

  Non-covered L4-focused robotaxi operator used to illustrate China’s cost advantage and overseas expansion.

  Strengths

  Low estimated Gen 7 cost, fleet scale, registered-user base, and international partnerships.

  Weaknesses

  Domestic monetization remains uncertain and overseas volumes remain small.

  Comparison

  L4-focused peer of Waymo and WeRide, with higher sensor redundancy than bottom-up players.

  Risks

  Regulatory approvals, geopolitical constraints, and intense domestic competition.

• WeRide (WRD.US; 800.HK)

  Non-covered L4-focused robotaxi operator used as an example of international expansion.

  Strengths

  Fleet exceeding 1,800 vehicles and permits across multiple international markets.

  Comparison

  Part of the L4-dedicated camp alongside Waymo and Pony.ai.

  Risks

  Overseas deployment remains subject to regulatory and geopolitical constraints.

• Momenta (6880.HK)

  Non-covered third-party ADAS provider pursuing both L3 and robotaxi expansion.

  Strengths

  219 design wins across 26 OEMs, 114 production programs, and Germany-wide L4 urban-road testing approval.

  Weaknesses

  L3 launches remain subject to regulatory approval.

  Comparison

  Bottom-up player alongside Tesla and XPeng, while also entering robotaxi applications.

  Risks

  Safety proof, regulatory approval, and commercialization execution.

Key data

• China L2+ penetration22%Share of 2025 new vehicle sales, versus 6% in the U.S. and 1% in Europe and Japan.
• Pony.ai Gen 7 robotaxi costc.US$40kEstimated current cost; projected at US$33–34k by end-2026.
• Pony.ai cost advantage versus Waymo Ojaic.US$50–60kEstimated lower cost for Pony’s Gen 7.
• China automated-driving safety standard effective date1 July 2027GB 44721-2026 is China’s first mandatory national safety standard for automated-driving systems.
• Waymo paid ridesMore than 500,000 per weekAcross 11 U.S. cities.
• Pony.ai fleetApproximately 2,000 robotaxisIn operation by 2Q26.
• Momenta design wins219 across 26 OEMsIncluding 114 programs already in production.

Impact & implications

Bernstein sees China’s adoption scale, local cost structure, and regulatory trajectory as supportive of broader L3 deployment and long-term autonomous-mobility growth. It cautions that competitive intensity, uncertain monetization, foreign-market constraints, and unresolved proof of L2+-to-L4 safety convergence may determine which business models ultimately scale.

Risks

• Domestic competition may compress returns and leave autonomous-driving monetization uncertain.
• International deployment is subject to regulatory approvals and geopolitical constraints.
• L2+-first systems have not yet demonstrated safety comparable to dedicated L4 systems, particularly in rare long-tail scenarios.
• Large-scale commercialization also depends on liability frameworks, insurance, fleet operations, and consumer acceptance.
• Extreme weather and community-safety situations remain difficult for autonomous systems.

What to watch

• Broader Chinese L3 commercialization approvals after the mandatory safety standard takes effect on July 1, 2027.
• Momenta’s expected 2027 L3 vehicle launches and the Audi E7X launch timetable.
• XPeng’s targeted wider global ADAS rollout by the fourth quarter of 2026 and its early-2027 driverless-operation target.
• The pace and regulatory success of Pony.ai, WeRide, and Momenta’s overseas deployments.
• Comparable, large-scale public safety data for L2+-first and L4-dedicated systems.