PONY Pony AI Inc.
$3.99
Pony AI Inc. Q3 F2026 Earnings Call Transcript
AI Conference Call Analysis
Sign in or subscribe to read.Conference Operator
Operator
Thank you for standing by and welcome to Pony AI Inc.'s third quarter 2026 earnings conference call. At this time, all participants are in listen-only mode. After the management's prepared remarks, there will be a question and answer session. As a reminder, today's conference call is being recorded and a webcast replay will be available on the company's investor relations website at ir.pony.ai. I will now turn the call over to your host, George Shao. Head of Capital Markets and Investor Relations at Pony.ai. Please go ahead, George.
George Shao
Head of Capital Markets and Investor Relations
Thank you, Operator. And hello, everyone. We appreciate you joining us today for Pony AI's third quarter 2026 earnings call. Earlier today, we issued a press release with our financial and operating results, which is now available on our
Conference Operator
Operator
Pardon me, everyone.
Conference Operator
Operator
This is the conference operator. It looks like we've lost our speaker connection here, so I'm going to put on some more music, and we will be right back with you. Please stand by while I reconnect.
George Shao
Head of Capital Markets and Investor Relations
Are we back on?
Conference Operator
Operator
And pardon me, everyone. This is the operator. I've rejoined the speaker location. Sir, please proceed.
George Shao
Head of Capital Markets and Investor Relations
Yeah, sorry for the technical issue. I was saying thank you, operator. And we appreciate you joining us today for Tony Ad's third quarter 2026 earnings call. Earlier today, we issued a press release with our financial and operating results which is available on our Investor Relations website. An earnings presentation which we will refer to during this conference call can also be accessed and downloaded on our Investor Relations website. Joining me on the call are Dr. James Peng, Chairman of the Board and the CEO, and Dr. Tiancheng Lou, CTO, and Dr. Liu Wang, CFO of the company. They will provide prepared remarks followed by a Q&A session. Before we begin, please refer to the Safe Harbor Statement in our earnings release, which applies to this call as we will be making forward-looking statements. Please also note that we will discuss non-GAAP measures today, which are more thoroughly explained and reconciled to the most comparable measures reported under GAAP in our earnings release available on our investor relations website. and Fowlings with SEC and Hong Kong Stock Exchange. I will now hand over to our Chairman and CEO, Dr. James Kong. Please go ahead.
Dr. James Peng
Chairman of the Board and Chief Executive Officer
Thank you, George. Hello, everyone. Thank you for joining our Endings call. For the Q3, we actually delivered another amazing quarter. Let me walk you through four highlights for this quarter. total revenue grew 75% year over year with robo taxi revenue more than doubled and reaching 40% of total revenue. Second, our robo taxi fleet now surpassed 2,700 vehicles expanding across both domestic and international markets. This massive scale up allowed us to interconnect Our Guangzhou footprint into a single contiguous coverage area spanning over 1,000 square kilometers. Third, we continue to expand globally with Europe's first fully driverless operations in Zagreb, Croatia, and our entry into key UK markets. Fourth, we achieved strong growth beyond just the passenger vehicles. RoboTruck revenue grew to 76% year over year. We also initiated Gen4 RoboTruck commercial operations and made solid progress on RoboLed trucks. This is our real edge. One core stack can generalize across vehicle platforms and operating environments, turning R&D into multiple revenue streams. Let me first start with our progress in China. As regulation normalizes and entry standards rise, our first mover advantage becomes more valuable. With scaled operation, our safety record keeps on improving, and as a result, insurers are now actually pricing in our lower risk profile. Recently, we renewed our vehicle insurance, and the insurance premium per Gen 7 robot taxi came in at just 40% of those for the average human-driven taxis, down from 50% last year. This is a clear indicator of our superior safety track record. In terms of Robotaxi operation, we continue to deepen our leading position in high-value mobility markets. In Guangzhou, building on the city-wide commercial permit we secured earlier this April, we have taken the lead in providing cross-district Robotaxi services, linking Nansha District with the downtown central area. Our previously separated operating areas in the city's central districts and Nansha now form a single contiguous service area