Looking at recent remarks from Xin Zhou Wu, head of NVIDIA’s automotive division, my long-held conclusion has become even clearer. Tesla is undoubtedly the leader in L2++. (This is why I’ve consistently said that the supervised FSD, where the driver remains ultimately responsible for accidents, represents Tesla’s limit.) Wu acknowledges this point—but that’s where it ends. Scaling a vision-only system like Cybercab across a large fleet is unlikely to be viable along the current path. Instead, it’s far more probable that Tesla will resort to limited-area demonstrations, concealing remote interventions and operational exceptions to buy time. Wu’s reasoning is straightforward: L2++ and L4 are not the same game. In L2++, the driver remains the ultimate responsible party. If the cameras get confused, the human can step in. Tesla’s advantage here is clear: it has long adhered to the vision-only approach and amassed vast amounts of data through its massive consumer fleet. Wu even stated, “In basic L2++, Elon is ahead of almost everyone.” L4 is different. There is no driver. If one sensor fails or one condition collapses, the vehicle must autonomously come to a safe stop. In Wu’s words, the fundamental principle of L4 is: “The vehicle must remain safe even in the event of a single point of failure.” That’s why NVIDIA designs its Hyperion base configuration with cameras and radar for L2++, but adds LiDAR for the advanced L4 variant. When I asked Wu whether LiDAR is necessary for L4, his short answer was “Yes.” He didn’t claim it’s 100% theoretically impossible without LiDAR—but he noted that achieving L4 with vision alone would inevitably shrink the Operational Design Domain (ODD). For customers demanding broad, real-world conditions, LiDAR is simply far superior. That single line is the core insight. It’s not that vision-only is “absolutely impossible.” It’s that scalable L4 is extremely difficult. You can mimic it under narrow geofences, good weather, well-maintained roads, and with remote operators ready to intervene immediately. But that’s not large-scale fleet operation. A true large-scale fleet faces rain, glare, construction zones, fallen objects, low-reflectivity obstacles, sensor contamination, nighttime sensor degradation, and unexpected road geometries—every single day. Cameras are passive sensors dependent on light; they don’t directly measure depth—they infer it. Inference performs well within learned distributions but collapses abruptly outside them. L4 cannot afford those moments of failure. Wu’s assessment that Waymo is an “already proven player” while Tesla is still “finding its path” fits this same logic. Waymo has already established repeatable, driverless service operations—even if limited to specific cities. Tesla’s vision-only Cybercab has not yet demonstrated such operational viability. Improved consumer FSD and a large-scale commercial fleet without safety drivers are entirely different challenges. The former succeeds by reducing intervention rates; the latter requires eliminating interventions entirely from the operational model. The data structure also works against Tesla. Tesla maintains a closed-loop fleet system—which is a powerful advantage for improving L2++. But for L4 safety certification and multi-OEM, multi-city expansion, different optimizations are needed. Wu emphasizes NVIDIA’s focus on enabling data and simulation sharing across OEMs. L4 cannot rely solely on one company’s limited set of driving scenarios—it must cover rare accidents, sensor failures, regulatory requirements, and regional traffic cultures. Closed loops may be fast, but they weaken verifiable redundancy. The cost logic also no longer clearly favors Tesla over time. The biggest advantage of vision-only is low hardware cost. But as you move toward L4, models grow larger, inference and redundancy computations increase, and transitioning to next-generation chips (AI5-class) dramatically raises computational demands. Wu acknowledged this point. To fill safety gaps left by cameras alone with larger models and more computation means the initial savings on sensor costs are quickly offset by rising expenses in chips, power, cooling, and validation. Add to that remote support personnel, operational restrictions, accident response protocols, insurance costs, and city-specific exceptions—and “cheap hardware” rapidly becomes “expensive operations.” Therefore, the realistic path I see for Cybercab is this: First: Limited operation on restricted routes and during restricted hours in select cities. Second: Rely on remote intervention and vehicle recovery when problems arise. Third: Maintain the illusion of a “driverless robotaxi” while repeatedly pushing software updates.Fourth, the comprehensive large-scale fleet expansion is continually postponed to the next quarter, the next city, the next chip. This is a delay strategy that refuses to acknowledge failure—because it is difficult to compensate for the physical redundancy and regulatory-grade safety validation required for L4 commercial operations using camera-based reasoning alone. The ODD limitations U mentioned are precisely this point. Demonstrations are possible. Limited success is achievable. But that does not equate to the everyday operation of a large-scale fleet. There is no need to diminish Tesla’s achievements in L2++. Those achievements are real. However, equating them with the completion of an L4 cybercap leads to logical collapse. The significance of Wu’s remarks lies not in NVIDIA dismissing Tesla, but in clearly distinguishing Tesla’s strengths from its limitations. Tesla remains the leader in L2++. Proven operational deployment of L4 is still closer to Waymo. A vision-only cybercap cannot instantly bridge that gap through data and computation alone. This is exactly as I anticipated. The cybercap will appear “functional” for some time—footage looks impressive, and certain segments perform well in practice. But to become a true large-scale autonomous transportation network capable of routinely handling weather, urban complexity, and edge cases, the current architecture is insufficient. While Tesla buys time and the market maintains its expectations during this gap (naive Tesla fans will continue to be misled), Wu’s interview points to the fact that this time cannot be extended indefinitely.