Why Does Nvidia Prioritize the Automotive Industry for Its Physical AI Rollout?

After the popularization of text and image generation by large language models, global tech giants are turning their attention to "Physical AI", which enables artificial intelligence to step out of digital screens, perceive, reason, make decisions and interact in the real physical environment, and directly drive machines to act.

Among the numerous carriers of Physical AI, intelligent vehicles that have achieved mass production and are equipped with cameras and radar sensors across the entire vehicle have become the first test field to run through the commercial closed loop.

Wu Peidong, Managing Director of Wan Venture Partners, told a reporter from CLS that currently only the automotive industry can meet the three conditions of data scale, supply chain and cash flow at the same time. More than 20 million new passenger vehicles hit the road in China every year, the marginal collection cost of data return is close to zero, sensors, batteries and other components are already mature, and consumers are willing to pay for intelligent functions; in contrast, the global shipment of humanoid robots in 2025 was about 18,000 units, and data collection still relies on teleoperation and motion capture.

In fiscal year 2026 (ended January 25, 2026), NVIDIA's automotive business revenue reached 2.349 billion US dollars, a year-on-year increase of 39%. In China's smart electric vehicle market, the competition faced by NVIDIA's in-vehicle chips is intensifying. Automakers including NIO (09866.HK), XPeng Inc. (09868.HK), Li Auto (02015.HK) have successively promoted the deployment of self-developed intelligent driving chips, while local intelligent driving chip vendors such as Horizon Robotics (09660.HK) are also expanding their designated project scale in the cost-effective solution market.

Data from NE Times shows that in August 2026, the total installed volume of domestic passenger vehicle assisted driving domain control chips was about 810,000 units. NVIDIA ranked first with a 30.2% market share, followed by Horizon Robotics and Huawei at 19.8% and 14.8% respectively; from January to February 2026, NVIDIA's share stood at 45.1%, leading Horizon Robotics by 35.8 percentage points, and the gap had narrowed to 10.4 percentage points by August.

Meanwhile, many leading automakers are shifting to the "world model" and "vision-language-action" large model routes this year, and the focus of competition has extended from in-vehicle chips to cloud-based model training.

Recently, CLS reporters communicated with a number of relevant NVIDIA experts and learned that centering on intelligent vehicles, the core carrier of Physical AI, NVIDIA's technical route and business logic have undergone obvious changes. On the cloud side, NVIDIA opens the underlying world models and simulation tools to the industry for free, and gains returns more from the demand for computing power; on the vehicle side, NVIDIA carries out software optimization around automakers' cost reduction needs. This technology stack will be further extended to unmanned driving and robots in the next step.

Behind the Free Open-Source Models

Since the beginning of this year, end-to-end large models have become the standard configuration for intelligent driving upgrades of domestic automakers. However, in actual road conditions, existing end-to-end models only directly convert camera images into steering wheel and brake actions, without understanding the causal logic of the physical world.

For example, when a vehicle encounters a traffic police officer gesturing to direct traffic at an intersection, an ordinary end-to-end model can only recognize that "there is a person in front that needs to be avoided", but cannot understand the meaning of the gesture. After introducing a "world model" that understands physical laws and common sense, as well as a model with reasoning capabilities, vehicles are expected to understand complex road conditions and make predictions just like experienced mature drivers.

However, the threshold for training a foundational large model that understands physical laws from scratch is extremely high. Public information shows that as of the first quarter of 2026, Tesla's put-into-operation Cortex 1 training cluster is equivalent to more than 100,000 NVIDIA H100 Graphics Processing Units (GPUs), and the Cortex 2 cluster in the initial ramp-up phase is equivalent to more than 130,000 units; XPeng has also built an intelligent computing cluster with a scale of over 10,000 cards. Most automakers and intelligent driving enterprises cannot afford such a large investment independently.

In response, NVIDIA's approach is to first train the underlying foundational model and open it to the world for free.

At present, NVIDIA has opened four core models for Physical AI, including the language model for pre-training (Nemotron), the world foundational model for understanding physical laws (Cosmos), GR00T for robots in the post-training stage, and Alpamayo for autonomous driving respectively.

Relevant NVIDIA staff told CLS reporters that NVIDIA hopes to use its own resources and technology to help the industry achieve a 60-point performance for the foundational model through open models, so that automakers only need to use a small amount of computing power and their own road test data for post-training fine-tuning; if platform enterprises do not take the lead in developing the foundational model, it will be very difficult for Physical AI to be truly implemented in the industry.

In addition to cloud model training, it is also difficult to collect all kinds of rare sudden dangerous road conditions through real road tests. How to verify safety at low cost before the model is put on the road after training is also a difficult problem faced by all major automakers.