covering more than 1,000 square kilometers. That's actually roughly nine times the size of San Francisco. Operating a fleet at this scale is a clear demonstration of our leading technology and the solid dispatching capability. Through the Pony Pilot app, Users can now book rides across five districts, including Nansha, Haiju, Tianhe, Huangpu, and Panyu, directly lifting our trip distance and average price per ride. In Shenzhen, our service at Bao'an International Airport is already open to users. We are now the only company operating a commercial, fully driverless robot taxi service at an international airport in Chinese tier-one cities. Shenzhen is actually hosting APAC these days, and our robot taxis are from the center in the city's showcase of innovation to the world. We are proud to be a living example of openness, innovation, cooperation, and green mobility that AIPAC Forum and Shenzhen champion to the world. Meanwhile, in Shanghai, we have secured additional commercial permits for fully driverless operations in the core area of Pudong District. Supported by these new licenses and expanded fleet, we will very soon extend our coverage into key economic hubs such as Zangjiang Science City. This is home to numerous high-tech semiconductor and biopharmaceutical companies in Shanghai. Beyond the tier one cities, we are also entering into new high commercial value cities. We have recently initiated deployments in Hangzhou, Chongqing, and Zhuhai. laying the groundwork for further expansion of our operational footprint. We should also see RoboTaxi launches in several other cities soon. Now let's switch the gears to overseas market. We are accelerating penetration into high-value cities, especially roads in Europe and the Middle East. Backed by a proven safety record and a business model, We continue to capture shares in key markets. Under our joint deployment model, we secured mobility partners and started generating high margin revenues. In Zagreb, Croatia, we achieved Europe's first fully driverless passenger rides, going from launch to fully driverless in just five months. This is our joint deployment model in action. Our core technology and operating know-how provide the foundation, while our partners, such as Uber and Vern, facilitate local user acquisition, service operations, and regulatory adaption. Our operations there have accumulated more than 200,000 kilometers and completed thousands of passenger trips. Our success in Croatia has become a replicable template for us to enter other global markets. We continue to implement our previously announced pipeline of over 4,000 robotaxis. In UK, by collaborating with Uber, we are advancing deployment in London and Manchester. The pace keeps on accelerating as road testing is already underway. We are also prepared to launch RoboTaxi in several key markets in Europe. In Dubai, we are accelerating our commercialization progress. Last month, we taxed 2026 in the Dubai World Trade Center. We were the only RoboTaxi partner showcased alongside Dubai's road and transport authority, laying the groundwork for further expansion. some other international efforts, including our current efforts in entering into Central Asia for the first time. Together with InDrive, we will launch RoboTaxi operations in Kazakhstan. Let me now turn to RoboTruck, where we deploy the same underlying technology and the safety framework that we developed for RoboTaxi across more vehicles and use cases. We unveiled a new Gen 4 heavy-duty robot truck model in collaborating with GAC Commercial Vehicle Corporation for the broadening our Gen 4 product lineup. The new vehicle is built on the T9 battery electric heavy-duty truck platform. We are now entering operations at Jiangmen Port and Malan Port, both in Guangdong Province. This is China's largest economy and the biggest export hub. The operation will generate ongoing revenue. And with success in these two ports, we have built a solid deployment pipeline with various leading port operators globally. Our target is to reach hundreds of vehicles across ports next year. In addition, we're tapping into urban logistics with our light duty trucks, which recently entered production. We're also advancing commercial deployments across key logistics corridors in the tier one cities in China. At the beginning of this year, we have set three targets for the year. First, we'll deliver 3.5 times Robotaxi revenue over prior year level. Second, our fleet is scaling to 3,500 vehicles. And third, our global Robotaxi deployment will cover 20 cities worldwide. As we approach year end, we are on track to reach all three ambitious targets that we planned earlier.
Dr. Liu Wang
Chief Financial Officer
Looking ahead to 2027, we will scale up our Robotaxi and Robotruck businesses with full momentum.
