Prepared remarks
Hello, ladies and gentlemen. Thank you for standing by for the Fourth Quarter and Fiscal Year 2025 Earnings Conference Call for XPeng Inc. Today's conference call is being recorded. I will now turn the call over to your host, Mr. Alex Xie, Head of Investor Relations and Capital Markets of the company. Please go ahead, Alex.
Thank you. Hello, everyone, and welcome to XPeng's Fourth Quarter and Fiscal Year 2025 Earnings Conference Call. Our financial and operating results were issued by newswire services earlier today and are available online. You can also view the earnings press release by visiting the IR section of our website at ir.xiaopeng.com. Participants on today's call from our management team will include Co-Founder, Chairman and CEO, Mr. He Xiaopeng; Vice Chairman and President, Dr. Brian Gu; Vice President, Mr. Charles Zhang; Vice President of Finance and Accounting, Mr. James Wu; and myself. Management will begin with prepared remarks, and the call will conclude with a Q&A session. A webcast replay of this conference call will be available in the IR section of our website. Before we continue, please note that today's discussion will contain forward-looking statements made under the safe harbor provisions of the U.S. Private Securities Litigation Reform Act of 1995. Forward-looking statements involve inherent risks and uncertainties. As such, the company's results may be materially different from the views expressed today. Further information regarding these and other risks and uncertainties is included in the relevant public filings of the company as filed with the U.S. Securities and Exchange Commission. The company does not assume any obligation to update any forward-looking statements, except as required under applicable law. Please also note that XPeng's earnings press release and this conference call include the disclosure of unaudited GAAP financial measures as well as unaudited non-GAAP financial measures. XPeng's earnings press release contains a reconciliation of the unaudited non-GAAP measures to the unaudited GAAP measures. I will now turn the call over to our Co-Founder, Chairman and CEO, Mr. He Xiaopeng. Please go ahead.
Hello, everyone. Reflecting on 2025, XPeng delivered a series of milestone breakthroughs across multiple fronts. Our annual delivery reached 429,445, up 126% year-over-year. MONA M03 became the best-selling battery electric sedan in the RMB 100,000 to RMB 200,000 segment. P7+ ranked #1 among pure electric sedans in the RMB 150,000 to RMB 200,000 segment. The Kunpeng Super Extended Range EV X9 entered mass production and started our one vehicle, dual energy era. Our quality management system and customer satisfaction as measured by NPS improved rapidly. Overseas deliveries nearly doubled to 45,000 units, with revenue from overseas markets contributing over 15% of total revenue. In addition, benefiting from enhanced organizational capabilities centered on physical AI, we successfully brought our Turing AI SoC into mass production and began deliveries to Volkswagen. Our VLA 2.0 saw an emergence and the debut of our humanoid robot IRON garnered significant attention globally. In 2025, we achieved not only high growth in scale, but also significant improvement of operational capabilities. Our gross margin for fiscal year 2025 reached 18.9%, an increase of 4.6 percentage points year-over-year. In response to national initiatives, we substantially reduced the turnover days of accounts payable to suppliers by 50 days. We had a free cash flow inflow of approximately RMB 5 billion in 2025 and ended the year with RMB 47.7 billion cash on hand, providing robust financial support for unwavering investment in physical AI R&D. In Q4, XPeng achieved its first ever quarterly profit with net profit exceeding RMB 380 million. This marks the emergence of a business model driven by technology leadership, which is a path to profitability distinct from that of traditional automakers. I'm glad to see that with years of hardware investment and physical AI transformation, XPeng is at a pivotal inflection point in the application of physical AI. After years of sustained and significant R&D investment, we have built a full stack in-house developed technology system, including SoC, foundation model, data, EE architecture and AI infrastructure. We're now driving upward integration across autonomous driving and smart cockpit systems while enabling cross-domain integration downward into powertrain and chassis systems. Looking ahead, we'll continue to push the technological frontier, accelerate the development of AI, expedite deployment and commercialization of our product technology innovations. In early March 2026, our VLA 2.0 successfully passed the physical touring test for autonomous driving and passengers can hardly distinguish if it's a human or AI who's driving the car; the vehicle has effectively evolved into a physical AI agent. XPeng's VLA 2.0 will soon cross the inflection point of both technology and large-scale deployment. Fully autonomous driving can be expected to come in the next 1 to 3 years. AI-powered vehicles and humanoid robots will soon fundamentally reshape how everyone travels, works and lives. 