Dialogue with Pony AI's Wang Haojun: Robotaxi is entering the stage from 1 to 1000

Author | Zhou Zhiyu

Editor | Zhang Xiaoling

In 2025, the global intelligent driving industry is undergoing a paradigm shift. Over the past decade, autonomous driving has been a code game in laboratories, a dream built on demos and PowerPoint presentations; now, this business has officially fallen from the void into reality, starting to clash head-on on financial statements.

As the once-glamorous L4 unicorns find themselves stalled due to their inability to cross the scale survival line, pioneers have quietly knocked on the door of profitability. In the second quarter of 2025, Baidu's Apollo Go achieved breakeven in Wuhan; in November, Pony AI announced that its seventh-generation Robotaxi achieved a positive unit economic model (UE) in Guangzhou.

Wang Haojun, co-founder and CFO of Pony AI, stated in a recent interview with Wall Street Insights that achieving a positive UE in Guangzhou means that Pony AI has gradually refined a standard operating process during the scaling process, which can empower its partners.

Wang believes that the commercialization of Robotaxi in previous years was still largely in the 0 to 1 stage, and it has now gradually entered the 1 to 100 and 1 to 1000 stages.

A clear commercialization timeline has emerged: aiming for a fleet of 1,000 vehicles by the end of 2025, increasing to 3,000 vehicles in 2026, and reaching a scale of 100,000 vehicles by 2030, Robotaxi will become a part of people's daily lives.

This means that the competitive main battlefield for Robotaxi has shifted. As the hardware cost per vehicle drops to the survival line of 250,000 RMB, a new competition for profits has begun, while AI is irreversibly reshaping the operational rules of the physical world.

Business Closed Loop

Achieving a positive unit economic model (UE) in Guangzhou means that Pony AI can realize a business closed loop in the regional market. This is a thrilling leap from "laboratory research and development" to "business closed loop."

The so-called UE positive means that after excluding total R&D investment, the income from each vehicle running on the street can cover hardware depreciation and operating expenses.

In the past, Robotaxi was a "burning money furnace" that gave investors headaches, with each test vehicle on the street burdened by exorbitant transformation costs and operating losses. This logic of "the more you invest, the more you lose" was an enduring shadow over the industry.

However, the numbers in Wang Haojun's ledger indicate that the turning point has arrived. Pony AI's seventh-generation Robotaxi achieved an average of 23 orders per day in Guangzhou, with daily revenue of approximately 299 RMB.

According to industry-standard operations, when the average daily order volume reaches 24, a positive cycle can form on both the user and revenue sides, and the figure of 299 RMB is already sufficient to cover all expenses, including hardware depreciation and operating costs.

The ability to balance the books relies on cost control and continuous optimization of operational capabilities.

Compared to the sixth generation, the BOM cost of the autonomous driving suite for Pony's seventh-generation vehicle has dropped by 70%. Pony has completely abandoned the traditional Robotaxi's industrial control computers (IPC), which often cost tens of thousands of RMB and consumed significant power, instead opting for its self-developed domain controller based on four NVIDIA Orin X automotive-grade SoC chips, becoming the industry's first L4 player to achieve large-scale implementation of such a solution By leveraging the scale effect of domestically produced L2+ mass-production vehicles, Pony AI has reduced the cost of LiDAR by 68%. The seventh-generation vehicle no longer relies on expensive customized mechanical radars but instead uses automotive-grade solid-state radars like Hesai AT128, which have entered large-scale production. The vehicle is equipped with 34 sensors that achieve 100% automotive-grade components.

At the same time, there is continuous optimization of algorithms. Wang Haojun revealed that software-level optimizations have improved the system's processing capability by 30 times when dealing with sensor "noise." This means that Pony AI can achieve safer and smoother performance than the previous generation using cheaper and more standardized hardware.

In addition to hardware, every penny of operating costs must be meticulously calculated. Wang Haojun told Wall Street Insight that, thanks to a safety record far exceeding that of human drivers, the annual commercial insurance premium for its Robotaxi is 50% lower than that of traditional taxis, which essentially represents an endorsement from insurance companies for the safety of AI drivers.

