I have watched several in-depth interviews recently, and three core players in the new energy vehicle sector are all betting on the same direction—VLA (Visual Language Action Model) driver large model. This is not a coincidence but a technological inflection point that the entire industry will face by 2025.

**Is VLA really the iPhone 4 moment for intelligent driving?**

Li Xiang, founder of Li Auto, directly states: VLA represents the "iPhone 4 moment" for automotive intelligent driving. His logic is clear—advancing VLA in three stages: from information tool → reasoner → intelligent agent, using a combination of RLHF, world model training, and collision feedback. He also observes that the gap between domestic large models (like DeepSeek, Qianwen) and those in the US has already narrowed significantly.

Xpeng Motors Chairman He Xiaopeng’s judgment is slightly more conservative, but the conclusion is similar: 2025-2027 will be the true inflection point for L3+ autonomous driving, and automakers that cannot keep up will be eliminated. It took him over two months to decide to bet on VLA, which shows the weight of this choice. He emphasizes one point many overlook—the response time for autonomous driving needs to be in milliseconds, which is completely different from ordinary large models (non-real-time, low reliability).

**What is the true moat?**

Horizon founder Yu Kai provides another dimension of answer. He says he has witnessed China’s AI from catching up to running parallel. But his positioning of Horizon is very clear—it’s not a pure chip company but a "chip + software" system-level intelligent driving technology company. Horizon’s chips have shipped over 8 million units so far.

His entrepreneurial philosophy is quite interesting: "The Jianghu (martial arts world) is not about fighting and killing, but about human relationships." He believes the long-term moat for intelligent driving relies on the accumulation of hard, dirty, and tiring work—vehicle-grade quality, massive scenario validation, organizational culture sedimentation—these cannot be achieved overnight. He also shared a sobering statistic: high-level intelligent driving requires a team of thousands and an annual investment of over 1 billion yuan.

**The cost and price of survival**

Li Xiang emphasizes the importance of capital operation and equity structure, noting that Li Auto has the best governance and cash management among new forces. He Xiaopeng is more straightforward— the industry has entered a survival-of-the-fittest race, no one dares to lie flat, and Xpeng must learn to swim in the "sea of blood." They are all doing the same thing: expanding from a price range of 200,000-500,000 yuan to 100,000-500,000 yuan, and achieving profitable, high-quality products at the critical 150,000 yuan price point. In other words, doing higher quality at lower costs—that’s the logic of survival.

**Large models are just a means**

Li Xiang also has an interesting thought: AI is responsible for "intelligence" (capability can be infinitely improved), while humans are responsible for "wisdom" (the relationship with all things). Regarding safety, they set very high goals—VLA aims to reduce major casualties by over 90%. But He Xiaopeng is more pragmatic—if 10% of new car users in first-tier cities use high-level intelligent driving each month, it will trigger a real trust inflection point.

The consensus from these three interviews is clear: large models, VLA, and chips are all important, but true competitiveness comes from systematic capability accumulation, team, capital efficiency, and grasp of industry rhythm. 2025 is not a watershed for technology but a watershed for business models and execution.