Pony.ai prepares domestic chip options for autonomous driving

Chinese autonomous-vehicle company Pony.ai is preparing alternatives to Nvidia computing hardware as it manages supply constraints and the prospect of tighter export controls. Chief Financial Officer Leo Wang said demand for Nvidia compute already exceeds supply, creating a pain point for the company’s work on autonomous-driving technology.
Pony.ai’s response has two parts: engineers are being asked to use training resources as efficiently as possible, and the company is increasing its use of chips produced in China. Wang said a deployment may use midtier computing resources rather than high-end hardware when that is appropriate for the training workload.
A contingency plan for AI compute
Nvidia is a significant supplier of AI computing technology for the autonomous-vehicle sector. Companies use its automotive processors to run automated-driving software, its GPUs to train driving models, or both through broader technology partnerships. Mercedes-Benz, Zoox and Wayve are among the companies cited as having full-stack partnerships involving Nvidia.
For Chinese developers, the issue is not simply whether a single component can be replaced. Training and deployment choices must be adapted to the compute available. Wang said domestic chips have received increased attention and development in recent years, giving companies another option for AI training and more flexibility when preferred Nvidia hardware is constrained.
Alternatives are becoming more mature
Wang said Chinese companies have been preparing for this scenario for several years. If restrictions on Nvidia chip shipments to China become tighter, alternatives already exist, although he said they are probably not yet the best choice on cost-effectiveness grounds. Their maturity, however, has improved compared with a few years ago.
The wider mobility sector remains heavily shaped by compute choices. As autonomous vehicle development spans software and platforms illustrates, autonomous-vehicle development spans software, vehicle platforms and operating models, while access to suitable AI infrastructure influences how quickly teams can train and deploy their systems.
What businesses should assess
The immediate implication is that autonomous-driving companies should identify where their model training and in-vehicle software depend on scarce hardware. They can then evaluate whether more efficient deployment, midtier compute or domestic chips can support defined workloads, while accounting for the cost trade-offs Wang described.
For businesses building AI-intensive transport products, a practical contingency plan is to test alternative compute paths before supply conditions require a rapid change in training or deployment decisions.