On the remote assistance side, the efficiency of human resources is also improving: the current ratio of remote assistants to vehicles has reached 1:20, with plans to push it to 1:30 by the end of the year. Ground maintenance has been standardized through a digital work order system, with cleaning and maintenance of vehicles after returning to the yard handled in bulk by support staff, allowing one person to manage nearly 20 vehicles.

It is this series of cost reductions that has enabled Pony AI to achieve positive unit economics in Guangzhou. Wang Haojun stated that this also gives Pony AI the confidence to deploy more vehicles, further driving growth in financial reports through network effects.

Wang Haojun believes that from a global market perspective, the Middle East market is currently the most noteworthy. As this market enters the demonstration operation phase, it will become a market that generates sustainable revenue.

While running the model, Pony AI is also shifting from heavy assets to light assets. Wang Haojun candidly stated that the company cannot stubbornly pursue self-operation, as the capital expenditure (CAPEX) required to hold 100,000 vehicles would be "financial suicide" for any startup.

Thus, Pony AI has built a value chain of "shared benefits": manufacturers produce vehicles, asset companies like Xihu Group and Yangguang Travel purchase and hold them, traffic platforms like Amap and Ruqi are responsible for order distribution, while Pony AI outputs the core "AI driver" brain.

In this model, Pony AI's revenue has shifted to sales revenue, technology licensing fees, and service sharing for each order. This marks Pony AI's official transition from a heavy asset explorer to an ecosystem enabler, with a goal of expanding its fleet to 3,000 vehicles by 2026.

Industry Turmoil

The Robotaxi track in 2025 has become a chaotic industry battle involving tech giants, mobility platforms, and manufacturers. At this trillion-dollar table, everyone is vying for a ticket to the finals through alliances, dimensionality reduction attacks, and ecosystem restructuring.

Waymo represents the American-style "engineering perfectionism" stronghold. The new round of $15 billion fundraising plans has pushed this pioneer’s valuation toward the $100 billion mark.

Waymo pursues extreme redundancy and standardized solutions, with weekly orders exceeding 450,000 and over 14 million rides since 2025, proving the feasibility of the "long-term cash flow" narrative to the capital market, but this model also comes with a heavy financial burden Wang Haojun pointed out that the new models to be launched by Waymo still cost more than 4 to 5 times that of similar models in China. This heavy asset approach, while building technological barriers, also leaves a window for Chinese players with strong supply chain resilience to launch cost-effective attacks.

Tesla's Robotaxi is also accelerating its rollout. Musk officially removed the safety driver in Austin, attempting to completely reshape the dispatch rules of the physical world through a very low-cost pure vision solution and an ambition of scaling to 1 million units. The reason Tesla dares to enter the market is essentially a shift towards reinforcement learning and generative data within its cloud training framework.

Even Uber, which once withdrew from developing its own brain, has returned to the role of a "neural network" dispatch platform, establishing collaborations with Waymo, WeRide, Pony AI, and others. This "OEM + technology provider + platform" trinity collaboration network is forming a siege against isolated technology companies.

The "hunting" by new domestic entrants is becoming increasingly fierce. XPeng announced that it will mass-produce three models of pure vision Robotaxi without LiDAR by 2026 and will start operations in cities like Guangzhou using second-generation VLA+VLM technology. Hello has partnered with Dongfeng Qichen and Horizon Robotics, aiming to deploy 50,000 units by 2027.

Meanwhile, Baidu's Apollo Go has exceeded 250,000 orders, with cumulative service orders surpassing 17 million.

Players like Cao Cao Mobility and Xiangdao Mobility are also quickly entering the core areas of first-tier cities through deep cross-industry collaborations with Qianli Technology and Momenta.

This round of chaotic competition marks a complete transformation of industry logic: Robotaxi has evolved from a mere "black technology" into a comprehensive melee of capital, supply chains, and operational efficiency.

This also reflects the automotive industry's entry into the second half of intelligent transformation. UBS Investment Bank's head of China automotive industry research, Gong Min, analyzed that the current price of Robotaxi can be compressed to below 300,000 yuan, leading companies to achieve scales of thousands of units and gradually begin to turn a profit on a single vehicle basis in certain areas.

The new technologies arising from industrialization and excellent cost control are gradually becoming affordable for everyone.

