XPENG Drives Physical AI To Next Level

XPENG reported second-quarter revenue of $2.91 billion, up 8.0% year-over-year and 51.5% from the first quarter, with gross margin rising to 20.7%, though vehicle margin fell to 12.1% and the company posted a net loss of $200 million due to a 35% increase in R&D spending. The company also launched VLA 2.0 Version 6.3.0, an over-the-air update featuring Infini-VLA, Streaming Inference, and X-Foresight, claiming a 300% improvement in latency and response speed for its intelligent driving system.

XPENG Drives Physical AI To Next Level Support CleanTechnica's work through a Substack subscription https://cleantechnica.substack.com/subscribe , on Patreon https://www.patreon.com/cleantechnica , or on Stripe https://cleantechnica.fundjournalism.org/contribute/ . Help us produce all of the high-quality, original content we publish week after week https://cleantechnica.com/2026/07/14/10/ despite the challenges of content-scraping AI, antisocial media, inflation, and other hurdles.The next evolution XPENG’s physical AI is beginning to emerge. Announcements over the past week indicate major developments for IRON humanoid robot, VLA 2.0 intelligent driving and the business overall. Many of these developments have not yet been realized in revenue and profits, but are poised to drive the rapidly expanding technology company to the next level in their development. Earlier last week, XPENG released their 2Q financial results. Revenue was $2.91 billion, up 8.0% YoY and 51.5% over 1Q. Gross margin hit 20.7%, up from 17.3% in 2025. For comparison, Tesla had a gross margin of 16.8%. However, a large portion of XPENG’s gross margin came from services, with vehicle margin falling to 12.1%, down from 14% last year. Part of that fall comes from costs associated with launching new models, but the services profitability that drove up overall gross margin is perhaps more interesting. The bulk of the services revenue comes from “technical R&D services” supplied to Volkswagen Group. In terms of profits, XPENG truly seems to be becoming a technology company, rather than just an automobile manufacturer. However, technology also hit their bottom line for the quarter. Overall, they posted a net loss of $200 million. Most of that loss was due to a 35% increase in R&D expenditure. In addition, administrative and selling costs were up, largely due to their global expansion and multiple launches of new models. Those impressive launch events, like the July L03 launch that I attended in Munich, can get expensive. However, with models still ramping up and deliveries starting, the business impact of those R&D and marketing investments will increasingly be felt in the overall business moving forward. XPENG Launches VLA 2.0 Version 6.3.0 On Thursday, XPENG announced the first major upgrade since the release of VLA 2.0, to rollout as an OTA update in the coming weeks. When making improvements to intelligent driving systems, they need to be safe, reliable and react quickly. And the model needs to operate within limitations of latency, computing power and power consumption. It is not enough to offer a software upgrade that cannot work well on existing hardware. Beyond understanding the space around it, the updated VLA 2.0 model incorporates time to perceive in 4D. Factoring in what has happened to predict what will happen. This means the model needs to process a very long temporal sequence. Infini-VLA Supports infinitely long historical timelines while decision making. However, historical memory is limited to 30 seconds, as XPENG found that more is typically unnecessary when driving. However, a moving vehicle does not have the luxury of being able to stop to make sense of the world around it or plan its next action. Updated Streaming Inference enables seeing, thinking and outputting trajectory tokens simultaneously and continuously. This lets XPENG improve latency and response speed by a claimed 300%. X-Foresight uses historical data to infer and predict 6 seconds into the future, providing “proactive reasoning.” The development of this technology was announced in June https://cleantechnica.com/2026/07/01/xpeng-unveils-x-mind-empowering-autonomous-driving-with-a-future-foresight-brain/ , but it is now launching to the public. It can support predictions up to 21 seconds in the future but is limited to 6 seconds to conserve computing power. Meanwhile, Flow-Matching converts data into multiple future paths and makes probabilistic decisions for the best outcome. All this combines for a 20x improvement in safety performance. Capabilities “better than human” for the human-like intelligent driving system. XPENG offers L4 Robotaxi operation using the same hardware. Over 2000 Robotaxi https://cleantechnica.com/2026/05/20/xpeng-starts-producing-riotaxis/ orders were completed on public roads in the past period to internal customers. Of course, Chinese intelligent driving regulations are stricter than those in the US, but broader commercialization is expected soon. While the most advanced intelligent driving capabilities are reserved for models with multiple in-house developed Turing chips, VLA capability will also be offered on single chip models. However, when going from two chips to one, it is not simply half the capability. Using HybridViT, the system streamlines processing needs but retains as much of the underlying capabilities as possible, minimizing tradeoffs. XPENG made changes to the underlying architecture, rather than simply trimming features. Master Agent uses an “Omni multimodal model” to allow for human-like interaction and full voice control of the vehicle. XPENG claims