Dialogue with Pony.ai’s Haojun Wang: 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 was a code game in the laboratory, a fantasy built with demos and PowerPoints; but now, this business has officially fallen from the void to the ground and has started to clash head-to-head on financial statements.

As once-celebrated L4 unicorns stall due to their inability to cross the threshold of scale, pioneers have quietly knocked on the door of profitability. In Q2 2025, Baidu's Apollo Go achieved break-even in Wuhan; in November, Pony.ai announced that its seventh-generation Robotaxi in Guangzhou achieved a positive unit economics (UE) model.

Pony.ai co-founder and CFO Wang Haojun told Wallstreetcn in a recent interview that being able to achieve positive UE in Guangzhou means Pony.ai has gradually honed a standard operational process during its scaling, which can empower its partners.

Wang Haojun believes that in previous years, Robotaxi commercialization was largely at the “0 to 1” stage, but now it is gradually moving to the “1 to 100” or even “1 to 1000” stage.

A clear commercialization timetable has emerged: rushing toward a fleet of over a thousand vehicles by the end of 2025, increasing to 3,000 vehicles in 2026, and reaching 100,000 vehicles by 2030, Robotaxi will become part of daily life.

This means the main battlefield for Robotaxi competition has shifted. As single-vehicle hardware costs drop to the critical threshold of 250,000 RMB, a new race for profits has begun; meanwhile, AI is irreversibly reshaping the rules of the physical world.

Commercial Closed Loop

Achieving positive unit economics (UE) in Guangzhou means Pony.ai can realize a commercial closed loop in the regional market. It is a thrilling leap from “lab R&D” to “commercial closed loop.”

So-called positive UE means that, after deducting total R&D investment, each car running on the street is now generating enough income to cover hardware depreciation and operating expenses.

In the past, Robotaxi was a “cash-burning furnace” that troubled investors, with every tested car on the street carrying huge modification costs and operating losses—the logic of “the more you invest, the more you lose” has long haunted the industry.

But the numbers in Wang Haojun’s ledger show that the turning point has arrived. Pony.ai's seventh-generation Robotaxi in Guangzhou averages 23 orders per day with a daily revenue of about 299 RMB.

According to industry benchmarks for normalized operations, when average daily orders reach 24 per car, both user and revenue ends can form a strong positive cycle, and the figure of 299 RMB is enough to cover all costs, including hardware depreciation and operation.

Balancing the books depends on cost control and ongoing operational optimization.

Compared with the sixth generation, Pony’s seventh-generation vehicle has cut the bill of materials (BOM) cost for autonomous driving by 70%. Pony has completely abandoned the traditional industrial PCs (IPC) used in Robotaxis, which cost tens of thousands of RMB and consume huge power, switching to large-scale use of its own domain controller based on four Nvidia Orin X automotive SoC chips—becoming the first L4 player in the industry to achieve mass production with such a plan.

By leveraging the scale effect of locally produced Chinese L2+ vehicles, Pony has lowered lidar costs by 68%. The seventh generation no longer relies on expensive customized mechanical lidar, instead adopting automotive-grade solid-state sensors like Hesai AT128 that have entered large-scale production cycles. All 34 sensors in the vehicle now use 100% automotive-grade components.

Meanwhile, algorithmic optimization continues. Wang revealed that software optimization has improved the system’s ability to handle “noise” in sensor data by 30 times. This means Pony can use cheaper, more standard hardware while achieving even safer and smoother performance than the previous generation.

Beyond hardware, every operational expense is carefully calculated. Wang Haojun told Wallstreetcn that thanks to a safety record far surpassing human drivers, Pony’s Robotaxi commercial insurance annual fee is 50% lower than traditional taxis—essentially an endorsement by insurers of AI drivers’ safety, in cash.

In remote assistance, human efficiency is also improving: currently the remote assist staff to vehicle ratio is 1:20, and is planned to reach 1:30 by year-end. Maintenance is standardized through a digital work order system—after returning to the depot, cleaning and maintenance for the vehicles is batch-processed, achieving near-20 vehicles per staff member.

This series of cost reductions has allowed Pony.ai to realize positive UE in Guangzhou. Wang Haojun said this also gives Pony confidence to deploy more vehicles and further drive financial growth through network effects.

Wang believes that globally, the Middle East is currently most worth watching. As its market enters the demonstration operation phase, the Middle East will become a market for sustainable revenue.

While running standard models, Pony is also shifting from heavy assets to light assets. Wang is blunt: the company cannot stubbornly stick to self-operation, for the capital expenditure (CAPEX) required to own 100,000 vehicles is “financial suicide” for any startup.

As a result, Pony has constructed a “shared benefit” value chain: OEMs produce the vehicles, asset companies like Xihu Group and Sunshine Mobility buy and hold them, traffic platforms like Amap and Ruqi handle order distribution, while Pony provides the core “AI driver” brain.

