When James Peng founded Pony.ai in 2016, he expected robotaxis, or autonomous taxi services, to need at least a decade of work before moving from long-term vision to large-scale deployment. The technology had to mature. Laws and regulations had to catch up. Public acceptance also had to be tested.

That timeline has largely held. Waymo vehicles now move through the streets of San Francisco, while Pony.ai’s fleet accepts ride-hailing orders in Shenzhen’s busy Nanshan district. Autonomous vehicles have entered public traffic and begun commercial operations.

What Peng may not have anticipated ten years ago was that one of the key constraints on rapid fleet expansion would be a set of routine operations and maintenance tasks that become more complicated once the driver is removed.

When a human is behind the wheel, the driver handles charging, car washing, vehicle maintenance, and even helping passengers with luggage. Once autonomous vehicles enter service, these invisible errands become operational problems of their own.

In an interview, Pony.ai CEO James Peng spoke with 36Kr about robotaxi deployment and operations, the company’s choices around mass production for its Level 2 driver assistance business, its decision not to rush into embodied intelligence, and its technical approach.

Peng told 36Kr that Pony.ai has already built an operations team and developed a set of operations and maintenance standards to support autonomous vehicles. “This workforce includes remote safety operators, as well as ground support and logistics staff,” he said.

Even if the fleet expands rapidly, Peng said the staff-to-vehicle ratio will not rise significantly, breaking from the traditional assumption that more vehicles require more people. He believes this is an operational blind spot that today’s ride-hailing companies and automakers may overlook when entering the robotaxi market.

After nearly a decade of development, driverless vehicle technology has entered a stage of commercial operation across several fields as automotive intelligence has matured across the supply chain. Autonomous logistics vehicles are one example. Rino.ai, Neolix, Zelos, and others have attracted investor backing and become prominent names in China’s logistics sector.

Robotaxis, one of the most visible commercial applications for driverless vehicles, have attracted automakers, ride-hailing platforms, and other companies. Tesla, Xpeng, Geely, and others have announced plans to operate driverless vehicles.

One mainstream view is that automakers will become key robotaxi players because they have full-vehicle engineering capabilities and driver assistance-to-autonomous driving technology systems that can generate a data flywheel, where more vehicles on the road produce more data to improve the system. But Peng told 36Kr that manufacturing capacity is already abundant in China. For robotaxis, he said, any company that has not operated such a service before is still nearly starting from zero.

Peng believes technical capability answers the zero-to-one question for robotaxis, namely whether a company can build the system at all. Operating capability, by contrast, determines efficiency. Pony.ai is now exploring what it sees as an increasingly clear business model.

The road has not always been smooth. The sector has long faced commercialization challenges, and many autonomous driving companies have shifted their focus toward supplying Level 2 driver assistance systems, which still require human supervision, as they searched for more revenue channels.

Pony.ai is one of the few companies that has remained committed to robotaxis. Its seventh-generation vehicles have already achieved per-vehicle profitability in Guangzhou and Shenzhen. After validating its financial model, Pony.ai plans to expand its fleet to 3,500 vehicles this year.

Looking back at the choices it has made, Pony.ai has often taken a path that runs counter to prevailing assumptions.

Peng said the company briefly tried the large-scale Level 2 driver assistance business, but quickly concluded that it was likely to be low margin because “the technical threshold is low, the user experience is not standardized, automakers hold the bargaining power, and intelligent driving companies can easily fall into price wars.”

On the push by automakers and ride-hailing platforms into robotaxis, Peng was more pointed. “Making an announcement is always easy,” he said. “Tesla has been talking about it for ten years. Has it done it?”

Automakers generally hope to emulate Tesla by using existing end-to-end algorithm capabilities as a common technology base for both driver assistance and robotaxi businesses.

But Peng offered a technical view that differs from the mainstream. Pony.ai, he said, has not followed the trajectory of large language models by building a unified autonomous driving algorithm with a vast number of parameters. Instead, it has continued to use a strategy that combines multiple smaller models, improving operating efficiency and reducing dependence on computing power.

As driver assistance developers move away from maps, Pony.ai has again chosen not to follow the crowd. Peng said the company will continue to use a lightweight mapping approach for the long term.

“You’re familiar with the roads you drive often, so driving them is easier. Places you haven’t been to before are tiring. That’s normal. So why not add maps?” Peng said. “Even Tesla uses maps.”

Pony.ai has also decided against entering embodied intelligence at this stage, choosing instead to watch from the sidelines.

“This is another thing that will take at least ten years,” Peng said directly. “There will always be opportunities to do it, but you still have to see clearly and think clearly.”

KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by 36Kr Auto.