Nvidia is offering its physical AI models and simulation tools to automakers for free, betting that the demand for cloud computing that follows will more than repay the investment. The company’s automotive revenue reached $2.349 billion in its fiscal 2026 year, which ended January 25, 2026, up 39% from the previous year.
Physical AI refers to machines that can perceive their surroundings, make judgments and act in the real world. After generative AI spread through text and images, technology companies turned to this field, and cars are among the largest targets. Most current driver-assistance systems convert camera images directly into steering and braking commands without understanding cause and effect. A vehicle may recognize a person at an intersection but cannot interpret a traffic officer’s hand signal to mean go or stop. Building models that understand physical laws is extremely expensive. Tesla’s Cortex 1 training cluster was equivalent to more than 100,000 Nvidia H100 GPUs as of early 2026, and its Cortex 2 cluster exceeds 130,000. Few automakers can afford that spending alone.
Nvidia’s answer is to train foundation models and release them to developers. Its portfolio includes Nemotron for pre-training, Cosmos, a world model that understands physical laws, GR00T for robots, and Alpamayo for autonomous driving. Automakers need only modest computing power and their own road-test data to fine-tune these models. Nvidia also offers the OVX platform for virtual simulation and a tool that converts ordinary road-test video into editable 3D scenes.
Free Tools Lower Barriers, Not Costs
Analysts say the free tools change the starting point but not the cost of development. Wu Peidong, managing director of Wanchuang Investment Bank, said the tools “change the starting point, not the finish line.” Developing a high-end driver-assistance system once took two to three years and several billion yuan. Teams can now begin with a working architecture and differentiate later, cutting development time by more than a year. However, spending on computing, data and calibration remains necessary.
Nvidia has also expanded its definition of automotive developers to include all companies that build their own software, including automakers and suppliers. The company argues that as customers build their own AI, they will use more of its platforms, which will sustain demand for cloud training and simulation servers.
Cost Pressure in China
China’s price-competitive market also demands cheaper hardware. Nvidia’s Orin chip is widely deployed, and its next-generation Thor chip is rated at 2,000 TFLOPS. Nvidia said that on Thor, 4-bit floating-point quantization can run models at 176 tokens per second, compared with 70 tokens per second for a 7-billion-parameter multimodal model on Orin. Many Chinese models still run on Orin, which does not natively support FP4 or FP8 formats.
Memory costs are also rising. Automotive-grade memory chip prices rose about 180% between March and June 2026. NIO chairman William Li said higher memory and raw material prices added CNY 3,000 to 5,000 to the cost of each high-end electric vehicle. To counter this, Nvidia’s China engineering team developed a lightweight C++ inference framework that, combined with FP4 quantization, helps some automakers cut domain controller memory from 32GB to 16GB without reducing driving functions. Some automakers also asked to remove the radar processing chip. Nvidia modified its middleware so the main chip’s vision accelerator handles raw radar signals, and the change has been added to its global software release.
Market Share Is Narrowing
Nvidia’s lead in China is shrinking. Its share of the passenger-car assisted-driving domain control chip market fell from 45.1% in January and February 2026 to 30.2% in August, when about 810,000 units were installed. Horizon Robotics ranked second with 19.8%, and Huawei third with 14.8%. The gap between Nvidia and Horizon narrowed from 35.8 percentage points to 10.4. NIO, XPeng and Li Auto are installing in-house chips, while Horizon is expanding in cost-effective solutions. Wu Peidong believes these optimizations can delay automakers’ shift to in-house chips by only one or two vehicle generations.
Regulation and Levels of Autonomy
China’s mandatory national standard for automated driving systems, published in August 2026, takes effect on July 1, 2027. It covers passenger and cargo vehicles with L3 and L4 systems. Nvidia‘s hardware reference design targets L4, with two independent power and sensor sets and dual Thor chips in a primary-backup arrangement. Mercedes-Benz and BMW have dropped L3 systems from recent models in favor of L2++ assisted driving, and Stellantis shelved its L3 project in 2025. Nvidia’s personnel have noted that L3 is difficult to make safe because a distracted driver may need several seconds to regain awareness of the road.
Extending Into Robotics
Nvidia plans to extend the approach to robots, using the Thor chip and a shared software development kit in both vehicles and humanoid robots. Wu Peidong estimates that 60% to 70% of the hardware can be reused, but the remaining 30% to 40%, including dexterous hands and force sensing, is the most expensive part. A Counterpoint Research report published in October 2026 said Nvidia dominates the humanoid robot GPU market, but Tesla and XPeng, which rely on in-house chips, could reshape competition. Horizon spinoff Digen Robot and Rockchip are also pursuing the market.