Under this model, Pony's income shifts to sales revenue, technology licensing fees, and service commission from each order. This marks a switch from asset-heavy explorer to ecosystem enabler, aiming for fleet expansion to 3,000 vehicles by 2026.

Industry Melee

By 2025, the Robotaxi track has become an all-out industry battle involving tech giants, mobility platforms, and automakers. On this trillion-level gambling table, everyone is seeking a spot in the finals through alliances, cross-dimension attacks, and ecosystem restructuring.

Waymo represents the high ground of American “engineering perfectionism.” A new $15 billion fundraising round pushes the pioneer’s valuation to $100 billion.

Waymo pursues extreme redundancy and standardization. Its weekly orders have surpassed 450,000, and since 2025, it has provided over 14 million trips, proving the feasibility of the “long-term cash flow” narrative to capital markets—but this model also comes with a heavy financial burden.

Wang Haojun points out that Waymo’s soon-to-be-deployed new models still cost four to five times as much as similar Chinese models. This heavy asset approach builds technical barriers but leaves a window of opportunity for cost-effective disruption by Chinese players with robust supply chains.

Tesla's Robotaxi is also accelerating deployment. Musk has officially removed safety operators in Austin, aiming to reshape the world's dispatch rules with a low-cost pure-vision solution and ambition for a million-vehicle scale. At its core, Tesla’s bold entry is enabled by shifting its cloud training framework to reinforcement learning and generative data.

Even Uber, which had momentarily abandoned self-developed brains, is returning as a “neural network” dispatch platform, collaborating with Waymo, WeRide, Pony and others. This “OEM+tech provider+platform” triad network is squeezing the prospects of isolated tech firms.

The “hunting” by new domestic entrants is heating up. XPeng Motors announced that by 2026, it would mass produce three pure-vision Robotaxi models (no lidar), using second-generation VLA+VLM tech and launching in Guangzhou and elsewhere. Hello Inc, in alliance with Dongfeng Venucia and Horizon Robotics, aims for a scale of 50,000 vehicles by 2027.

On the side, Baidu's Apollo Go services now exceed 250,000 orders per week, with over 17 million cumulative trips served.

Meanwhile, Caocao and Jiandao Mobility are quickly moving into core city areas by deeply partnering with Qianli Tech, Momenta and others.

This new round of industry melee marks a fundamental change: Robotaxi, once “black tech,” is now a comprehensive brawl of capital, supply chain, and operational efficiency.

This also sketches the “second half” of automotive industry intelligence. UBS Investment Bank’s China Auto head Gong Min says that Robotaxi prices can now drop below 300,000 RMB, leading companies have reached a scale of thousands, and are beginning to break even per vehicle in certain regions.

Industrialization and advanced cost-control tech are making new technologies ever more affordable.

If tech companies can't quickly bind deeply with OEMs and platforms, they will wither under dual exhaustion of data and funds. Ultimately, this is no longer a contest of algorithmic elegance among engineers—it’s a battle for who can first occupy the physical space of cities and achieve a dispatch scale in the millions.

System Competition

No one wants to miss the trillion-level Robotaxi market. Especially as technology hits the tipping point, hyper-fast development and even overnight mass adoption may follow.

Gong Min told Wallstreetcn that by 2030, if deployed in some Chinese first-tier cities, the Robotaxi market size will be $8 billion; if nationwide in China, $183 billion; counting all overseas markets outside the USA, $394 billion.

Once technology crosses the feasibility threshold, algorithms are no longer the only trump card. Robotaxi is indisputably entering the “second half,” dominated by operational efficiency.

Why does operations matter so much? UBS reports that as hardware costs fall, the share of operational expenses (maintenance, insurance, recharging) in per-vehicle cost will rise from 48% to 55%. Thus, future competitive leadership shifts from software engineers to those who truly understand urban dispatch.

The “golden node” for scale effectiveness is set at 100,000 vehicles. Wang predicts Pony’s goal for 2030 is 100,000 cars, equating to 5~10% share of first-tier city markets in China. Upon reaching this threshold, powerful network effects will bring qualitative change—in user acceptance and further scaling.

To prepare, leading players are moving resolutely away from the dead end of imitation learning. Wang emphasizes L4 must be 10 times safer than a human; this requires reinforcement learning and a generative “world model.” Such algorithms are not just to drive well, but to train a general AI with chain-of-thought (COT) and logical reasoning abilities.

From a grander narrative, the real battle on wheels for smart driving companies is actually the early field fight for AGI (Artificial General Intelligence). Today’s smart cars are essentially robots on wheels. Horizon Robotics’ “intuition system” aims to give AI human-like intuition, while Pony.ai’s VLA empowers machines with the ability to understand road signs and reason about complex tidal lanes.

They’re all aiming at the same outcome: enabling AI to perceive and interact with the physical world through deep understanding of physical rules, not just command execution.