2026 marks a major year of product portfolio expansion and upgrade in our product capabilities. We plan to launch 4 new models, expanding into both large and compact sized vehicles. Designed by our newly restructured design team, these models are crafted for global markets and built on dual energy platforms, supporting the evolution of autonomous driving capabilities from L2+ to L4. As deliveries of new models ramp up, we'll achieve strong quarter-over-quarter growth in volume. In Q3 2026, we'll start preorders of our first flagship 6-seat full-size SUV, the XPeng GX, bringing in a truly uncompromised best-in-class experience model to family users. The GX delivers MPV comfort and spaciousness and features steer-by-wire and rear-wheel steering. It will be our first model designed to support L4 level hardware and software capabilities. Starting from this year, we'll further expand our AI vehicles' global market share and bring our VLA 2.0 model to global markets. At the same time, our lead teams will bring both robotaxi and advanced humanoid robot into mass production. With global product sales increasing, in 2026, we'll further solidify our global scale-up capabilities across production, supply, sales and service. Our supply chain, manufacturing, logistics, spare parts, sales and service systems are moving towards greater global synergies, underpinning sustained improvement in quality and brand perception in international markets. Our overseas deliveries goal is set to double in 2026 year-over-year with international business contributing over 20% of total revenue. We plan to introduce 4 new models to global markets. In the key SUV segment, we aim to launch more global flagship models. We target to have 680 overseas stores for sales and service, doubling the number of stores from the end of 2025. XPeng's self-operated ultrafast charging network will also expand beyond China to 10 key international markets. This will bring our industry-leading 5G ultrafast charging experience to users globally. With enhanced overseas competencies, our global expansion will further accelerate significantly through 2027 and 2028 and revenue from overseas markets will become one of the core drivers of the company's profitability. To make the leap from L2+ to L4 and enable global deployment, we have fundamentally redesigned the architecture of our autonomous driving technology. Starting yesterday, we began the gradual rollout of our VLA 2.0 to users. This is our first release designed for the L4 autonomous driving era, delivering a safe and smooth experience with driving performance comparable to experienced human drivers and representing a generational gap ahead of the industry. The VLA 2.0 is transforming advanced autonomous driving from an early adopter feature into a truly mainstream mass market feature that even everyday drivers can trust, rely on and enjoy with confidence. All of our 732 stores nationwide now offer VLA 2.0 test drives. So far in March, the number of our daily test drives has doubled month-over-month and the percentage of the Ultra and Ultra S trims among all trims has also more than doubled. Over the next 3 to 6 months, VLA 2.0 will notably improve test drive conversion rate and substantially increase user engagement and retention for XPeng. The VLA 2.0 has broken down traditional rules and validated the applicability of scaling laws in the physical world. Next, we will continue to scale up to aggressively widen our lead. By the end of the year, we target to increase the number of parameters on the edge from several billion to the 20 billion level. This will increase average miles per takeover by 25x and the safety-critical miles per takeover by 50x. At this pace of improvement, the era of fully autonomous driving is on track to emerge within the next 1 to 3 years. Autonomous driving will become part of everyday mobility. Our GX Robotaxi powered by VLA 2.0 has received official road testing approval in Guangzhou and is now conducting ongoing L4 corporate road tests. In the second half of the year, we plan to launch pilot passenger operations for our mobile taxi service to validate the technology, user experience and the business model. We'll also begin overseas road testing of the VLA 2.0. As technology advances and regulatory frameworks evolve, XPeng will be one of the few companies globally that is capable of scaling autonomous driving efficiently across multiple global markets. The strong performance of our VLA 2.0 is underpinned by the joint optimization of our in-house Turing SoC, proprietary foundational large model and compiler, delivering a 10x improvement in effective compute even though the plan was only to achieve 3x improvement. Since entering mass production and being deployed in vehicles in Q3 last year, the XPeng Turing SoC has actually shipped over 200,000 units. Starting from the second quarter of this year, all XPeng models, including the Max trims, will fully transition to shipment with our in-house Turing SoC. Shipments of Turing SoC are targeted to reach nearly 1 million units this year. The Turing SoC is positioned to become the leading high-compute AI SoC deployed on the edge by shipment volume. Volkswagen is our first external customer for both our Turing SoC and VLA 2.0. Our deep collaboration with leading global automakers, together with our ability to mass produce efficiently, demonstrates that our technology is scalable, replicable and globally competitive. We welcome more automakers, OEMs and Tier 1 suppliers to adopt our Turing SoC and integrate intelligent solutions for their customers. In my view, the ultimate competition in physical AI will be determined by fundamental