that they effectively have turned the voice interaction into a robot. The AI system utilizes a dedicated Turing chip to process voice interactions locally and interpret intent in context of the surrounding environment. Using conversational interaction, the system can identify destinations and routes. It does not require the user to say the correct pre-set commands. Changes and modifications can be made while travelling. If you ask it to pull over when it finds a safe spot, it will find a suitable spot and pull over. No need to navigate through touch screens or type out destinations. In a video, a user asked it to navigate to a restaurant based on the type of food with a description of a hard to pronounce name and to stop at a big black building in a neighborhood. Virtually all functions of the car can be controlled by voice. It is the closest to KITT https://cleantechnica.com/2026/07/10/xpeng-robotaxi-employee-tests-tesla-competition-us/ from Knight Rider that we have seen so far and seems to be a dramatic departure from the typical voice assistants that many of us are used to. Of course, we will need to experience it in person to know for sure. XPENG sees voice control as a requirement for Robotaxi, providing safe operation and a needed level of interaction when a driver is no longer present. And the strong generalization capabilities of the model let it quickly adapt to new markets. With the most recent announcement, the usage of the multi-Turing chip configurations becomes clearer. The first chip is for intelligent driving, with the second chip expanding capabilities. A third chip is dedicated to voice control and communication. On Robotaxi models, a fourth chip is for redundancy, providing added safety when operating without a driver. Each of these chips has 750 TOPS, which is more than the total computing power of Tesla HW4. With more vehicles on the road, more data is collected. Training data has increased to 110 million video clips. Meanwhile, X-World https://cleantechnica.com/2026/04/28/xpeng-releases-world-model-technical-report-powering-vla-2-0-model-rd-and-verification/ has increased simulation models generated per day by 290% from June. All this data and simulation helps the model address a wide range of edge cases, including navigating constructions zones and taking ferries, as seen in the video clips below. Ultra and Ultra SE models will receive the new generation VLA 2.0 update in September. The Max trim with a single Turing chip will also receive VLA Lite updates next month. Older models with dual NVIDIA Orin chips will also receive an update later this year. No word yet on the single Orin vehicles. Overall, XPENG is offering capabilities to existing customers. They “don’t want to simply reduce the number of model parameters, but rather give everyone a more consistent experience.” And they are offering greater capabilities. Capabilities beyond what they initially promised customers. It is a stark contrast compared to a company that charges extra to unlock capabilities within their existing hardware. A dramatic difference compared to a company that over promised and have yet to deliver. Having sampled VLA 2.0 on several https://cleantechnica.com/2026/04/30/xpeng-p7-with-vla-2-0-a-sporty-drive-that-can-confidently-drive-you/ occasions https://cleantechnica.com/2026/07/30/xpeng-vla-2-0-in-munich-taking-global-intelligent-driving-lead/ it was already impressive with its humanlike https://cleantechnica.com/2026/05/24/xpeng-offers-more-human-like-autonomous-driving/ driving and seemed to learn fast. If the new updates live up to their promise, XPENG could noticeably pull ahead of competitors in intelligent driving. Of course, by not charging for subscriptions, the revenue impact of these updates will not be immediate. However, once many potential customers experience the capability, I have a feeling that they will find it compelling. That is poised to drive vehicle sales and sales of higher trim models with the most advanced capabilities, as well as sales of technology to other automakers seeking to offer similar capabilities. XPENG is delivering beyond what many expect would be available from their hardware. But intelligent driving isn’t the only area where XPENG is challenging expectations. IRON Cash Infusion Fuels Commercialization XPENG’s robotics division Dogotix raised $900 Million, for a $6.3 Billion valuation, which they claim is “the largest single-round private financing ever recorded in China’s embodied AI industry.” In 2026, the robots are to be deployed at XPENG stores for sales support, with deliveries to external customers in the retail and service sectors set to start in 2027. Production capacity of several thousand units per month will then scale with demand. The focus on human interaction makes sense, as IRON in one of the most human-like robots in form and design. In addition, a humanoid robot is not the ideal form factor for many industrial applications. However, when considering that Alibaba and Tencent were listed as strategic investors, this also opens up other possibilities. As the companies behind Alipay and WeChat Pay that dominate day-to-day transactions in China, I could see robots increasingly taking on sales and customer service roles. The separate funding means that Dogotix will still be controlled by XPENG, but it will no longer be a wholly owned subsidiary. This means additional reporting and additional visibility into the performance of the robotics division. As sales revenue begins growing next year, it will be interesting to see the margins and profitability. XPENG indicated that lifetime revenue per unit between sales