Robotaxi is the largest-scale, most real-time, and lowest-tolerance “combat zone” for AGI in the real world. The logical reasoning and game-theory algorithms forged in this “meat grinder” will be reused in the wider field of embodied intelligence. Seizing the Robotaxi high ground means getting the ultimate ticket to rewrite the rules of intelligence for the physical world.

In this brutally competitive quest for survival and sovereignty, the key to breakthroughs will lie with those who deeply understand vertical industry depth and achieve extreme operational efficiency. The transformation has begun.

The following is an edited transcript of Wallstreetcn’s interview with Wang Haojun:

Q: Why is Pony.ai pursuing an “asset-light model” now? How does revenue sharing work?

Wang Haojun: Actually, this comes back to the difference between L4 and L2+. When you’re offering an L2+ product, it’s a partial driving function; the value varies for each user. But for L4, we’re offering a complete driving function—that is, doing the entire job of driving.

We expect in the L4 field, both license (authorization) and revenue sharing are highly recurring and important. This definitely exists.

As UE turns positive, more companies will want to get involved. What we’re focused on is whether, via more efficient capital operation, we can deploy more cars with less capital. We’ve also realized that many traditional operators are heavily asset-based—it’s their historical model. The new business model isn’t disruptive to them; it’s a good extension.

As for revenue sharing, we’re still in early exploration; our economic interests are tied to three parts: vehicle sales revenue, technology licensing, and service revenue sharing.

Q: There’s much discussion about the boundary between L3 and L4. What evolution do you foresee?

Wang Haojun: L3 can hardly provide service like L4. L3 still needs a human driver; as long as you need human presence, you can’t save the labor cost for Robotaxi, so positive UE (unit economics) is impossible.

L3 does offer a way to move private cars upward in intelligence. If it succeeds, people may be more willing to accept private vehicles with L4 features based on Robotaxi success.

Q: Will Pony.ai make L4 products for the consumer market?

Wang Haojun: From a business point of view, never say never, but for now, it’s too early for us.

Q: What’s Pony.ai’s logic for overseas markets? Which regions are priorities?

Wang Haojun: Overseas and China’s Robotaxi capacity are at very different levels. I expect next year the total volume allowed for demonstration overseas will be a few hundred vehicles at most.

Overseas is still in its early expansion—true commercialization needs at least a thousand vehicles. Robotaxi is highly regulated—so local mileage must be accumulated. Overseas layouts need to be front-loaded, with safety records built first.

China and the US are the biggest markets, followed by the EU. Other priority areas are Japan, South Korea, Australia—places with high labor costs and mobility demand.

The Middle East is unique. Though its mobility demand isn’t the largest, there’s top-down policy will, strong support for high-tech, and strong policy momentum. Demonstration operation may begin as early as next year.

When expanding overseas, it’s crucial to find good local partners. Pony won’t do vertical operation alone abroad, but will empower local partners with resources and willingness—also via the asset-light model.

Q: With Robotaxi nearing profitability, is Pony aiming to be the first in the industry to turn a profit?

Wang Haojun: That’s not my main focus now. My priority is achieving positive UE. After that, I’m more confident about deploying more vehicles. The most important thing is volume and growth, not just breakeven.

Q: There’s been much discussion about technical routes; what are your thoughts on VLA and world models?

Wang Haojun: For Pony, BEV and end-to-end at the vehicle side—VLA can be one line; the other is the world model, a different thinking from reinforcement vs imitation learning.

We started talking about this two to three years ago: for cloud training frameworks, L4 requirements are clear—safety must surpass human drivers by a lot. Only then will regulators allow demonstration operation.

So imitation learning doesn’t work; we need reinforcement learning to be much better than humans in some respects.

That’s why we shifted to reinforcement learning as our training framework five or six years ago; it’s consistent with today’s emphasis on world models. For L4 or Robotaxi safety, world models, or reinforcement learning, are most important.

Q: How do you view carmakers (e.g., Tesla, XPeng) entering L4?

Wang Haojun: More players is a good thing—it means people believe in the industry and commercialization is coming.

But safety is the key for L4. Recently Musk himself admitted FSD has two versions: one for Robotaxi, one for Model Y. That means one set of things can’t just evolve from L2+ FSD to full L4. XPeng faces the same issue.

L4 is highly regulated—even if carmakers have good safety records at L2+, it doesn’t help get L4 licenses. Regulators require specific mileage from L4 systems. New entrants need time and capital to accumulate this, giving Pony a window of opportunity.

Q: Carmakers highlight data advantages. Is this a weakness for Pony?

Wang Haojun: If they're talking about data advantages, they're still doing imitation learning—using more real-world data for imitation. For L2+, this is fine; the goal is to drive like a human, and that’s a good product.

But for L4, imitation learning doesn’t work. What matters is a good generative data architecture.

Risk Reminder and DisclaimerThe market involves risks—investments require caution. This article does not constitute individual investment advice, nor does it take into account individual users' investment goals, financial situation, or needs. Users should consider whether any opinions, views or conclusions in this article are appropriate for their specific situations. Invest accordingly, at your own risk. ```