If technology companies cannot quickly achieve deep bundling with OEMs and platforms, they will slowly wither away in the dual depletion of data and funds. Ultimately, this is no longer a competition among engineers over whose algorithm is more elegant, but a power struggle over who can first occupy urban physical space and achieve real operational dispatch at a scale of millions.

System Competition

No one wants to miss out on the trillion-dollar market of Robotaxi. Especially when technology reaches a critical point, there will be rapid development, even a swift popularization overnight.

Gong Min analyzed for Wall Street Insight that by 2030, if partially deployed in first-tier cities in China, the Robotaxi market size will be $8 billion; if deployed nationwide in China, it could reach $183 billion; and if overseas markets outside the U.S. are included, it could reach $394 billion.

Once technology crosses the feasibility threshold, algorithms are no longer the only trump card. Robotaxi is undeniably entering the "second half" centered on operational efficiency Why has operations become so important? UBS's research report predicts that as hardware costs decline, operating expenses (maintenance, insurance, and energy replenishment) will increase from about 48% of the cost structure per vehicle to 55%. This means that the true competitive dominance in the future is shifting from engineers who write code to builders who understand urban scheduling best.

The "golden node" for economies of scale is set at 100,000 units. Wang Haojun predicts that by 2030, Pony AI's goal is to reach a scale of 100,000 vehicles, which corresponds to a 5% to 10% share of the mobility market in China's first-tier cities. Once this threshold is reached, powerful network effects will lead to qualitative changes, whether in user acceptance or in the continued growth of scale.

To prepare for this moment, leading players are resolutely abandoning the dead end of "imitative learning." Wang Haojun emphasizes that the mission of L4 is to be ten times safer than humans, which must rely on reinforcement learning and generative "world models." This algorithm is not only for driving well but also for training a general AI brain with a chain of thought (COT) and logical reasoning capabilities.

From a broader narrative perspective, the battle being fought by autonomous driving companies is actually a preliminary skirmish for AGI (Artificial General Intelligence). Today's smart cars are essentially robots with four wheels. Horizon's "intuitive system" aims to give AI human-like intuition, while Yuanrong Qixing's VLA model endows machines with the ability to understand road signs and reason through complex tidal lanes.

They all point to the same ultimate goal: to enable AI to perceive and interact through a deep understanding of physical rules rather than mere instruction execution, much like biological entities.

Robotaxi is the largest, most real-time, and least fault-tolerant practical training ground for AGI in the real world. The logical reasoning capabilities and game-theoretic algorithms refined in this "meat grinder" will ultimately be reused in broader embodied intelligence fields. Securing the high ground of Robotaxi means obtaining the ultimate ticket to reshape the intelligent rules of the physical world.

In this brutal contest concerning survival sovereignty, those who hold the key to breaking the deadlock will inevitably be the ecological builders who can deeply understand the industry's depth while possessing extreme operational efficiency. The transformation has already begun.

The following is a transcript of the dialogue between Wall Street Insights and Wang Haojun (edited):

Q: Why does Pony AI choose to implement a "light asset model" at this time? How is the revenue sharing calculated?

Wang Haojun: This actually goes back to the difference between L4 and L2+. When you have an L2+ product, you provide partial driving functions, and the value of that function varies among different users. However, for L4, I provide a complete driving function, meaning I take care of the driving task.

We expect that in the L4 field, whether through licenses or revenue sharing, this recurring part is very important. And I believe this will definitely exist.

As the current UE becomes formalized, more companies will be willing to engage in this. One point we focus on in this process is whether we can deploy more vehicles with less capital through a more effective capital operation method We also realize that many traditional operating companies are asset-heavy. For them, this has historically been the case. This new business model is not a disruption to them, but rather a good extension.

As for revenue sharing, we are still in the early exploration stage. From the perspective of Pony AI's economic interests, it cannot be separated from three aspects: vehicle sales revenue, technology licensing fees, and service revenue sharing.

Q: There is a lot of discussion in the industry about the boundaries between L3 and L4. What do you think the future evolution trend will be?

Wang Haojun: L3 is difficult to provide services like L4. In the end, L3 still requires a human driver. As long as there is a human driver present, the labor cost for running Robotaxi cannot be eliminated, and the unit economics (UE) cannot turn positive.