To this end, between the computers responsible for cloud training (DGX) and the computers responsible for vehicle-side inference (AGX), NVIDIA has specially built computers responsible for virtual simulation (OVX), and opened the 3D Gaussian Splatting restoration tool (NuRec) and closed-loop simulation tool (AlpaSim) to developers. According to NVIDIA's design, automakers can convert a section of ordinary road test video into an editable 3D virtual scene, and repeatedly test the model's response in the virtual environment to shorten the real vehicle verification cycle.

Wu Peidong believes that these open tools "change the starting point, not the end point". In the past, it took two to three years and an investment of at least tens of billions of yuan to develop a set of high-level intelligent driving systems. Now teams can first run through the closed loop and then make differentiation, and the starting cycle can be shortened by more than one year. He also said that leading automakers will take these tools as the benchmark, and still develop the core technology stack independently; second and third-tier automakers lack engineering teams and data backflow, and these tools give them a deliverable lower limit. However, free does not mean zero cost, and the investment in computing power, data and calibration cannot be saved at all.

After opening the models and simulation tools that cost huge sums of money to train, where does NVIDIA's commercial return come from? The company said that this year it has redefined "developers" in the automotive industry as "all enterprises that develop and sell software on their own", including automakers and intelligent driving suppliers.

The aforementioned relevant NVIDIA staff told reporters that only by helping developers build their own AI can customers use NVIDIA's platforms and products more. According to the introduction, in the process of promoting autonomous driving to Robotaxi, major manufacturers in the automotive industry are currently using a large number of NVIDIA's open technologies and toolchains; as long as the whole industry uses this open framework when evolving towards high-level intelligent driving, it will eventually be converted into continuous demand for NVIDIA's cloud training and simulation server computing power.

Automakers Are "Cutting Costs" on the Vehicle Side

After the large model completes training and simulation verification on the cloud, it must finally be installed in the end-side chip of the vehicle to run in real time. In China's fiercely competitive and price-sensitive automotive market, automakers not only require the model to run fast, but also require to reduce hardware costs. Li Bin, Chairman of NIO, said in April this year that NIO used to spend 300 million US dollars a year on NVIDIA chips at its peak, and self-developed chips have saved the company a lot of money.

In the vehicle-side inference link, following the widely mass-produced Orin chip (with a computing power of 254 TOPS, i.e. 254 trillion operations per second) in China, NVIDIA launched the new generation in-vehicle chip Thor with a computing power of up to 2000 TFLOPS (2000 trillion floating-point operations per second), as well as a lightweight inference model with only 2 billion (2B) parameters, and a world model with 4 billion parameters that can be deployed on the edge side.

However, in the actual mass production process of automakers, the nominal peak computing power of the chip does not represent the final experience. What is more critical is how to make full use of the existing computing power and reduce the total vehicle bill of materials cost through underlying software and system optimization.

Due to the huge number of parameters of large models, the first problem faced when deploying to the vehicle side is that the computing power of existing vehicle models is insufficient. At present, a large number of main models in China are still using the previous-generation Orin chip, whose hardware itself does not support native FP4 or FP8 (8-bit floating-point) computing, and cannot directly run large models compressed with such low-precision formats.

CLS reporters learned from NVIDIA that the underlying software stacks of the Orin and Thor chips are consistent. Combined with speculative sampling technology (using a small model to predict multiple tokens at the same time to speed up output) and other means, the Orin platform can reach a throughput speed of 70 tokens per second when running a 7 billion parameter (7B) multimodal model, and the Thor platform can reach 176 tokens per second after combining 4-bit floating-point quantization (NVFP4).

The second problem for vehicle-side deployment is the cost pressure brought by the continuous rise of memory chip prices. Public information shows that from March to June 2026, the price of automotive-grade memory chips rose by about 180%; Li Bin said in March this year that the rise in the prices of storage, memory and other raw materials had an impact of 3,000 to 5,000 yuan respectively on the cost of high-end new energy vehicles, with a total impact of nearly 10,000 yuan.

Previously, the intelligent driving domain controllers (ECUs) of domestic automakers generally needed to be equipped with 32GB of memory. CLS reporters learned that NVIDIA's China engineering team has developed a lightweight open-source inference framework based on C++ language. Combined with FP4 quantization technology, it has assisted some domestic automakers to reduce the domain controller memory from 32GB to 16GB without reducing driving functions and accuracy.

In addition to reducing the memory load, domestic automakers are also turning their attention to peripheral sensors for cost reduction. Some domestic automakers proposed to directly remove the processing chip inside the millimeter-wave radar module, and send the unprocessed original signals directly to the main chip. To this end, NVIDIA's China team modified the underlying software middleware, and called the idle programmable visual accelerator inside the main chip to undertake the radar original signal operation, so as to save the cost of the radar chip. It is reported that this cost reduction solution derived from Chinese automakers has now been merged into NVIDIA's global standard software version.

