How XPeng turned driver-assist into an AI business

On 19 March 2021, a convoy of XPeng P7 sedans left Guangzhou carrying journalists and independent testers rather than the company's own engineers. Over the following week the cars covered 3,675 kilometres through ten cities in six provinces, 3,145 of them on highways, changing lanes, overtaking and exiting at motorway junctions under a feature called Navigation Guided Pilot, or NGP, that no other Chinese carmaker had yet put into a production vehicle. By the time the convoy set off, XPeng owners had already logged more than a million kilometres under NGP in the seven weeks since its launch. [1]

Five years on, the same lineage of software has a different job. XPeng licenses a descendant of it to Volkswagen, runs a small fleet of self-built robotaxis in Guangzhou, and spends close to a billion dollars a year on the artificial intelligence behind it. Getting from a highway pilot feature to something worth licensing took three hardware generations, one rewritten software architecture, a government rewriting what the industry could call any of it, and an arithmetic argument for why a research bill this size makes more sense shared with a much bigger carmaker than carried alone.

01Six levels, one confusing word

Cut through the branding and 'autonomous driving' is not one thing but six, defined since 2021 by the SAE J3016 standard that regulators and safety engineers actually use. [2] Level 0 is no driving automation. Levels 1 and 2 keep a human fully responsible at all times: the car can steer, brake and accelerate, but a lapse in attention is the driver's problem, not the vehicle's. Level 3 shifts that responsibility to the car within a limited set of conditions, which is why almost nobody sells it yet: Mercedes-Benz's Drive Pilot remains about the only Level 3 system in commercial use anywhere, and BMW discontinued its own 'Personal Pilot L3' in April 2026, following Mercedes' earlier retreat from a wider rollout, both citing high cost against weak demand. [28] The legal groundwork for Level 3 has existed longer than the business case for it: even the EU, which has had a type-approval framework for automated lane-keeping since 2022, has not seen that translate into a durable consumer product. Levels 4 and 5 remove the human from the driving task altogether, first within a defined area, then everywhere.

Every version of NGP and XNGP that has reached a customer's driveway, from the 2021 highway pilot to the VLA 2.0 rollout in 2026, has been a Level 2 system: increasingly capable, but never, formally, the one driving. XPeng's own marketing has occasionally reached ahead of that classification — both XPeng and Nvidia described XPILOT 3.0 as an 'L3' system when it was previewed in 2019 and launched in 2020 [3][4] — but as the next sections show, Chinese regulators have since made that kind of language a great deal harder to use loosely. The Robotaxi is the one part of the business built to a Level 4 target, and even that ran with a safety operator on board as of mid-2026, with driverless operation targeted for early 2027. That distinction matters for everything that follows: this is the history of a car that still, formally, expects someone to be paying attention.

02From lane-keeping to point A-to-B

XPeng was founded in 2014 and delivered its first production model, the G3, in 2018, running XPILOT 2.0 and then 2.5: adaptive cruise control, lane centring, automated parking. The system took six major over-the-air updates through 2019 to add lane-change assistance, and XPeng's autonomous-driving vice-president, Dr Xinzhou Wu, told Nvidia's GTC China conference that December that XPeng was "the first and only Chinese automaker to have end-to-end closed-loop R&D capability for autonomous driving." [3]

The step change came with the P7 sedan, which began customer deliveries in mid-2020 running XPILOT 3.0 on a single Nvidia Drive AGX Xavier chip rated at 30 TOPS, paired with 12 ultrasonic sensors, five radars and 14 cameras. [4] Navigation Guided Pilot launched on top of that hardware on 26 January 2021, the first point-to-point highway driving feature in a Chinese production car, and the one the March expedition above was built to prove. [1]

03The map problem

City driving was the harder problem, and XPeng approached it the way most of the industry did at the time: with high-definition maps. The lidar-equipped P5 introduced City NGP, limited to Guangzhou, in 2022; the G9, launched on 21 September 2022, unified highway and city assistance into XNGP on dual Nvidia Orin-X chips delivering 508 TOPS combined, with 31 sensors including dual lidar. [5] A more capable version reached the P7i and G9 in April 2023, but only in Guangzhou, Shanghai and Shenzhen. A map had to exist before the car could use it there. [6]

By late January 2024, that map covered 569,000 kilometres of road across 243 Chinese cities, with more than 3.7 million kilometres of cumulative real-world verification behind it. He Xiaopeng told an event that month that the accident rate under XPeng's assisted driving ran at a tenth of manual driving's, with one takeover roughly every 2,000 kilometres of highway driving, and that XPeng's annual AI and smart-driving R&D had reached RMB 3.5 billion (about $490 million) with a team of more than 3,000 people. [7] A month later, on 29 February 2024, XPeng dropped the high-definition-map requirement entirely, citing a 384-fold increase in the speed of adding a new area — from months to one or two days — and expanding beyond those 243 cities to the whole country. Average highway speeds under NGP rose 13%; 94.7% of eligible drivers used the highway feature, against 83.2% for the city portion, and the system was by then driving 40% of all the kilometres its users covered. [6]

Five months later, on 5 August 2024, XPeng removed the last restrictions on routes and road conditions, alongside a Tianji XOS 5.2 software update carrying 484 feature changes; the company said it was by then selling in 30 countries. [8]

XPeng's self-driving system was fantabulously good nearly two years ago when I tested it virtually for a long drive into a busy city.