organizational capabilities and AI infrastructure. Recently, we completed a critical organizational upgrade, integrating our autonomous driving center and smart cabin center into the General Intelligence Center. This is more than a structural change. It represents a paradigm shift in building intelligent systems. Driving decisions and human-vehicle interactions will no longer exist in isolation. How the vehicle drives on road and how humans interact with it now share the same physical AI foundation model and infrastructure. R&D efficiency and effectiveness have improved substantially. XPeng has established a fully integrated AI infrastructure stack, including computing power on the cloud and on the edge, data compilation, quantization, deployment, simulation and real-world testing at great scale. Our physical AI foundation model will evolve at an exponential speed. It will power vehicles, robotaxi and humanoid robots and become a unified super foundation model for all of our physical AI agents. Our next-generation humanoid robot, IRON, is targeted to enter mass production by the end of 2026. Powered by 3 Turing AI SoCs, IRON's computing power on the edge far outpaces other robots in the industry. Our technology focuses on the robot's brain and cerebellum. The brain enables independent thinking and decision-making while the cerebellum controls emotions. At the same time, we continuously accumulate data for embodied intelligence models. Our VLA 2.0 technology stack is now running successfully on our robots. Paired with the fourth-generation motion control system, IRON will deliver industry-leading agility and motion control in the second half of the year, establishing a clear generational lead in intelligence capabilities. IRON will focus on 3 key application scenarios: commercial, industrial and household. Initial deployment will support reception, guidance and retail assistance across XPeng stores and campuses in China and overseas. In the first quarter of the year, we began construction of our humanoid robots mass production base in Guangzhou. IRON aims for a monthly production target of over 1,000 units by the end of this year. Leveraging IRON's advanced intelligence and our strength in mass production with quality and supply chain management, XPeng will become one of the world's largest and most valuable humanoid robot companies. Because we already saw the tremendous disruption that the physical AI world has brought to the original software development paradigm in 2025, we invested RMB 9.5 billion in R&D, including RMB 4.5 billion in AI. Our sustained and efficient investment in AI R&D over the past few years has enabled us to build an industry-leading fully in-house physical AI technology stack, including SoCs, foundation models and AI infrastructure. Powered by this technology stack, the pace of advancement in physical AI will accelerate significantly starting 2026. We'll see the mass production and application of physical AI agents with their scale poised to enter a steep growth curve. Over the next 5 to 10 years, the market for physical AI is expected to surpass that of the automotive industry. Both robotaxi and humanoid robots represent $1 trillion to $10 trillion level global market opportunities. In the future, the sales rankings of physical AI agents may matter even more than the ICE vehicle volume charts a decade ago or NEV volume charts today, and they will change very, very fast. We're confident that XPeng will become the global leader of physical AI agents. Pushing the boundaries of physical AI is very exciting for my team and me. We're more committed than ever to intensifying our R&D investments. This year, in addition to vehicle development, investment in physical AI-related R&D will increase to RMB 7 billion. In my view, this investment will not only help us increase our competitive advantage and deliver substantial long-term returns. Scale allows us to survive in competition, but it is sustained leadership in physical AI technology and commercialization that will define our core competitive advantage. For the first quarter of 2026, we expect deliveries to be between 61,000 and 66,000 units. Revenue is expected to be between RMB 12.2 billion and RMB 13.2 billion. March deliveries are expected to grow sequentially by 69% to 101% month-over-month. As the VLA 2.0 and four new models enter mass production, we expect quarterly sales to continuously trend upward and achieve year-over-year growth in the second half of the year to significantly outpace the industry. We are on the cusp of large-scale mass production of our fourth physical AI agent. Going forward, I will dedicate more of my efforts to develop strategies, establish organization and operational capabilities for XPeng's globalization and commercialization with a focus on translating our technological leadership into commercial success. The innovative business models of physical AI will add network effect and agent effect on top of the traditional automotive economies of scale, meaningfully raising market entry barriers and industry concentration. As a leader in physical AI, we will possess a long-term sustainable advantage, securing a higher share in a vastly expanded market and achieving greater corporate value. Thank you, everyone. With that, I'll now turn the call over to our Vice President of Finance and Accounting, James, who will walk you through our financial performance for the fourth quarter of 2025.
