Autonomous driving has never been a simple technical challenge.

Some say that NVIDIA's latest development kit can help traditional automakers catch up with Tesla's FSD. But this analogy is quite fitting—Lego launching a space shuttle set doesn't threaten the real Falcon 9 rocket at all.

NVIDIA has indeed done a lot of useful work. Multiple generations of ADAS development kits, hardware tools, prototype computing frameworks based on its chips... These things alone do not constitute a complete autonomous driving system; they merely lay a starting point for developers. Quickly assembling a demo system with these tools is possible—effects can be seen within weeks—but mass production is another matter. From concept to implementation, these kits might only have completed 0.01% of the work. The most difficult parts—real-time decision-making, redundancy systems, coverage of millions of scenarios—almost all have to be developed independently.

The new generation of kits claims to use VLA as the core architecture. VLA indeed offers many development conveniences, but its computational load is enormous, making it impractical for direct use in mass production.

What does reality look like? By the end of this year, Tesla's cumulative expenditure on NVIDIA training hardware alone will approach $10 billion. They have also developed their own AI4 chips to handle massive video data, otherwise costs would double. Currently, they produce over 2 million vehicles annually, all equipped with dual SoC AI4, 8 cameras, redundant steering actuators, and high-bandwidth communication systems. This is a complete, validated system.

Traditional automakers starting from scratch? Not next year, and probably not even in five years will they see a mature mass production solution. It’s not a matter of technical impossibility, but requires massive investment, years of real-world testing data accumulation, and bearing significant business and legal risks. If one company truly succeeds, it would be almost a miracle.

The good news is that the process of replacing human drivers with autonomous driving requires the participation of hundreds of millions of vehicles. Relying on a single company to complete this task within a reasonable timeframe is impossible. So, if other manufacturers invest seriously, it’s a good thing for the entire industry. But the return cycle for such investments will be long—even if successful, it will take a decade before roads are truly transformed.

In this process, fleet size and the moat built from real data remain the strongest defenses.