Zachary Shahan, CleanTechnica, 5 August 2024

CarNewsChina's Mark Andrews, testing the nationwide rollout that same week, was more equivocal: the system worked well in Shanghai in a P7i, he wrote, but not in Guangzhou in a G6. [6]

04Building the brain in-house

XPeng's next move was to stop renting its intelligence. At an event on 6 November 2024, He Xiaopeng unveiled the Turing chip: a 40-core, dual-NPU processor built to run models of up to 30 billion parameters locally, rated at roughly the equivalent of three Nvidia Orin-X chips apiece. [9] The first cars to use it, G7 Ultra models delivered from June 2025, carried three Turing chips for 2,200 TOPS combined; more than 10,000 pre-orders arrived in the first 46 minutes, part of a run that helped lift full-year deliveries from 190,068 in 2024 toward a 380,000-unit target for 2025. [10][11] By the Next P7's debut in February 2026, the same three-chip layout was rated at 2,250 TOPS combined; the Robotaxi unveiled that May carries four Turing chips for an effective 3,000 TOPS.

Thirty TOPS in 2020 to 2,250 in 2026 is a 75-fold increase in five years and nine months — a compound growth rate above 110% a year, or a doubling roughly every eleven months, computed from the manufacturer's own published chip specifications rather than read off a slide. Buying that curve from Nvidia every two years was one strategy. Owning the chip design is another, and it lets XPeng set its own schedule for how much smarter the next car gets rather than waiting on someone else's.

05A model that stops translating

The ground shifted under the whole industry in April 2025. After a high-profile fatal crash elsewhere in the industry involving a car in assisted-driving mode, China's Ministry of Industry and Information Technology convened carmakers at a 16 April meeting on intelligent connected vehicles and set out new rules XPeng and everyone else had been marketing under: no more 'autonomous', 'self-driving' or 'intelligent driving' in advertising, only a plain SAE level number; no more public 'pioneer user' beta programmes without prior approval; remote summoning and unsupervised valet parking banned outright; hands-off-the-wheel detection capped at 60 seconds before the car had to slow down or pull over; and any OTA update touching how the car drives now needs a safety validation report and government sign-off before it ships. [12] None of this reached XPeng's hardware roadmap; it landed entirely on the software: what the company could call the system, and how fast it could ship changes to it. It was the regulatory echo of the map problem two years earlier, a legal boundary where the first had been geographic.

Every driver-assistance system XPeng had shipped through 2024 worked the way most of the industry's did: camera and radar data turned into a language-like intermediate representation, which a separate model then translated into steering and braking commands: vision, then language, then action. At an AI Day event on 5 November 2025, XPeng replaced that pipeline with VLA 2.0, a single model trained on close to 100 million video clips that generates driving actions directly from visual input, skipping the translation step. [13] Pioneer owners got early access that December, now under the new approval regime; a full rollout to XPeng's Ultra-trim models followed in the first quarter of 2026.

XPeng had named Volkswagen as VLA 2.0's first outside customer when it unveiled the model at its AI Day in November 2025. [13] On 2 March 2026, at a media event in Guangzhou, it put dates to the arrangement: Volkswagen would be the system's first customer in China, with global delivery targeted for 2027, even as VLA 2.0 began public road testing on one of XPeng's own robotaxis. He Xiaopeng told the room:

XPENG's VLA 2.0 is the first version designed to achieve full autonomous driving and will iterate at an unprecedented pace. We believe that full autonomy will arrive within the next one to three years, making autonomous driving a natural part of people's daily travel.

He Xiaopeng, XPeng VLA Media Experience Day, Guangzhou, 2 March 2026

XPeng's own measurements from that event put the new system's driving efficiency, tested through Guangzhou's evening rush hour, 23% ahead of the Level 2 systems it replaced. By 3 June 2026, at CVPR in Denver, the company said VLA 2.0 had entered formal mass production and that assisted driving accounted for more than half of all the kilometres its equipped cars covered in the system's first month on the road, alongside a 4,360% gain in the single-job training efficiency of its underlying model over the preceding year. Behind that headline figure, driver interventions per 100 kilometres fell 25.87% month on month, and trips completed entirely under the system's control rose 27.84%. [14][24]

During China's five-day May holiday alone, XPeng owners logged 84.46 million kilometres under the system at a 93.21% daily adoption rate, with the single longest continuous trip reaching 5,441 kilometres. [14] Engagement on that scale is what a driver-assistance business now has to publish to be taken seriously, and what a partner would want to see before licensing the technology under its own badge.