and upgrades could be higher than automotive. However, they still use the same infrastructure and software systems across different platforms to maximize R&D spending. While the applications may look different, XPENG used the analogy of an iceberg, stating that the 95% that you don’t see is the shared infrastructure. But There’s More Based on the presentation, XPENG has taken the approach of focusing on a complete physical AI R&D system. Tackling the difficult problems first, with the end in mind. Taking the perspective that when the difficult problems are addressed first, that the easier problems will be increasingly easy to address. Not going after the low-hanging fruit first clearly didn’t generate the immediate results sought by some short-term investors. Given the rapid speed of competitive industry in China, this was likely also not an easy commitment to make. XPENG is also rapidly iterating to solve the “Unknown Unknown” problems. Problems that nobody can reasonably anticipate in planning. Addressing the edge cases within edge cases. XPENG claims that only by building a complete AI system can they rapidly iterate to solve those unknown problems. XPENG mentioned that this focus creates an “AI Flywheel”: Better cars generate more data, that creates better software, that then sells even more cars, that then cover more edge cases, that creates better models, that then extend to other products… Local operation was needed to work within anticipated intelligent driving regulations, requiring effective use of available computing power and an efficient AI model. Other systems from competitors rely more heavily on sensors to overcome a lack of processing power or rely on centralized data centers that run into privacy and regulatory challenges. By creating an efficient AI model that can run locally, the potential applications scale beyond just physical AI. That creates the potential to apply the Turing chip and AI model to any ecosystem. At the VLA 2.0 6.3.0 launch, XPENG announced XLLM architecture, enabling a large language model LLM to efficiently run on a single Turing chip. XPENG claims XLLM offers efficient inference, similar to ChatGPT. The system optimizes “a currently common large language model” turning it into a local application with a generation rate of over 20 tokens per second. While not the fastest token generation of any LLM, it is a useable speed and all the processing happens on a single chip, rather than in a massive, power-hungry data center. Using XLLM technology, we could see LLM capabilities offered as a local application in a wide array of potential use cases. We could even see AI data centers using the optimizations to become far more efficient, with the potential to dramatically reduce energy consumption. As such, solving how to locally run physical AI in a car first is opening up new applications for the technology. Solving the most difficult challenges first also opens up emerging revenue streams. By being ahead on technology, technology services can increasingly become a major source of income. With Porsche teaming up with XPENG on emissions pooling in Europe another potential revenue source the Volkswagen Group technology partnership is likely to deepen, and we could see VLA 2.0 offered in a growing number of Volkswagen Group vehicles globally. With the success of that partnership, other partnerships could open. In addition, as the intelligent driving system is tied to electric vehicles, sales motivated by intelligent driving capabilities natural will drive EV volume. This increased scale helps support EV platform development. The intelligent capabilities also help drive global vehicle electrification. Being ahead also has broader benefits. XPENG is positioned to be the first to meet harmonized UN DCAS regulations and provide truly global intelligent driving. As customers experience the benefit of intelligent driving, more will be likely to consider XPENG cars that offer the capability, increasing sales revenue. Other automakers may also want to offer that capability, increasing technology services revenue. The evolving AI capabilities, driven by better data, can then further improve and be reapplied to other applications. With any technology company, investment is based on what is coming next. XPENG is on the verge of turning long-term R&D into revenue across several products. That revenue could lead to significant profits and/or be used to develop even more technologies. This is not intended to be investment advice, but there is potential for both the business and technology to develop. It will be interesting to see where this current period of evolution leads with XPENG. Sign up for CleanTechnica's Weekly Substack for Zach and Scott's in-depth analyses and high level summaries https://cleantechnica.substack.com/subscribe , sign up for our daily newsletter https://mailchi.mp/cleantechnica/daily-newsletter , and follow us on Google News https://news.google.com/publications/CAAqLQgKIidDQklTRndnTWFoTUtFV05zWldGdWRHVmphRzVwWTJFdVkyOXRLQUFQAQ Have a tip for CleanTechnica? Want to advertise? Want to suggest a guest for our CleanTech Talk podcast? Contact us here https://cleantechnica.com/contact/ . Sign up for our daily newsletter for 15 new cleantech stories a day https://mailchi.mp/cleantechnica/daily-newsletter . Or sign up for our weekly one on top stories of the week https://mailchi.mp/cleantechnica/weekly-newsletter if daily is too frequent. CleanTechnica uses affiliate links. See our policy here https://cleantechnica.com/cleantechnica-editorial-ethics/ . CleanTechnica's Comment Policy https://cleantechnica.com/cleantechnica-comment-policy/