However, Wu Peidong told CLS reporters that optimizations such as reducing memory and saving radar chips are indeed effective, but they can only delay the pace of automakers switching to self-developed or local chips by one to two generations of models, and cannot change the general direction. In his view, what NVIDIA can retain is the group of automakers that want to develop self-owned chips but find it not cost-effective after calculation, for a period of about two to three years; NVIDIA's own investment has also been tilted towards Robotaxi, overseas automakers and L4-level unmanned driving.

Extending from L4 Unmanned Driving to Robots

After the large model is deployed on the vehicle side and the hardware cost is reduced through engineering optimization, truly allowing Physical AI to take over the steering wheel on public roads still faces the last threshold of safety and regulations.

The mandatory national standard "Safety Requirements for Autonomous Driving Systems of Intelligent Connected Vehicles" announced in August this year is scheduled to be implemented on July 1, 2027. It applies to passenger and cargo vehicles equipped with conditional autonomous driving (L3) and highly autonomous driving (L4) systems, among which the L3 system should also have the function of monitoring the driver's takeover capability.

At present, the safety responsible subject of the rapidly popularized high-level assisted driving (L2++) in China is still the human driver, and a single main chip is usually configured on the vehicle side; once entering the L3 and L4 stages, the safety responsible subject begins to shift to the system itself.

Relevant NVIDIA staff told CLS reporters that there are many debates in the industry about L3. L3 requires the human driver to take over the vehicle within a few seconds, but distracted drivers can hardly recover road condition perception in a short time; in those few seconds before the takeover, the requirements of L3 for hardware safety redundancy are almost no different from those of L4.

Public information shows that since the beginning of this year, Mercedes-Benz and BMW have cancelled the L3 system on their facelifted S-Class and facelifted 7 Series respectively, and turned to L2++ level assisted driving. BMW gave the reason that the market demand is insufficient and the benefits for customers are limited; Stellantis also shelved the AutoDrive L3 project in 2025.

Different from the automakers' move to shrink L3, NVIDIA's hardware reference architecture is designed with L4 as the goal. In order to meet the requirement of "fail-operational" for unmanned driving, it is equipped with two independent power supplies and sensors, as well as a dual-Thor chip master-slave backup system.

According to the design, under normal driving conditions, the two chips can run at the same time or be controlled by the main chip. Once the main chip or the corresponding sensor fails, the backup chip and independent sensors can take over the vehicle. At the algorithm level, relevant NVIDIA staff introduced that some automakers currently use the open Alpamayo or Cosmos inference model as an "examiner" to evaluate whether the driving behavior is reasonable.

In terms of the underlying operating system, overseas automakers generally choose the QNX real-time operating system that has passed safety certification but has a high development threshold, while domestic automakers prefer to use the Linux system. It is reported that in the full-stack safety system (Halos), NVIDIA has ported the safety diagnosis service on QNX to the Linux system for domestic automakers to use as a transition.

In NVIDIA's overall plan, this set of software and hardware architecture from cloud to vehicle side will eventually be extended to the field of embodied intelligent robots. From the perspective of underlying architecture, intelligent vehicles and robots are highly homologous at the Physical AI level.

Specifically, NVIDIA's solution allows intelligent vehicles and robots to share a set of world foundational models on the cloud; on the end side, whether it is an autonomous driving car or a humanoid robot, the underlying Thor chips and unified inference software development kit are used, the only difference is that the upper-layer algorithm finally controls "four wheels" or "hands and feet". By running through the closed loop from cloud training, virtual simulation to end-side safety inference on the mass-produced intelligent vehicles first, NVIDIA is trying to replicate this Physical AI solution to the robot market.

However, not all capabilities can be directly reused from vehicles to robots.

Wu Peidong said that the industry's relatively consistent view is that 60% to 70% of the hardware can be reused, but the remaining 30% to 40% is precisely the most expensive part, and links such as dexterous hands, high-degree-of-freedom joints, force tactile and motion control cannot be migrated. In his view, the logic of "vehicles supporting robots" holds in terms of manufacturing and cost, but does not hold in terms of capabilities and data.

Meanwhile, a report released by market research institution Counterpoint Research in October this year said that NVIDIA currently dominates the humanoid robot GPU market, while Tesla and XPeng are expected to change the market pattern of humanoid robot AI chips relying on self-developed chips and proprietary AI models. Dige Robot, spun off from Horizon Robotics, has reached cooperation with manufacturers such as Leju Robot and Extant Dynamics, and Rockchip (603893.SH) also said that its co-processor RK182X multi-product line has entered the stage of product release and project mass production. When NVIDIA brings this set of solutions to the robot market, it still faces competition from automakers' self-development and local chip manufacturers.

This article is from the WeChat Official Account "CLS", Author: Wang Biwei, published with authorization by 36Kr.