Thank you, Xiaopeng. Now let me provide a brief overview of our financial results for the fourth quarter of 2025. I will reference RMB only in my discussion today, unless otherwise stated. Our total revenues were RMB 22.25 billion for the fourth quarter of 2025, an increase of 38.2% year-over-year and an increase of 9.2% quarter-over-quarter. Revenues from vehicle sales were RMB 19.07 billion for the fourth quarter of 2025, an increase of 30% year-over-year and an increase of 5.6% quarter-over-quarter. The year-over-year and quarter-over-quarter increases were mainly attributable to higher deliveries. Revenues from services and others were RMB 3.18 billion for the fourth quarter of 2025, representing an increase of 121.9% year-over-year and an increase of 36.7% quarter-over-quarter. The year-over-year and quarter-over-quarter increases were primarily attributable to the increased revenues from, firstly, the technical R&D services rendered to the Volkswagen Group due to the successful achievement of certain key milestones in the current quarter; secondly, parts and accessories sales in line with higher accumulated vehicle sales; and lastly, carbon credit trading. Gross margin was 21.3% for the fourth quarter of 2025, compared with 14.4% for the same period of 2024 and 20.1% for the third quarter of 2025. Vehicle margin was 13% for the fourth quarter of 2025 compared with 10% for the same period of 2024 and 13.1% for the third quarter of 2025. The year-over-year increase was primarily attributable to ongoing cost reduction and improvements in product mix of models. R&D expenses were RMB 2.87 billion for the fourth quarter of 2025, representing an increase of 43.2% year-over-year and an increase of 18.3% quarter-over-quarter. The year-over-year and quarter-over-quarter increases were mainly due to higher expenses related to the development of new vehicle models and technologies as the company expanded its product portfolio to support future growth. SG&A expenses were RMB 2.79 billion for the fourth quarter of 2025, representing an increase of 22.7% year-over-year and an increase of 12% quarter-over-quarter. The year-over-year and quarter-over-quarter increases were primarily due to higher commission to the franchise stores related to the sales volume and the launch of new models. The year-over-year increase was further due to higher marketing and advertising expenses. As a result of the foregoing, loss from operations was RMB 0.04 billion for the fourth quarter of 2025 compared with RMB 1.56 billion year-over-year and RMB 0.75 billion quarter-over-quarter. Net profit was RMB 0.38 billion for the fourth quarter of 2025 compared with net loss of RMB 1.33 billion year-over-year and net loss of RMB 0.38 billion quarter-over-quarter. The company recorded positive net profit for a single quarter for the first time in Q4 of 2025. As of December 31, 2025, our cash position was RMB 47.66 billion. To be mindful of the length of the earnings call, I would encourage listeners to refer to our earnings press release for more details on our fourth quarter and full year 2025 financial results. This concludes our prepared remarks. We'll now open the call to questions. Operator, please go ahead.
Questions and answers
Your first question today comes from Tim Hsiao from Morgan Stanley.
I have two questions both are related to smart driving. The first one, what type of major upgrade should we expect for XPeng's VLA 2.0 in the coming months? How do you anticipate VLA 2.0 will impact order conversion and user retention in the following quarters? That's my first question.