Dr. James Peng
Chairman of the Board and Chief Executive Officer
Robotaxi will expand across three fronts, deepening operations in existing cities, ramping up in new Chinese cities, and accelerating overseas in strategic markets. Our dual-engine domestic and global growth strategy will increase commercial vehicles in operation, directly driving top-line growth. GlobalTruck will deliver another powerful engine as Gen4 models and new initiatives begin unlocking new revenue streams. This is where our flywheel becomes truly powerful. A greater scale yields superior technology, higher efficiency, and stronger economics. With that, I'll turn the call over to our CTO, Tiancheng, to elaborate how our technology powers this flywheel. Tiancheng, please go ahead.
Dr. Tiancheng Lou
Chief Technology Officer
Thank you, James. So hello, everyone. This is Tiancheng. I will elaborate more on technology. Let me start with where we are. Back in early 2025, we were already running over 100 fully-drived robot taxis and continued to extend. Since then, we scaled that operation more than tenfold. The real story is how we scale. Alpha robot taxis and robot trucks operate in the physical world, where safety is non-negotiable. Every second, they interact with vehicles and pedestrians in real time. When a decision goes wrong, the damage is done. There is no rollback, no retry. That fundamentally separates us from digital AI applications, where models can fail and retry endlessly, and from L2-assisted driving, where a human driver remains behind the wheel as the ultimate safety format. Like frontier large-language models, physical AI has also two stages of training, pre-training and post-training. Pre-training on massive multimodal data established basic driving skills, but its performance ceiling remained bounded by human driving data, leaving long-tail hallucinations unresolved. Breaking beyond the human safety ceiling requires post-training by reinforced learning. In the latest AI engineering, over 90% to 95% of core work happens here. This is where our world model comes in. The whole new world is our closed-loop virtual driving school. Inside it, our AI driver executes billions of reinforced learnings and counterfactual scenarios to eliminate transformation and solve long-tail safety problems. However, physical AI faces a critical challenge that digital AI, like coding, does not. The same-to-real gap. In coding, the environment is domestic. for coding, the training environment and the real-world environment are the same one. But in autonomous driving, over-world model must mirror exact probability distribution of the physical world. Education cannot appear too rarely or the model won't learn how to handle it. Nothing cannot be artificially exaggerated or the driving policy become too overly conservative. This simulated distribution have to mirror reality. This is our decisive edge. We built an AI-driven, self-evolving cold-suit pipeline that continues to calibrate and upgrade Pony World using real-world fleet feedback. Every edge case encountered on a public road automatically sharpens the world model's fidelity. As a result, Pony World not only evolves at a fast speed, but breaks through a fundamental highest performance ceiling because it overcomes the limits of human comprehension and manual scenario design. Today, Pony World generates 10 billion kilometers of training every week, scaling post-training in the virtual world before deploying onto public roads. Because Pony World dominates post-training, it solves complex long-tail interactions with unmatched efficiency. directly empowering over commercial extension across four forms. First, we get through safety. Real-world risks come from rare extreme interactions by using intentional illumination inside the Pony World. We systematically generate all possible intentions and the trajectories of surrounding real users in a given scenario. It allows us to exhaustively stress test every extreme interaction path in simulation and regularly verify the continuous and safety boundary of our driving policy before deployment. This drives measurable safety improvement, reflecting our insurance premium dropped to 40% of those human-driven taxis compared with 50% last year. Second, rapid area extension without retraining Inside Pony World, reinforced learning trains our AI driver to master all kinds of generalizable scenarios inside a specific road or area. Whenever traffic rules and driving styles remain consistent, the model generalizes natively with zero retraining. Because over-safety is proven at scenario level, expanding our fully driverless coverage to over 1,000 square kilometers in Guangzhou and scaling to tens of thousands square kilometers next is purely a function of deploying more vehicles, not repeating data collection or R&D cycles. Third, scalable global replication. Enter new countries require adopting two distinct traffic laws, left and right-hand driving and the local driving patterns. Holding world high fidelity interactive smart agents accurately replicate local user behavior, given our AI driver localized interaction training prior to launch. This directly empowers Europeans' first fully driverless operation in Zagreb in just a few months and gives us a path to replicate internationally at scale. The fourth, cross-platform generalization. RoboTaxi and RoboTrucks share one unified core architecture. By sensing agnostic generalization, our model extracts a unified physical representation across diverse vehicle types and sensor setups, effectively adopting biopulse training to heavy-duty trucks and light trucks. One IMD engine unblocks multiple revenue streams with minimal incremental cost. So finally, unlike Unlike other digital large-language models that expose public APIs and can be easily distilled, our models operate strictly inside our closed-source system. Anyone who wants to replicate it must build an entire post-training infrastructure and the real-world data firewall from scratch. This gives us a much deeper mode than large-language models, making it far harder for newcomers to catch up. Looking forward, this technical edge will keep compounding. As more real-world operating data flows back into the polling world, our simulation fidelity and training efficiency will continue to advance. That means when we expand into larger operating areas, enter more overseas cities, and extend into more vehicle platforms, the marginal R&D cost will keep coming down. The speed of commercialization will keep growing up. So with that, we hand the call over to our CFO, Dr. Leo Wang, to review our financial results. Leo, please go ahead.
