L2 uses cameras and radar; autonomous vehicles go as far as LiDAR—the standard for autonomous driving is becoming a hierarchical multi-sensor system. NVIDIA DRIVE’s short video is not a spec sheet. On a Mercedes dashboard driving through downtown San Francisco, a reasoning prompt appears: “Yield to pedestrians on the crosswalk before turning right.” The message is clear: it’s not about how the car turns, but why it made that decision. Looking at the underlying hardware reveals more clearly where the industry is converging. It’s not one set—it’s two layers. The Tesla debate long centered on whether cameras alone are sufficient. But real-world products have already moved beyond this binary. The NVIDIA and Mercedes L2++ demo (CLA / https://t.co/7lbuc4Uypz Assist Pro) currently on the road includes roughly 10 cameras, 5 radars, and 12 ultrasonic sensors—no LiDAR. Since the driver remains the ultimate responsible party in this supervised system, cost and complexity are deliberately capped here. In contrast, NVIDIA’s high-end DRIVE Hyperion system for L4 and robotaxis includes 14 cameras, 9 radars, 1 LiDAR, and 12 ultrasonic sensors. This configuration ensures the vehicle won’t stop even if one sensor fails. The same architecture underpins the next-generation S-Class’s L4-ready design: “Sensor diversity + computational redundancy + software redundancy” is the foundation. 🚨 The reason this division is becoming the standard is simple: L2 and L4 carry vastly different costs of failure. In supervised systems, the human is the shield. In fully autonomous systems, the company is the shield. Applying a single-sensor philosophy to a product without a shield isn’t a matter of technical preference—it’s a question of accountability structure. Tesla FSD’s honest multiplier isn’t 7x—it’s 2x. Tesla’s safety page cites figures compared against the U.S. national average. These include older vehicles, different road types, and varying definitions of accidents. When comparing the same Tesla driven manually (with active safety enabled), the story changes. Recent North American data shows that for major collisions (airbag-deploying events), FSD-supervised driving achieves approximately 5.69 million miles per incident, while manual driving of the same vehicle averages about 2.08 million miles. That’s a gain of roughly 2.7x—up from around 2.4x in prior periods. In European non-highway conditions, the improvement is also around 2.7x. But this 2x gain reflects performance in a system where a human is watching. Tesla counts a trip as FSD-assisted if FSD was active up to five seconds before a collision. Even so, the fact that FSD still delivers a 2x safety margin over manual driving suggests that the safety buffer when transitioning to full autonomy may be thinner than assumed. This is precisely why Reuters questioned Tesla’s methodology: it’s not about inflated multipliers—it’s about framing supervised performance as if it were evidence of full autonomy capability. The fact that ADAS-related accident reports filed with NHTSA are disproportionately attributed to Tesla is a separate trust issue. Higher adoption rates naturally lead to higher absolute numbers. At the same time, Tesla makes it difficult for external parties to verify its denominator—the actual miles driven—thus eroding regulatory and public trust. Even if we cannot definitively say risk per mile is worsening, the mere lack of transparency alone undermines confidence. Data from fully autonomous segments is already tilting decisively toward multi-sensor approaches. Full autonomy is not yet the “coming soon” future Tesla claims it to be. Waymo is already operating commercially with over 200 million miles of paid robotaxi rides and 500,000 trips per week. Based on this experience, Waymo acknowledges cameras are “excellent” but concludes they are “not sufficient alone.” It cites cases where LiDAR detected pedestrians beside the road during dust storms that rendered cameras effectively blind. Cameras are passive sensors that rely on ambient light; LiDAR and radar are active sensors that emit their own signals—they differ fundamentally in physical properties. Tesla has also begun deploying fully autonomous robotaxis in Austin and other locations since early 2026 and has announced one million fully autonomous miles—a meaningful start. But scale and validation time still lag behind Waymo by at least an order of magnitude. Consumer FSD remains supervised. It is still too early to claim “vision-only has already solved autonomy.” The old argument that LiDAR is too expensive to deploy has weakened significantly. Automotive-grade LiDAR costs have dropped to around $200 per unit. Cost advantage alone can no longer justify sticking exclusively to cameras—at least not for L4 systems.In China, more vehicles are incorporating LiDAR even at urban auxiliary driving stages; even L2 may not be fixed as “LiDAR-free.” Still, hierarchies remain—determined by where cost-effective sensors are placed and how responsibility levels are assigned. The standard, the leftover, the foldable: To summarize the future standard in one sentence: Supervised systems use cost-effective multi-sensors (camera + radar); fully autonomous systems, where the company assumes responsibility, include LiDAR-based multi-sensors. The significance of NVIDIA Alpamayo lies in layering software on top of this hierarchy that can output not just decisions, but the reasoning behind them. Regulators and insurers don’t ask, “Does it seem smart?”—they ask, “Can you reproduce why it failed?” Inference traces and sensor redundancy answer this question in the same direction.