XPeng starts robotaxi mass production in Guangzhou with full-stack in-house technology

XPeng starts its robotaxi line in Guangzhou, three years behind the leaders and ahead of any other Chinese automaker

The first car off the line on Monday marks the start of a slow ramp. Public pilots arrive in the second half, fully driverless by early 2027. XPeng said on Monday that the first robotaxi had rolled off its production line in Guangzhou, making the Chinese carmaker, in its own framing, the first automaker in China to begin mass production of a robotaxi developed entirely with in-house technology. The vehicle is built on XPeng's new GX platform and is engineered, according to the company, to support Level 4 autonomous driving from the factory floor rather than through aftermarket retrofits. The headline is precise in a way that matters. XPeng is not the first company in China to put robotaxis on the road; Baidu's Apollo Go runs more than 1,000 vehicles across 22 cities and is the country's largest commercial operator, while Pony.ai's fleet reached 1,446 vehicles by late March and is on track for 3,000 by year-end. WeRide operates over 1,000 across more than 30 cities and 11 countries. What XPeng can claim is that it is the first traditional automaker, as opposed to a pure-play autonomous-driving company, to put a robotaxi-grade vehicle into series production in China. That distinction is part marketing, part substance. The pure-play operators have, until now, retrofitted consumer vehicles, often Toyotas or domestic EVs, with sensor stacks and onboard compute supplied by the AV company. XPeng's GX-based robotaxi is the inverse: a vehicle designed from the platform up with the compute, drive-by-wire chassis, and redundancy required for Level 4 operation. The car carries four of XPeng's own Turing AI chips delivering a combined 3,000 TOPS of compute, a Bosch next-generation steer-by-wire system that eliminates the mechanical steering shaft, and what XPeng describes as aviation-grade redundancy across the safety-critical systems . The software running on top is the company's VLA 2.0 stack, a vision-language-action model that compresses end-to-end response latency to under 80 milliseconds. Volume will be modest at first. Brian Gu, XPeng's president, told Reuters last month that the company will likely produce hundreds to thousands of robotaxis over the next 12 to 18 months. Pilot operations with safety drivers begin in the second half of this year. Fully driverless commercial service is targeted for early 2027. The robotaxi line in Guangzhou holds an intelligent-connected-vehicle road-testing licence, with L4 road tests already running on public roads. The commercial logic is partly defensive. XPeng's robotaxi push sits inside an aggressive 2026 diversification plan that also includes humanoid robots and a modular flying car, all powered by the same Turing silicon and the same vision-language-action software stack. The company has been clear that the strategy is to monetise the AI stack across multiple form factors rather than rely on EV margins alone in a market where prices are falling and Chinese government subsidies have been wound back. Volkswagen is the most visible external customer for that strategy. The German automaker, which took a 4.99% stake in XPeng for $700m in July 2023, named XPeng's Turing chip and the VLA 2.0 system as its first external commercial customer this year. It is the first time a major Western carmaker has adopted Chinese-developed autonomous-driving software at this depth. XPeng has indicated it is targeting roughly one million Turing chip shipments in 2026 and plans to open its robotaxi SDK to additional fleet operators over time. The timing is awkward in one respect. Monday's production-line milestone comes against a softer near-term backdrop. XPeng has guided first-quarter 2026 deliveries to 61,000-66,000 units, a year-on-year decline of roughly 30% that the company has attributed to reduced government subsidies and a longer Lunar New Year. The company reports Q1 earnings on Wednesday. Its first-ever quarterly net profit, RMB 380m posted for Q4 2025, arrived alongside that softer Q1 guide. It is also coming at a moment when the operational risks of large autonomous fleets have become more visible. In late March, more than 100 Baidu Apollo Go robotaxis froze mid-traffic in Wuhan in a correlated software failure, stranding passengers and surfacing a category of risk, simultaneous failure across a managed fleet, that conventional vehicle regulation was not designed for. Pony.ai, despite reaching city-level profitability in Guangzhou, has had to contend with similar concerns as it scales. XPeng enters the same operating environment with a hardware advantage and a software stack it has built and tested itself, but it also enters as a software operator and not just as a manufacturer, with the same governance questions about fleet-scale failure modes that the existing players are still working through. What XPeng has done this week is start the line. The first car has been built. The second-half pilots will test whether the platform performs in mixed traffic with a safety driver on board, and the early-2027 target will test whether the company can remove that driver. Hundreds to thousands of units over 18 months is a deliberately modest ramp by the standards of Baidu and Pony, both of which are targeting fleets in the low thousands by the end of this year. XPeng is betting that integration, owning the chip, the software, the platform, and the chassis, will produce a more reliable robotaxi than the retrofitted models that currently dominate the market.