Dr. Liu Wang
Chief Financial Officer
Thank you, Tiancheng. Hello, everyone. This is Leo. I will now review our financial results for the third quarter of 2026. As a reminder, all comparisons are on a year-over-year basis, unless otherwise stated. Please refer to our earnings release for detailed financial results and the reconciliation of US GAAP and non-GAAP results. Total revenues increased by 75% to $44.6 million in the third quarter of 2016. Together, these two business lines accounted for 80% of total revenues compared with 66% a year ago. As our business continued to scale, revenues grew much faster than operating expenses, reflecting improved operating leverage. Our operating margin improved from negative 274% to negative 154%. Let me start with Robotaxi. Robotaxi's services revenue maintained a strong growth momentum reaching 17.8 million US dollars, up 165% year-over-year, and 47% quarter-over-quarter. This made Robotech the largest contributor to our year-over-year revenue increase. Robotech revenues accounted for 40% of total revenue, up from 26% in the prior year period, Along with this revenue growth, our fleet continues to expand. As of September 30, 2026, our global robot taxi fleet exceeded 2,700 vehicles. Just as a quick recap, our robot taxi business operates through two primary models, our self-owned fleet and the joint deployment model. In China, we primarily generate fare charging revenues from our self-owned fleet. Overseas, we strategically partner with local fleet operators and mobility platforms under the joint deployment model, generating revenue from vehicle sales as well as virtual driver licensing fees. Drilling down for fare charging revenues, we recorded approximately 400% year-over-year growth in the third quarter supported by the continued expansion of operating areas and the service offerings. In Guangzhou, operations in Haiju, Tianhe, Huangpu, and Panyu District, which we highlighted in our previous earnings release, generated incremental revenue. In Shenzhen, our services to and from Bao'an International Airport and the Sherco Cruise Port also contributed to third quarter fare charging revenues. Under the joint deployment model, we will continue to advance deployment in market with higher fares while accelerating the rollout of our previous announced pipeline of more than 4,000 robot taxis with Uber and other partners.
Conference Operator
Operator
As commercialization Pardon me, everyone.
Conference Operator
Operator
This is the operator. It does appear that we have lost the speaker connection again. I'm going to put the music back on and I will dial back out. Please stand by, everyone. Once again, we'll get started in just a moment. Thank you. Thanks for holding everyone. This is the operator. I've reconnected the main speaker line. Please proceed.
