The lines separating software companies from traditional automakers are increasingly indistinct. At the CVPR 2026 conference held in Denver, the Chinese company XPeng demonstrated its identity as an artificial intelligence (AI) firm rather than merely a vehicle manufacturer. Dr. Xianming Liu, head of XPeng's General Intelligence Center, revealed that the company invests approximately 300 million RMB, or about$41 million, each month to train AI models.
This substantial expenditure totals nearly $500 million annually, a striking figure for a company that sold around 200,000 electric vehicles last year. Nonetheless, XPeng's significant financial commitment is aimed at transitioning its VLA 2.0 software into mass production.
During the conference, Dr. Liu highlighted a distinctive choice in XPeng's software architecture, noting that language can serve as a bottleneck in autonomous driving systems. The first-generation system processed visual road data through language tokens before determining driving actions; however, VLA 2.0 eliminates this intermediary step.
XPeng engineers recognized that its system captures approximately two billion visual tokens per second from onboard cameras but requires only 10 to 20 tokens to control steering and acceleration. The necessity of translating roadway information into written language causes unnecessary computational burdens and introduces potentially dangerous delays. Instead, VLA 2.0 utilizes language merely as an input, enabling drivers to issue verbal commands that the vehicles can easily understand.
This innovative approach is already operational in consumer vehicles, as XPeng successfully transitioned VLA 2.0 from research to mass production. Within the first month of rolling out over-the-air updates, the software managed over 50% of the total assisted driving mileage among users.
The architecture accommodates both Level 2 driver assistance and Level 4 fully autonomous systems. Indeed, XPeng leveraged its new GX platform to create its first robotic taxi, which is already in production. This specialized vehicle boasts remarkable onboard computing power of 3,000 TOPS, enabling it to navigate intricate driving environments without human intervention.
XPeng does not depend exclusively on VLA 2.0 to navigate its electric vehicles. The software works in conjunction with what the company refers to as a "world model." While VLA 2.0 learns driving behavior by analyzing millions of hours of human actions, the world model focuses on understanding fundamental physical principles. It predicts changes in the surrounding environment and anticipates the movements of other drivers.
To fortify this system, XPeng created three specialized subprograms: X-World, X-Foresight, and X-Cache. X-Cache notably minimizes redundant computational processing by about 70% with minimal impact on image quality, enabling the system to operate 2.7 times more efficiently.
Many companies disagree on the optimal sensor configurations, yet XPeng has adopted a balanced approach. Models such as the P7 and G7 utilize a vision-only system for their primary driving intelligence, resembling the strategy of companies like Tesla, without completely abandoning traditional safety sensors.
Each electric vehicle is equipped with three mmWave radars and 12 ultrasonic sensors. Dr. Liu clarified that these additional sensors do not communicate with the primary AI driving system. Instead, they function independently, forming a separate safety system for automatic emergency braking and steering. This setup ensures that if the vision AI fails, a secondary physical layer is in place to protect passengers.
To optimize the functionality of these systems, XPeng is pushing the boundaries of the AI "Scaling Law," which posits that models become more intelligent as they receive more data and computational resources. The current VLA 2.0 system comprises billions of parameters and trains on hundreds of millions of video clips, processing over four trillion tokens during each update cycle.
Despite the scale of these operations, XPeng has effectively controlled its costs by enhancing efficiency. Over the past year, the company improved training efficiency for individual tasks by 4,360% and increased GPU hardware utilization from 40% to 90%. This efficiency maximizes the potential of the company's RMB 47.66 billion ($6.5 billion) cash reserve.
The rapid advancements of the software prompted a high-stakes friendly wager within the company. Last year, XPeng's CEO He Xiaopeng publicly bet Dr. Liu on the performance of VLA 2.0: if the system does not match Tesla's Full Self-Driving software by August 30, 2026, Dr. Liu must run naked across San Francisco's Golden Gate Bridge.
Fortunately for him, Dr. Liu is confident he will not have to fulfill this embarrassing bet. He indicated that internal testing in busy cities like Beijing demonstrated that VLA 2.0 has already reached parity with Tesla’s older systems and often surpasses newer versions, thanks to the intricate driving data available on Chinese roadways.
XPeng views its electric vehicles as tangible hardware designed to showcase its engineering capabilities and to gather data for its advanced VLA software. This perspective elucidates the company's recent decision to license the VLA 2.0 architecture to Volkswagen, which plans to implement it in vehicles set to launch in 2027.
Additionally, XPeng intends to extend its AI technologies beyond the automotive sector and is preparing its "IRON" humanoid robot for mass production by the end of 2026. The ambition is to employ these robots as shopping assistants in its retail locations beginning in the first quarter of 2027, thus evolving the company from a traditional automaker into a physical AI enterprise.
Source