Thank you for your question. Let me answer your first question first. So basically, going forward, each quarter, we are expecting at least one major version of OTA. And I think we can share some of it, just to give you some examples. In our Q2 OTA, for example, the autonomous driving will be able to cover more roads, whereas currently it was only able to cover the major highways. In the future, for the Q2 OTA, we'll be able to cover smaller parking lots and campuses or communities in different parks. It is actually a very critical change, meaning that our capability is evolving from navigation-enabled public road to cover more areas. In addition to that, we would also like to highlight that autonomous driving in the future will become more AI agent oriented. If you look at all the scenarios of application for robot systems, it's essentially very similar to that. Also, the four levels of autonomous driving and navigation follow the same logic as well. This year, we also expect to do a lot more upgrades in terms of our capabilities. For example, if you look at our model on the vehicle or the edge side, we are increasing the parameters from the 1 billion level to actually up to the 20 billion level. We also hope that we can increase our miles per takeover by about 5 to 10x. In addition to autonomous capability upgrades, we also have added some multi-language support capabilities as well. So that's another level of upgrade for localization support, which will allow us to combine smart cockpit with autonomous driving. Now because our VLA 2.0 only started to be officially pushed to our users yesterday, and I mentioned in the prepared remarks that we started promoting it in early March, we could actually see that the market reacted very, very positively. For example, our test drives actually more than doubled sequentially across our stores and also Ultra and Ultra S trim sales also more than doubled as well. And I believe when we fully launch our VLA 2.0 to the market and to our users and continue to upgrade it, we are expecting a higher sales volume and also a higher level of conversion rate, which will also eventually raise the average selling price of our vehicles. Now when it comes to user retention or engagement, I would like to share my own case and experience as well. After trying out the VLA 2.0, there's no going back for me, honestly, because I travel a lot and a lot of the times I actually rely on professional drivers who drive me around. After trying VLA 2.0, I realized when I now use other types of autonomous driving versions or when I ride in professional drivers' cars, they are not nearly as good as VLA 2.0 because VLA 2.0 now gives me the smoothness of driving and also peace of mind. That kind of experience is unparalleled. It's been only two to three months, and I'm no longer the same person — there's no going back really. So I believe that in the future, advanced autonomous driving will definitely be part of everyone's daily necessity. Usage rate potentially will reach very high levels and I feel this very, very strongly. Recently, my professional drivers have been asking me a couple of times when their profession might eventually be phased out. I would like to add one brief last point, which is that this represents a completely new paradigm. Our current priority is to really address a lot of safety issues and many of our weaknesses first. However, since the launch of our second-generation VLA, the rules have completely changed. It's never the same again. So I believe that with our weaknesses being compensated for, now we can actually spend a great deal of our time on enhancing our overall strength and deployment globally.
My second question is also related to smart driving. What is the deployment roadmap for XPeng's Ultra model and VLA 2.0 in the overseas market? How will the overseas expansion of smart driving affect XPeng's global sales? And does the smart driving software have the potential to be monetized along the way? That's my second question.
Let me briefly address this question. First of all, we have already launched the preparation for testing our VLA 2.0. By the end of this year or beginning of next year, we are going to gradually roll out the VLA second-generation testing at great scale and also its delivery or deployment across different regions. I believe that VLA 2.0 has a significant advantage overseas. First, we found that it's actually very generalizable. In our testing across multiple markets, we realized that even without any overseas actual road data, VLA alone performed really well and achieved excellent results. Second, compared to some of our competition, for example, FSD out there, our VLA has a particular advantage over them in capabilities across smaller roads and also country roads or uncharted territories. We've tested and seen that in some overseas markets such as Southeast Asia and Europe. That strengthens our conviction that our VLA 2.0 can provide high-quality, safer autonomous driving for our users overseas at higher quality and lower cost. We have done hardware preparation, and so by the second half of this year, we are ready for larger tests and for future launches as well. We are also considering some business model upgrades or transitions for our smart driving software in the overseas market. I believe there are many opportunities for commercialization and monetization, and we have strong confidence in converting our technology into business opportunities and revenue.
Your next question comes from Nick Lai from JPMorgan.
My first question is related to the humanoid robot ambition and long-term strategy. Management mentioned that in the near term we'll see a product launch at the Beijing Auto Show followed by mass production by year-end. Aside from that, are there any major milestones that investors should be mindful of in the next one to two years? In addition, when we expand our footprint from smart vehicles to humanoid robots, what can we leverage regarding R&D, production and the supply chain?