Dr. Liu Wang
Chief Financial Officer
Okay, yeah, this is Leo again. Let's continue. Now let's shift into robot truck services. Third quarter revenues reached 18 million U.S. dollars, up 76% year over year and 35% quarter over quarter. The increase was primarily attributable to the growth in freight transportation services, supported by our collaboration with Sinotrans. In addition to the growth in our existing freight business, we are advancing robot truck deployments in Jiangmen and at Marwan port in Shenzhen, with commercial operations expected to generate ongoing transportation services revenue. As we expand into more ports, our project pipeline will continue to fuel top line growth. Lastly, Intelligent Solutions revenues were 8.9 million US dollars, broadly flat compared with the third quarter of 2025. Let me now turn to gross profit and operating expenses. Gross profit increased by 69% to 7.9 million US dollars as revenue grew. Gross margin was 17.8%, a slight improvement from 17.5% in the second quarter of 2026. On the cost side, our safety record has helped us secure more favorable insurance terms. Annual insurance premiums for our Gen 7 robot taxis were down to approximately 40% of those for human-driven taxis. compared with approximately 50% in 2025. While higher vehicle costs would normally push up insurance premiums because of higher repair costs, our safety record has more than offset this upward pressure. This translates into structural savings in fleet operating costs, strengthening our unique economics as we scale. We also maintained Expenses discipline. Gap operating expenses increased by only 3% year over year to 76.6 million US dollars. Non-gap operating expenses, which excluded share-based compensation expenses, were down 2% to 66.2 million US dollars. Meanwhile, the increase in gross profit more than offset the modest increased in operating expenses, helping narrowing our gap loss from operation to 68.7 million US dollars, compared with 69.7 million US dollars in the same period last year. The improvement was more pronounced on a non-gap basis, with non-gap loss from operation decreased by 8% year over year 58.3 million US dollars, reflecting stronger financial discipline. Gap net loss was 136.3 million US dollars, including a non-cash loss of 77 million US dollars from changes in the fair value of trading securities. Excluding this fair value loss and the share-based compensation expenses, non-gap net loss narrowed by 14% to $48.9 million from $56.8 million a year ago. Let me finish with our financial position and our capital allocation. We maintain substantial financial resources to support technology development and the commercial expansion. As of September 30, 2026, Cash and cash equivalent, restricted cash, short-term and long-term debt instrument for wealth management totaled approximately 1.15 billion U.S. dollars, compared with 1.21 billion U.S. dollars as of June 30, 2026, on the same basis. Capital expenditures were 17.7 million U.S. dollars, a decrease of 22, and Zeng Li. The decrease primarily reflected a disciplined pace of investment while we continue to support the mass production and the deployment of Gen 7 Robotaxis fleet. This quarter, revenue growth and improved operating performance reflect the progress of our commercialization efforts. Looking ahead to the fourth quarter, and 2027, we will continue directing our resource into core technology and key markets, translating our active pipeline and commercial progress into sustained revenue growth. At the same time, we maintain expense and capital discipline to deliver greater operating leverage as our revenue base expands. With that, I will now turn the call over to the operator for the Q&A session. Thank you.
Conference Operator
Operator
Thank you. We will now begin the question and answer session. To ask a question, please press star then one on your telephone keypad. To remove yourself from queue, please press star then two. For the benefit of all participants on today's call, please limit yourself to one question. If you have more questions, please enter the question queue. If you ask questions in Chinese, please repeat them in English. and today's first question comes from Tim Sao with Morgan Stanley.
Conference Operator
Operator
Please go ahead.
Dr. James Peng
Chairman of the Board and Chief Executive Officer
Hi, this is Tim from Morgan Stanley. Thanks for taking my questions and congratulations on the solid operating results. I just have one question. It's about overseas expansion.
Dr. Liu Wang
Chief Financial Officer
How does Pawnee stand out in overseas markets such as London? Despite the tough competition of multiple players,
Dr. James Peng
Chairman of the Board and Chief Executive Officer
are already in the area. So that's my question. Thank you. Thanks for the question. This is James. Let me take this one. Yes, there are a lot of competitors in the European market, but I feel confident about our overseas market position, especially in Europe. Globally, when we look at it outside of China and the U.S., Europe is one of the most attractive robo taxi markets, giving scale, pricing, and regulatory maturity. In that market, our competitive edge actually comes down to our proven technology, a strong safety system, and especially our very competitive unit economics. Because of our competitive edge, the leading global mobility players want to work with us which in turn can facilitate the regulatory approval and allows us to scale faster, attract more partners and create more values across the whole ecosystem. Let's look at Croatia as a great example. We launched there in April this year and by September we had already gone fully driverless. In just five months, we achieved Europe's first fully driverless passenger operations in more than 200,000 kilometers and thousands of passenger trips. Also, in Vagra, Croatia, that is a classic European cities, our robo-taxis actually operate in a mixture of scenarios with old town roads modern urban districts and the complex traffic conditions. So this successful experience definitely can replicate to other European cities. Now we are actually taking this playbook to the UK and other European countries. As Uber's partner, we are advancing deployments in London and Manchester to the two of UK's most premium mobility markets. We actually started road testing in Manchester shortly after our announcement, which also shows that our execution speed is great. So with our proven safety record, large-scale operation experience, attractive unit economics, and a strong partner network, we believe we are were all positioned to be a leading player in Europe and especially in the UK. So I'm sure that we will capture a leading market share as the market matures. So with that, back to the operator.