06Who pays for the brain

None of this was free to build, and for most of the past decade the bill has been paid the same way: XPeng doesn't sell driver-assistance as a subscription so much as it sells a pricier trim and includes the software. 'Max' and 'Ultra' trims carrying the extra sensors and chips have commanded a premium over standard versions of the same model for years, and even after XNGP's usage restrictions disappeared in August 2024, that trim premium remained the actual toll gate.

The bill behind that gate has been rising. XPeng's total R&D spending ran to $735.3 million (RMB 5.28 billion) in 2023, then climbed 22.4% to $899.6 million in 2024. [15] The company's 2025 plan, spanning XPeng and its AeroHT flying-car unit, calls for roughly $1.32 billion in total, of which about $626.6 million is earmarked for AI specifically. [16] All figures here are converted at a single reference rate of RMB 7.1812 per US dollar, the 2024 average, so the trend reads like-for-like rather than blurred by year-to-year currency moves. [17]

Isolate just the AI slice against XPeng's own delivery volume and the arithmetic gets uncomfortable. The RMB 3.5 billion He Xiaopeng said XPeng spent on AI and smart-driving R&D in 2024, divided across the 190,068 vehicles it delivered that year, comes to roughly $2,564 (RMB 18,414) of AI research spending per car, computed directly from those two disclosed figures. [7][11] Volkswagen Group delivered 9.03 million vehicles worldwide in 2024, 47.5 times XPeng's own volume. [18] Spread the same fixed RMB 3.5 billion bill across a base that size instead, and the per-vehicle cost falls to about $54 (RMB 388). That figure is an illustration, since no licence fee has been disclosed and Volkswagen pays no such per-car amount; what it captures is why a research bill this size looks different once someone else's factory volume is available to spread it across.

That is the logic behind the sequence running from 2023 to today, though the sequence is not as tidy as 'equity investor becomes licensee' suggests. Volkswagen paid $700 million (then about €632 million) for a 4.99% stake in XPeng in July 2023, alongside an agreement to jointly develop two mid-size EVs. [19] That grew, through 2024, into a shared electrical-and-electronic architecture built for Volkswagen's China Main Platform, and by August 2025 Volkswagen was calling it 'one system for all', extending the same architecture into its combustion and hybrid models built in China, with an upgrade planned for 2027 and a third generation for 2029; the group has said it wants most of its China-built cars running on it by 2030. [20] The first car built on it, the Volkswagen ID.UNYX 08, rolled off the line in March 2026. What Volkswagen kept, at almost the same moment, was the software: it had already become the first outside customer for VLA 2.0, XPeng's driving system, unveiled the previous November. Volkswagen stepped back from co-engineering the architecture at close to the same moment it agreed to buy XPeng's driving software outright — one relationship offering a real answer to which half of this business a much bigger carmaker actually wanted to pay for.

A third revenue line is newer still, and further from proven. The Robotaxi that rolled out in Guangzhou on 18 May 2026 runs XPeng's own Turing silicon and VLA 2.0 software, without lidar or high-definition maps, at a claimed response latency under 80 milliseconds. XPeng has opened a software development kit for the platform, with the mapping service Amap signed as its first ecosystem partner, and is targeting paid pilot operations through the second half of 2026, with a target of removing the safety operator entirely by early 2027. The fleet reached closed-beta testing on 9 July 2026, when He Xiaopeng took its first ride himself outside XPeng's Guangzhou headquarters, alongside internal staff and local riders — a step short of fare-paying service. [25] Third-party estimates of what that market could be worth vary by an order of magnitude depending on the assumptions: Goldman Sachs sees China's robotaxi market growing from $54 million in 2025 to $47 billion by 2035; UBS puts the ceiling, if every taxi and ride-hailing car in the country were eventually replaced, as high as $183 billion a year. [21][22] Neither number is a forecast for XPeng specifically, and the Guangzhou fleet still runs with a person in the driver's seat and no fare-paying passengers yet. In the same stretch of July 2026, He Xiaopeng personally completed a European acceptance test for VLA 2.0 in Germany, and XPeng launched the L03 globally from Munich on 16 July with Google Maps built in for navigation and future driver-assistance features outside China, becoming the first APAC automaker to ship with Google's Auto SDK — the geographic half of the same expansion, running in parallel with the licensing and robotaxi halves. [26] Days later, Bloomberg tied that push to a broader shift: with growth slowing in China's domestic EV market, XPeng is now aiming for half of its volume to come from overseas. [27] On 1 April 2026, XPeng Motors had already dropped 'Motors' from its Chinese name, becoming XPeng Group, as plain a signal as a company can send about which of these businesses it thinks it is now running.