
From Lin Dahua's Interview: 5.5 Million Units Shipped Cumulatively — Why Did SenseTime's SenseAuto Spin Off Its Autonomous Driving Business?
From Lin Dahua's Interview: 5.5 Million Units Shipped Cumulatively — Why Did SenseTime's SenseAuto Spin Off Its Autonomous Driving Business?

SenseAuto's booth at the Beijing Auto Show, with the Sage Box edge AI device beside the display car.
SenseTime has moved SenseAuto's loss-making autonomous driving business off its financial statements — that business now has to go find money in the market on its own.
At this April's Beijing Auto Show, SenseAuto presented a cockpit-driving integrated product, with the Sage Box edge AI device placed at the center of the booth.
Around the same time, SenseTime co-founder Lin Dahua gave a long interview, repeatedly talking about cost.
Forty thousand models a year — the math doesn't add up
SenseTime's prospectus at its 2021 listing mentioned that a platform called Model Factory produced 40,000 models a year.
Lin Dahua said these weren't 40,000 independent projects — many were repeated iterations along the same route.
A new client comes in, the industry changes, data has to be recollected, models retrained, and only after testing can they enter the production environment. After all that work, the model can still only handle a very narrow slice.
After completing the first benchmark client, the second one is just as expensive. Lin Dahua put it bluntly: the money earned can't cover such a long delivery chain.
By the end of 2025, SenseAuto had partnered with more than 30 automakers, including Chery, NIO, Honda, and BYD, covering 188 vehicle models with cumulative shipments approaching 5.5 million units.
In the same year, SenseTime's total revenue was 5.015 billion yuan, with a net loss of 1.782 billion yuan. The in-vehicle software sold didn't earn back what was invested.

Sage 32B on-device model: three sets of metrics — zero-cost tokens, always online, and scenario reasoning — plus specific performance on tool calling, planning, and inference speed.
Spinning off autonomous driving makes the numbers clearer
In the interview, Lin Dahua explained SenseTime's "1+X." The "1" refers to the two technological foundations of generative AI and computer vision AI; the "X" refers to separately operated businesses such as smart vehicles, healthcare, and robotics.
He gave an example: the same manager, within the group, could request 50 people, but going out to run things independently, they'd only take 5. Asking leadership for resources versus surviving in the market forces completely different judgment and efficiency.
SenseTime's financial report for the first half of 2025 showed that SenseAuto's loss-making intelligent driving and world model businesses were spun off and no longer counted in the group's statements starting in August.
The report showed "X innovation business" revenue of 302 million yuan, down 5.92% year-on-year. SenseTime's explanation was that autonomous driving had been moved off the statements.
The group's net loss narrowed accordingly to 1.782 billion yuan, 58.6% less than the previous year. Second-half EBITDA was 380 million yuan, positive for the first time since listing.
Lin Dahua didn't spell out this causal link. But the group's books did get lighter because of it: the loss-making business is no longer consolidated, and the money-burning has been handed off to the outside.
32B — bigger isn't always better
In the interview, Lin Dahua disagreed with the approach of piling on parameters. He said parameters determine the ceiling, but what companies need to handle is competition that must be settled within two or three months.
A several-hundred-B model, as long as training, architecture, and data are done solidly, iterates faster and in three months may well match a larger one.
He also wants the industry to go back and check data quality. Models with two or three trillion parameters losing to ones with a few hundred billion often comes down to impurities and conflicts in the data not being cleaned up properly.
The Sage 32B that SenseAuto unveiled at the auto show follows this thinking: it runs on the vehicle side, not connected to the cloud.
According to the vendor's figures, token cost is zero, response time is 0.5 seconds, one vehicle saves 30 yuan a day in cloud fees, and ten thousand vehicles save over 100 million yuan a year.

The Sage Box edge AI device, running a 32B model on the vehicle side.
Lin Dahua also said: a model is just a processor; what's valuable is the real-time information fed into it.
In the car, for an offline 32B model to be useful, the vehicle itself must understand the road, understand people, and understand what the driver wants to do.
Creating a segment and delivering a whole series are not the same thing
The interview touched on video models. Lin Dahua said what everyone compares now is mostly tens-of-seconds generated clips.
But what creators need to deliver is a hundred-episode short drama — scripts, storyboards, characters, and consistency between episodes all have to be managed.
"Creation and generation are completely different concepts," he said. SenseAuto's video platform therefore moves toward the creation workflow rather than stopping at generating clips.
The same thinking was carried over to the car: cockpit and autonomous driving done together.
R-UniAD 2.0 is what SenseAuto calls a unified L2-to-L4 foundation that can run on chip platforms such as J6P, J6M, and NVIDIA Thor. SenseAuto Go, meanwhile, packs autonomous driving, the cockpit agent, and the Sage Box into a single Robotaxi solution.
Physix Frontier