First of all, thank you for your question. I would like to clarify that we said mass production in the second half of this year. What you asked is very big, and I'll try to address it. Last year, many supply chain partners and companies asked us whether humanoid robot supply chains should be car grade. We believe that car grade is the minimum requirement because cars typically have one engine and if it breaks down it can cause serious accidents. Humanoid robots have at least 70 to 80 joints and any kind of damage to those joints will cause information or signal loss and could lead to very serious accidents and compromise safety. So we need to look at humanoid robots with at least car-grade standards. Since the beginning of last year, we started development for our humanoid robot adopting car-grade standards in preparation for manufacture. XPeng is unique in that sense because we are adopting the same high-level rigor in developing robots as we do in developing vehicles. We have to do a lot of in-house development covering not only the joints, but also the torso and other parts of the body — from the SoC to the foundational model, from data to all the training of the brain and cerebellum. We are using a very challenging methodology which can be an order of magnitude harder than developing robotaxi to achieve high-quality mass production. By the end of the year, our targeted production is at least 1,000 units per month. I think in future vehicle companies will also be robotic companies. However, one of the biggest differences between humanoid robots and vehicles is that robots actually generate a lot of value from the software; that value can start at 50% since day one, whereas cars may start slower with only 10% to 15% and take time to ramp up value creation. They come from the same logic and share the same origin. When it comes to R&D, mass production, supply chain management, quality and globalization management — there will be a lot of similarity between car and robot development, and that sets us apart because we have always been doing things in-house with a full-stack technology and cross-domain integration. When we develop our robots from elbow to hands to leg to feet, we do everything in-house or at least joint development, which again sets us apart from other car or robot companies.
My second question is also related to humanoid robots regarding cost and selling price. How should we expect the cost reduction after mass production from year-end? And should we expect any external sales after mass production from year-end?
This is a very good question. When thinking about the cost structure for humanoid robots, it's very different from cars. There are three major cost items for humanoid robots: hardware cost, R&D cost, and operational cost. Hardware cost of a robot is similar to a car. But R&D costs, especially AI-related R&D costs, will be much higher than that of vehicles, because in the AI era you cannot rely on previous rule-based systems and established third-party solutions the way you could in the past. For humanoid robot development, you need substantial in-house AI R&D. Operational costs are also a completely different world because data varies by scenario. When you deploy in a new scenario, you need to retrain the AI and the robot. There are two types of humanoid robots: general-purpose and specialized-purpose. General-purpose robots have transferable skills and general knowledge, while specialized robots require extensive additional training for particular tasks, like investment or medical practice. When it comes to mass production, we will continue to drive down BOM costs for humanoid robots, but software and operational costs will continue to be significant and will improve depending on the model and application scenarios. We are different from many other companies because we target deployment and commercialization early. We will first go into commercial scenarios before industrial and then household. Commercial applications require a lot of motion control, especially for full-body control, which is something we're really good at and leading the industry on. Industrial applications will focus more on hand movement and hand agility, which we are working towards. Household will be the most challenging scenario to deploy humanoid robots in, so we'll gradually move from commercial to industrial to household.
Your next question comes from Tina Hou from Goldman Sachs.
Congrats on a strong set of results. My first question is regarding our overseas expansion. Which key markets will drive 100% volume growth in overseas this year, and looking at 2027 and 2028, what kind of products or technology are we going to introduce to drive further expansion overseas? Also, when will we start to launch our Kunpeng Super Hybrid platform models in the overseas market?