L3 itself does provide a pathway for private cars to advance, reflecting the progress of intelligence. If L3 is successful, then in the future, based on the success of Robotaxi, more people may be willing to accept private cars that include L4 functions.

Q: Will Pony AI develop L4 aimed at the consumer market in the future?

Wang Haojun: Commercially, it can never be said to be impossible, but at present, it is still too early for us.

Q: What is Pony AI's logic for overseas market layout? Which regions are the focus?

Wang Haojun: The capacity for Robotaxi overseas and domestically is still at two different levels. I expect that next year, the total number of vehicles allowed for demonstration operations overseas will be a few hundred.

The overseas market is currently in the early expansion stage and has not yet reached the stage of true commercialization. At least a scale of a thousand vehicles is needed for commercialization opportunities. Robotaxi is a heavily regulated industry, and it is necessary to accumulate mileage locally. Therefore, the layout overseas needs to be done in advance, focusing on establishing safety records.

In terms of regions, China and the United States are the largest markets, followed by the European Union. Additionally, countries with high labor costs and high travel demand, such as Japan, South Korea, and Australia, are also prioritized.

The Middle East market is very unique. Although the travel demand is not the largest, there is a top-down policy will that welcomes high-tech implementation, with strong policy driving force. It may enter the demonstration operation phase next year.

An important point in entering overseas markets is whether we can find good local partners. Pony AI will not operate vertically one by one in overseas markets, but can empower local partners with resources and willingness, adopting a light asset model.

Q: With Robotaxi profitability on the horizon, will Pony AI aim to be the first company in the industry to achieve profitability?

Wang Haojun: This is not my primary consideration at the moment. In my view, what we are focusing on today is turning UE positive. After UE turns positive, I will indeed have greater confidence to deploy more vehicles. However, if you ask me about the next growth target, I think the more important aspect is the volume, the growth itself, rather than breakeven itself Question: Recently, there has been a lot of discussion about the technology routes for autonomous driving. What are your views on the trends and advantages/disadvantages of the VLA and world model?

Wang Haojun: For Pony AI, whether it's the vehicle-side including BEV or end-to-end, VLA itself can serve as one line. The other line is about the world model, which is fundamentally a different approach from reinforcement learning and imitation learning.

Pony AI has been saying for two to three years that one very important point is that on the cloud training framework, the requirements for L4 are very clear. The safety level must actually be much higher than that of human drivers. Only when the safety capability reaches a significantly higher level will regulators allow you to enter a demonstration operation phase.

Because of this requirement, we believe that imitation learning itself does not work, so reinforcement learning is needed. Only through reinforcement learning can certain performance aspects be significantly better than human drivers.

This is why we shifted to reinforcement learning as a training framework five or six years ago, which is consistent with what is currently being discussed about the world model. From the perspective of safety requirements for L4 or Robotaxi, the most important aspect is the world model or reinforcement learning.

Question: How do you view car manufacturers (such as Tesla and XPeng) entering the L4 track?

Wang Haojun: More players entering is a good thing, indicating that everyone is optimistic about this industry and commercialization is imminent.

However, the key to L4 is safety. Recently, Musk himself admitted that FSD currently has two versions: one for Robotaxi and one for Model Y, which means that his original idea of using the same system to transition from L2+ FSD to L4 capability is not valid. XPeng also needs to consider the same.

In addition, L4 is a heavily regulated industry. Even if car manufacturers have many safety records in L2+, this does not directly help in obtaining an L4 license. Regulators require the mileage accumulated by the L4 system itself. New players need to spend time and capital to accumulate L4 mileage to prove safety, which leaves a window of opportunity for Pony AI.

Question: Car manufacturers often talk about data advantages. Will this be a shortcoming for Pony AI?

Wang Haojun: If they are still talking about data advantages, it means they are still following imitation learning, which involves using more data from actual road conditions for imitation. From the perspective of L2+, this is fine because the ultimate goal of L2+ is to drive like a human driver, and the capability ceiling can be similar to that of a human driver, which is a good product.

However, for L4, imitation learning itself does not work. The key is whether there is a good generative data architecture