Conference Operator
Operator
Thank you. Our next question today comes from Ming Sun Li with B of A.
Conference Operator
Operator
Please go ahead.
Dr. Tiancheng Lou
Chief Technology Officer
Hi, James and our management team. Thank you for giving the opportunity to answer the question. So I also only have one question. So could you share some cover on the operational and the regulatory progress in China?
Dr. James Peng
Chairman of the Board and Chief Executive Officer
This is James again. I'd like to take this question again. So I believe that the current regulatory environment is by and large is actually turning into a positive direction. Under the 15th five-year plan, intelligent connected new energy vehicles are becoming part of the national policy agenda, targeting large-scale commercial deployment of autonomous driving for this five-year plan. This provides ongoing policy momentum as a central government to achieve the milestone. essentially have a regulatory base to shift RoboTaxi from local implementation towards a more standardized national framework. As the framework becomes more standardized, the safety bar certainly will be uniform and high for all the cities across the whole nation. This is because all the other nations, other cities that want to operate RoboTaxis They need to follow the practice first established in the tier one markets. That will actually directly benefit operators like us who have already established proven technology, strong safety records, and real operating experience. We have already built the track record in China's most complex tier one cities, and we are now scaling our active fleet expanding coverage in the existing core cities and entering new regions such as Hangzhou, Chongqing, and Zhuhai with many other cities in the pipeline. So therefore, my opinion is that as the industry enters a more standardized expansion phase, we believe Tony's first mover advantage becomes even stronger. This track record and regulatory trust help streamline the regulatory approvals, accelerate our engagement with both the central and the municipal authorities, and ultimately allow us to enter new markets faster and consolidate a leading position over time. With this, back to the operator.
Dr. Tiancheng Lou
Chief Technology Officer
Thank you.
Conference Operator
Operator
Our next question today comes from Yuquan Ding with HSBC.
Conference Operator
Operator
Please go ahead.
Yuquan Ding
Analyst, HSBC
Thank you. Thank you, Tim. I got two questions. The first, could you share with us on the update on the RoboTaxi unit economy? And second question, then, how should we think about the UE margin domestic versus overseas market as you are scaling up in both places? Thank you.
Dr. Liu Wang
Chief Financial Officer
Thank you for your question. This is Leo. I'll take this one. So I think, yes, UE is really, really important. And let me start with our China operation. So we have already achieved a positive UE margin in our Guangzhou and Shenzhen fleet. And with the network effect and also our continued effort on cost optimization, The trend is actually keeping improving. One direct example, as we mentioned in this earning call, is about the insurance. That per vehicle insurance premium for our Gen 7 fleet has come down by another 19% this year. So I want to also emphasize when we talk about UE, the revenue refers to the net value generated from ride-hailing services after deducting discounts and refunds. And also we cover seven cost items. First is the depreciation of the vehicle, including the battery and also the autonomy driving hardware. Second is the electricity charging. Third is the salary we pay to the remote assistant and also ground supporting staff. Fourth is the vehicle cleaning service and maintenance. Fifth is the insurance cost. And the sixth is the parking cost. And the final is the network communication cost. About half of the cost comes from the vehicle depreciation. And the other remaining half comes from the other six cost items. So when we talk about the UE, it's really a very comprehensive measure. and we're very proud that the Gen7 platform gives us a very meaningful cost advantage because it is fully automotive grade design. It is also helping us to lower our EV charging cost and also helping us to lower the service and maintenance cost. In addition, our operational efficiency we believe remains to be industry leading backed by very competitive and the staff-to-vehicle ratios for both remote assistant and the ground support staff. We will continue to improve our staff-to-vehicle ratio and this will keep the labor cost per vehicle low even when the fleet keeps scaling. And our overseas business carry a different structure which is higher UE margin driven by two factors. First, on the cost side, thanks to our stronger technology and optimized operation at a scale, our total cost per mile remains very, very competitive in the whole industry. Specifically, our very high ratio of staff to vehicles ratio and also our superior energy efficiency gives a very good cost edges in some developed market like Europe where the labor and electricity costs are very high. Second is the joint deployment model, which is actually asset-light for us. We don't bear the vehicle CapEx. We offload the depot infrastructure, EV charging and the daily vehicle cleaning and maintenance and other operational costs to our local partners, which would further improve our margin profile. We think that the joint deployment model will become a high margin business and that that can also scale. With that, I'll get back to the operator.