Tina, let me address your question. In terms of key markets, we started our global journey in Europe and the pan-European market represents our largest regional market, accounting for approximately 50% of our overseas volume. We saw our premium-positioned EVs lead in a number of categories, especially in Nordic countries. We also saw very encouraging growth in large markets like Germany, France and Great Britain. So Europe is a key focus for us. After Europe, Southeast Asia last year represented a very important growth market. We saw exciting growth in Thailand, Indonesia and Malaysia, and we continue to see those markets as key growth points. Outside these two large markets, emerging markets like the Middle East, Central Asia and Latin America also have great growth potential; we established operations there last year and see continued penetration leading to good growth prospects. How do we accelerate development? Several factors contribute. First, we are launching more products that cover wider market segments for these international markets. As we mentioned earlier, this year we intend to launch four new products; all four are global products and complement our current market segments, either at the lower end or the high end, and are attractive for international markets. Also, we intend to launch our extended range product line by the end of this year to select markets. That will help in countries where charging infrastructure and driving distances are barriers to EV penetration. We have been stepping up localization efforts. Last year, in key markets in Europe and Southeast Asia, we started local production for some products. Local production will be an important part of our growth strategy by providing more attractive products and better service closer to customers. We also intend to launch our smart driving technologies alongside expected DCAS regulation. By the end of the year or early next year, our autonomous driving VLA 2.0 system can be utilized in international markets such as Europe and select Southeast Asian markets. On top of that, we are developing organizational structure, recruiting talent, building brand and marketing, and starting to roll out proprietary charging facilities in select countries. All those efforts are intended to increase our competitiveness internationally and expand brand awareness. We have an exciting prospect for the next three to four years. We anticipate the overseas market will grow faster than our overall growth rates and become a core profit center for our business.
My second question is regarding our AI investment and overall computing power. You mentioned that physical AI investment will increase to RMB 7 billion this year. What is our compute power right now and the plan for the next few years? Will these investments be recorded in R&D or allocated to capital expenditures?
Thank you for your question. Since this involves long-term investment, I can provide a general response. In the coming years, R&D in the automotive sector will continue to converge and gradually stabilize, while investment in AI will gradually increase and trend upwards because of our conviction in AI development. However, R&D spending will become more efficient. In subsequent years, the growth rate of our AI R&D investment will not be as aggressive as in previous years. Our major investment will include autonomous driving, smart cockpit, robotaxi and humanoid robots across multiple categories. After mass production of humanoid robots begins, we will see greater investment in humanoid robots as well. All of those investments will be booked as R&D expenses and not CapEx. Regarding compute power, it's difficult to pinpoint exact long-term requirements for physical AI. Right now, we have tens of thousands of GPUs. In the future, I think we need to at least reach 100,000 units of computing power accumulation for physical AI data training purposes — at least that number. I believe that robotics training computing power requirements will be orders of magnitude greater than vehicle requirements. In the future, new methodologies will be used to address computing power and infrastructure challenges, including power generation. These will be long-term challenges for physical AI, but we believe we will figure out solutions to address them over time.
Your next question comes from Ming Lee from Bank of America.
What is our current progress on robotaxi testing? What do you expect as important milestones and timing for our robotaxi business with a safety driver and without a safety driver in operation?
Thank you. I think I touched upon this question earlier. We believe that fully autonomous driving capability on the software side will arrive in about one to three years, not only in China but around the world. Our hardware development is progressing smoothly. What remains is addressing regulatory requirements. It will take time to obtain testing licenses and to move from having a safety driver on board to going without a safety driver. We are also doing extra R&D for robotaxi global operation. I believe that by the second half of this year, we should be able to do some tests with a safety driver on board. By the beginning of next year, hopefully, we can operate without a safety driver on board. In the future, we are going to open the whole system to cover not only China but the world, partnering with reliable operational partners globally, using our platform, technology and products so that we can provide robotaxi or autonomous driving capability for users around the world.
What do you expect the scale or cost structure of our robotaxi business to be in the future? Which models will provide the robotaxi service?
The expansion and deployment of robotaxi services depend on regulatory development. In the future, the whole ecosystem needs to develop quickly. What kinds of vehicles will support robotaxi is something that will be explored over the next few years. Fundamentally, we need to determine whether robotaxi should be driven by cars or by robots. Over time, people will see clearer answers to that question.
As there are no further questions, I would like to turn the call back over to the company for any closing remarks.
Thank you once again for joining us today. If you have further questions, please feel free to contact XPeng's Investor Relations through the contact information provided on our website or Piacente Financial Communications.
This concludes today's conference call. You may now disconnect your line. Thank you.
Portions of this transcript that were marked as interpreted were spoken by an interpreter present on the live call.