NVIDIA's Xinzhou Wu: autonomous driving's ChatGPT moment has arrived, L4 mass production no longer a dream

In the rapidly evolving field of physical AI, autonomous driving is often viewed as the first major challenge to overcome. Recently, Wu Xinzhou, Vice President of NVIDIA, outlined the company's ambitious vision for intelligent driving at a Beijing communication event. He not only described the "five-layer cake" architecture supporting assisted driving but also provided a clear timeline for the rollout of Level 4 autonomous driving.

The 'Five-Layer Cake' Creates a Full-Stack Ecosystem

NVIDIA has moved beyond simply supplying chips, aiming instead to build a comprehensive service system through a trio of computing platforms: vehicle-side inference, cloud training, and simulation verification. Wu Xinzhou metaphorically refers to this as the "five-layer cake," spanning from the underlying Hyperion hardware platform, through the operating system, open model Alpamayo, and simulation toolchain, all the way up to the top-level cloud infrastructure.

At the heart of this system is a focus on lowering the development barrier for automakers. Particularly during the shift from modular to end-to-end architectures, NVIDIA leverages its powerful simulation capabilities to run two million scenario validations each day, significantly boosting model training efficiency. The company is now actively encouraging major car manufacturers to adopt the Hyperion platform, aiming for a major leap in standardization and scalability.

Vision-First Approach with Built-in Redundancy

On the technical front, Wu Xinzhou is a strong advocate of vision-based solutions. He argues that visual sensors offer pixel density and information limits far beyond those of LiDAR, making them more than capable of supporting high-level assisted driving. However, for more advanced L3 and L4 systems, NVIDIA still sees LiDAR as an essential safety redundancy. He revealed that the company is collaborating with suppliers in Europe and the US to secure stable hardware support for high-level intelligent driving solutions.

L4 Countdown: 30 Cities by 2028

Amid ongoing industry debate about skipping L3 to jump directly to L4, Wu Xinzhou takes a pragmatic view. He believes L3 offers immediate value by freeing human drivers from certain tasks, while L4 demands substantial cloud operations. According to NVIDIA’s roadmap: a mass-production project with Mercedes-Benz is set for 2025; an L4 pilot with Google is planned for 2027; and by 2028, NVIDIA intends to partner with Uber to offer driverless services during the Los Angeles Olympics, targeting coverage in 20 to 30 cities worldwide.

Physical AI Expands: From Cars to Robots