Conference Operator
Operator
Thank you. And our next question today comes from Lawrence Ren at Citi.
Conference Operator
Operator
Please go ahead.
Dr. James Peng
Chairman of the Board and Chief Executive Officer
Hi, management. Thank you for taking my question.
George Shao
Head of Capital Markets and Investor Relations
And this is Lawrence from Citi. I only have one question on the AI side.
Dr. James Peng
Chairman of the Board and Chief Executive Officer
As everyone in the AI industry is buying chips, with the recent news about the S-200 restrictions and the risk of more bans, where is your compute stand going and how are you different from the rest of the industry to stay safe from these bans?
Dr. Tiancheng Lou
Chief Technology Officer
Thanks. Thank you. This is Tiancheng. I can take this one. In short, I will say we are not bottlenecked by the AI compute. So let me break this down across the inference and training part. So on inference, so over Gen 7, Robotext had been running on automotive GUI chips since it launched last April. So these chips require much less computing power than general-purpose GPUs in data centers. We optimize the computer resource consumption of our own vehicle models by optimizing model parameter size and improving the efficiency of our inference engine. So this delivers low bump cost, lower energy consumption, and increased driving mileage per full charge. That's for inference. And on training, over-computer dynamics are dramatically different from large-language models for two reasons. First, our on-vehicle models are much smaller than large-language models, which means we don't need a huge cluster of cutting-edge GPUs to do the pre-training or post-training. Most post-training computes go in towards running simulation inside our world model, meaning in training, the driving model primarily relies on world model inference, which requires far less compute. Second, Over the past few years, we will keep making our world model iterate faster. We call it Pony World 2.0. It's a step for evolving. It takes the raw test data coming back from over L4 fleet, works out where the root-synth view gap is, and figures out where scenarios need to be better fidelity automatically. That cuts both the time and the compute we burn on training the world model by an order of magnitude. So in summary, over-on-vehicle model architecture is more refined, and over-engineering execution is stronger. That is what makes us the first company in the world to achieve L4 driving on automotive grid chips. Our per-vehicle bond cost is highly competitive. The same strengths also make our data center compute more economical. So as I mentioned earlier, Automatic driving models fundamentally cannot be distilled, so the competitor has to build the entire system from scratch. They also see that all the major L4 players today rely highly on vehicle compute. That means larger model parameter size and more compute consumption in training. For an AI company, both model and engineering capability matters. I believe our technology lead will drive faster fleet extension. It also gives us more flexibility when data center computers in short supply. In addition, we have evaluated more than one China-made computer chips for model training, and we have successfully deployed our training pipeline. If needed, we can train our models on those chips. With that, back to the operator.
Conference Operator
Operator
Thank you. And our next question today comes from Leo Yu at CLSA. Please go ahead.
George Shao
Head of Capital Markets and Investor Relations
Hi. Good evening, management.
Dr. Liu Wang
Chief Financial Officer
Congratulations on the strong results in both Robotaxi and beyond Robotaxi. So my question is, how do you think about diversifying the whole business? And it seems that you are moving to more logistics fronts.
Dr. Tiancheng Lou
Chief Technology Officer
And would you mind sharing more about that strategy? Thank you.