XPeng (XPEV) rolls first mass-produced robotaxi off the line, a China first

XPeng (NYSE: XPEV) announced today that it has rolled the first mass-produced unit of its robotaxi off the production line in Guangzhou. The milestone makes XPeng the first automaker in China to achieve mass production of a robotaxi built entirely through full-stack, in-house development. The purpose-built vehicle, engineered to L4 autonomous driving standards, is powered by four of XPeng's self-developed Turing AI chips delivering 3,000 TOPS of computing power -- and it doesn't use any LiDAR. It's important to understand what this robotaxi is -- and isn't. The vehicle is built on the same XPeng GX platform that underpins XPeng's $58,000 consumer flagship SUV, which launched last month at Auto China 2026. Both share the same core hardware: four in-house Turing AI chips at 3,000 TOPS, the VLA 2.0 autonomous driving system, Bosch steer-by-wire, and the aviation-grade six-layer safety redundancy architecture. But the robotaxi version is configured specifically for ride-hailing. Where the consumer GX is a six-seat AWD luxury SUV with 750 km of range that a buyer drives themselves, the robotaxi strips out the consumer-oriented layout and replaces it with a passenger-first cabin: privacy glass, gravity seats, rear entertainment screens, and voice-controlled cabin settings. XPeng has announced plans for three robotaxi variants -- a 5-seater, 6-seater, and 7-seater. This is a fundamentally different approach from Tesla's Cybercab or Geely's newly unveiled EVA Cab, both of which are purpose-built robotaxis designed from the ground up without a steering wheel or driver controls. XPeng's bet is that a shared platform between consumer vehicles and robotaxis reduces costs and speeds development -- you validate the hardware in millions of consumer cars while configuring the same platform for autonomous ride-hailing. The robotaxi relies on what XPeng calls a "pure vision solution." There are no LiDAR sensors and no high-definition maps -- instead, the system runs entirely on XPeng's VLA 2.0 end-to-end AI model. VLA 2.0 eliminates the language-translation step found in traditional vision-language-action architectures, compressing system response latency to under 80 milliseconds. XPeng claims the model offers 12x faster inference than the previous generation and roughly 5x better performance than competitors on takeover rates, driving smoothness, and scenario coverage. I tested Xpeng's level 2 VLA 2.0 for consumer vehicles in China last month and I was impressed: XPeng has been building toward this moment methodically. In January 2026, the company secured a road testing permit for L4 autonomous vehicles in Guangzhou. In March, it established a dedicated robotaxi business unit to oversee product development, testing, and commercialization. The timeline from here is aggressive but defined: XPeng plans to begin pilot robotaxi operations in the second half of 2026 to validate the technology, user acceptance, and its business model. The company aims to achieve fully autonomous operations -- no on-site safety officer -- by early 2027. On the ecosystem front, XPeng will open its robotaxi SDK to third-party developers, with Amap (Alibaba's mapping platform) already signed on as its first global ecosystem partner. The robotaxi sector is in the middle of a critical transition from technical validation to commercial scale, and XPeng is entering a crowded field. In the US, Tesla launched its robotaxi service in Austin in June 2025 with human safety monitors and began integrating unsupervised vehicles in January 2026. Tesla has since begun Cybercab production at Giga Texas and expanded its service to Dallas and Houston. Waymo, meanwhile, continues to operate the most mature commercial robotaxi service, with hundreds of thousands of weekly rides across multiple US cities. In China, Baidu's Apollo Go hit 250,000 weekly robotaxi rides by late 2025 -- closing in on Waymo's numbers -- and operates across more than 20 cities. Pony.ai and WeRide each run fleets of over 1,000 vehicles. In China specifically, Geely just unveiled the EVA Cab at the same Auto China 2026 show -- a purpose-built robotaxi with no driver seat, dual Nvidia Drive Thor-U chips at 1,400 TOPS, 43 sensors including LiDAR, and a 2027 deployment plan through CaoCao Mobility's existing 60-city ride-hailing network. The EVA Cab takes the opposite architectural approach from XPeng: a dedicated robotaxi chassis with no provision for human controls, similar to Waymo's vehicles and Tesla's Cybercab. What differentiates XPeng is the full-stack approach combined with platform sharing. While Baidu and Pony.ai are primarily software and fleet operators working with third-party vehicles, and Geely chose Nvidia's chips for the EVA Cab, XPeng designs the chips, the AI model, the vehicle platform, and the manufacturing all in-house. XPeng's 3,000 TOPS from four Turing chips also offers roughly 2x the compute headroom of Geely's Nvidia-based stack. That vertical integration could give it a significant cost advantage as the industry scales. The challenge is that XPeng is late to actual L4 road operations, but it has real-world experience with L2 in its consumer vehicles. Baidu has years of real-world ride data, Tesla's Austin service already has unsupervised vehicles on public roads, and Geely has CaoCao Mobility's existing fleet infrastructure ready to absorb its robotaxis. XPeng's pilot operations haven't started yet. This is a significant milestone, but let's be clear about what it is and isn't. Rolling the first unit off a production line is not the same as running a commercial robotaxi service. XPeng still needs to validate the technology in real-world pilot operations, build rider trust, and prove out the economics -- none of which has happened yet. That said, the full-stack approach is genuinely compelling. The fact that XPeng developed the Turing chips, the VLA 2.0 AI model, the GX vehicle platform, and the manufacturing process all in-house puts it in rare company globally. Tesla is the obvious comparison on the vertical integration front. The pure vision approach -- no LiDAR, no HD maps -- mirrors Tesla's strategy, but XPeng's VLA 2.0 model takes a different architectural path. The sub-80ms latency claim is impressive if it holds up in practice. We think the real test comes in the second half of 2026 when pilot operations begin. The technology can look great on a production line, but robotaxi businesses are made or broken by the millions of edge cases encountered on real city streets. XPeng's goal of removing safety officers by early 2027 is ambitious, and we'll be watching closely to see if the timeline holds.