
The Great Autonomous Driving Route Debate
VLA and world models have been debated in autonomous driving circles for quite a while. Looking through the public materials from Li Auto and XPeng, both companies have already put the two together.
Two years ago, the argument was about "whether to use LiDAR"—that was essentially a hardware cost calculation, clear and visible. Later, the debate shifted to model routes, which sounded a lot like an academic summit of sword masters. Many outsiders watched in confusion, feeling it was just engineers showing off their skills.
But if we pull our perspective away from technical papers and put it back into the business world, we'll find an iron rule: there has never been a purely academic debate in the business arena. Behind every war of words over technical routes is actually a direct projection of each company's resource endowments, business narratives, and survival anxiety.
What Are They Actually Arguing About?
Setting aside the obscure English acronyms, VLA (Vision-Language-Action model) and world models actually represent two completely different approaches to "building a brain."
VLA is like building a "left brain that understands common sense." Its core logic is stuffing a large language model into the car. Its greatest ability is "understanding semantics"—for example, if someone temporarily puts up a handwritten sign by the roadside saying "Road work ahead, borrow the left lane," or a traffic officer makes a complex hand gesture, or you casually say from the back seat "pull over at that Luckin Coffee up ahead," VLA can understand the meaning behind it like a human and respond.
A world model is like building a "right brain that understands physics." It doesn't care much about language and words; it cares about the physical laws in three-dimensional space. It's like a super simulator—seeing the road conditions ahead, it can instantly simulate in its mind the game trajectories of a dozen surrounding cars over the next 5 seconds: "If I change lanes to the left, what's the probability the car behind accelerates and hits me?" At the same time, it can generate countless extremely dangerous "ghost pedestrian" scenarios out of thin air in the cloud, letting the car practice a lifetime of dangerous situations in advance in the virtual world.

At the same construction intersection, the car must both understand temporary instructions and judge how surrounding vehicles might move next. Physix Frontier illustration.
Both sound reasonable, right? So why do the big shots still question each other, even publicly "undercut" one another?
The Underlying Logic of Taking Sides: Whatever Cards You Hold, That's the Religion You Believe In
In business competition, where you stand depends on where you sit—that's the norm, but more accurately: genes determine the route.
You'll find that those who clearly and firmly support VLA are mostly new carmakers directly facing C-end consumers (like Li Auto and XPeng).

Photo of Li Auto i8 driving. Source: Li Auto.
Why? Because for automakers, autonomous driving isn't just a safety feature—it's an extremely expensive consumer selling point. When consumers sit in a car that drives itself, the biggest psychological barrier isn't that the car doesn't drive fast enough, but the "sense of losing control"—I don't know why it suddenly slows down, and I don't know if it saw the traffic officer ahead.
VLA happens to provide a huge commercial dividend: the sense of trust brought by explainability. It can understand your plain language and tell you why it's driving the way it is. This essentially transforms the car from a "cold transportation tool" into a communicative "AI professional driver." In an extremely oversupplied new energy market, this visible, understandable interactive experience can directly translate into extremely high emotional value and product premium.

Li Auto disclosed that during the 2025 National Day holiday, VLA commands were used 1.22 million times.
Those who remain vigilant about VLA and heavily bet on world models are often giants with a strong engineering DNA or B2B genes (like Huawei and Momenta).
From their position, large language models have an innate fatal flaw called "hallucination." When a chatbot talks nonsense, it gets screenshotted and laughed at for a couple of days at most; but when it comes to driving, the margin for error is zero. If the model suddenly "hallucinates" non-existent obstacles on the highway, or mistakes a large truck for a white cloud, the consequences are devastating.
More importantly, world models rely extremely heavily on massive cloud computing clusters and enormous data simulation capabilities. This asset-heavy, high-threshold approach that competes on underlying engineering certainty happens to be exactly where tech giants are best at building moats. For them, calculating the laws of the physical world to the extreme and pushing the safety bottom line to the highest is the hard currency that makes automakers willing to pay and regulators feel at ease.

Momenta R7 sharing session at the 2026 Beijing Auto Show. Source: Momenta.
The Truth Behind the War of Words: Fighting for Valuation Anchors and the Right to Define the Industry
Having broken it down to this point, you can probably see that this route debate is actually far more than just "whose technology is more advanced."
In capital markets and talent markets, narrative itself is productivity. Whoever can tell their technical route as "the only ticket to ultimate autonomous driving" can get a higher valuation in the next funding round and attract the top AI talent in the recruitment war.
Physix Frontier