Dr. James Peng
Chairman of the Board and Chief Executive Officer
This is James, and I'm glad to take this one. As most of you know, since our founding, our vision has always been autonomous mobility everywhere. And the logistics market sits at the center of our vision. Logistics itself is a massive market, and autonomous driving is a great solution for lowering down the costs and also feeling the labor shortage. We leverage the same L4 technology and the operation across vehicle platforms and the environments. We use exactly the same robot taxi stack that we have developed over the years to power the robot trucks and also our robot light trucks. Of course, several scenarios such as ports, long-haul corridors, and intercity logistics. We actually have already seen the results. Robotruck revenue grew 76% year-over-year this quarter. By collaborating with Sinotrans, we are deepening our long-haul operations, such as in the Beijing-Tianjin-Tangshan corridor. In ports, our Gen4, are in operation at Jiangmen and also Mawen ports with an ongoing pipeline of hundreds of vehicles across the leading global port operators. In the urban logistics, we provide our solutions as last-duty robo trucks compared with the more popular, currently operating low-speed robo vans Our light truck actually offers three to four times the cargo capacity and two to two times the speed. So essentially it's a great application for much higher valued applications such as supermarkets, cold chain logistics, and also the full urban supply chain. So as a result, the potential for commercialization for long haul trucks and light duty trucks is actually enormous. That's the power of our generalized platform. We can replicate, deploy, and monetize the same alpha capability across many different markets, which is exactly how we diversify. Of course, all of those efforts actually fit into our vision of the autonomous mobility everywhere. With this, I'll back to the operator.
Conference Operator
Operator
Thank you.
Conference Operator
Operator
And our next question comes from Robin Leong at Bloomberg Intelligence. Please go ahead.
George Shao
Head of Capital Markets and Investor Relations
Hi, thanks management for checking my question. Robin here from Bloomberg Intelligence.
Dr. James Peng
Chairman of the Board and Chief Executive Officer
I have one quick question on the technology side. How do you think about companies plugging GPC-6 type models into embodied AI and eventually coming after the Robotexy market? And if autonomous driving algorithms starts to converge and look more like a commodity, how doable is this mode and where does the advantage get? Thank you.
Dr. Tiancheng Lou
Chief Technology Officer
Thank you for your question. This is Tiancheng. I will take this one. So in embodied AI, industry solos still have not reached large-scale commercial deployment. So that's why a gender-purpose foundation model like GPT-6 can match it in selective scenarios. Massive pre-training gives a broad world knowledge and enables impressive demos, such as robotic interactions and manipulations. However, as I noted in my remarks, Pre-training only establish a baseline of human-like common sense. It does not eliminate long-tail hallucinations across the non-industrial city bar of the physical world. In L4 driving, over 90% to 95% of the core engineering happens in post-training. Operating on public roads means interacting with vehicles and pedestrians in real time, where there's no robot and no retry. That's why a few minutes demo proves nothing about opportunity safety at scale. Furthermore, training L4 autonomous driving directly on top of a large-language model is fundamentally inefficient. Given the same computer budget, language is a slow and a lossy intermediate presentation for the physical control. Even when humans drive, we rely on intuition and understanding rather than translating every perception action into words. This is also why L4 algorithms will not simply converge and commoditize. Reaching commercial scale requires mastering the post-training program and closing the same-to-real gap. Pony World provides a high-fidelity, closed-loop, reinforced learning environment to scale post-training in simulation. while our fleet of driverless vehicles continuously feed real-world data back to calibrate and sharpen point-of-worth accuracy. So in short, foundation model pre-training is only the starting line. Unlike digital large-language models with public APIs, our coastal post-training ecosystem cannot be distilled by other parties. Thank you. And that concludes our question and answer session. I'd like to turn the conference back over to your host for closing remarks.
George Shao
Head of Capital Markets and Investor Relations
Okay, thank you once again for joining us on the call today. If you guys have any questions, feel free to reach out to our IR team. Look forward to speaking with you in the next quarter.
Conference Operator
Operator
Thank you. That concludes today's conference call. Thank you all for attending today's presentation. You may now disconnect your lines and have a wonderful day.