Li Auto Explores Self-Developed Cloud Inference Chip as of August 2026

As of August 2026, Li Auto is exploring the self-development of cloud inference chips, with the project at an early stage. In response, Li Auto said it had not heard of such a move. Cloud inference chips are essentially AI accelerator chips deployed in data centers or cloud servers. AI models consist of two stages: training and inference. Inference involves operations such as answering each query, recognizing images, or generating plans. Cloud inference chips are designed to enhance the speed, stability, and cost efficiency of these AI calls. The agency predicts that this proportion will rise to 59% by 2027. In the intelligent vehicle sector, cloud inference chips are not directly used for real-time driving decisions; perception, planning, and control during vehicle operation rely primarily on in-vehicle computing power. Cloud inference capabilities serve more in model iteration, data loops, and intelligent service systems, such as batch recognition and filtering of road data, retrieval of long-tail scenarios, automatic annotation, model invocations in simulation evaluation, as well as in-car intelligent assistants and cabin large models. Li Auto has already made public moves in in-vehicle AI inference chips, with its self-developed Mach M100 chip using a dataflow architecture and positioned as an AI inference chip. Cloud inference chips have different requirements from vehicle-mounted chips; the cloud emphasizes large-scale data center deployment, high-concurrency throughput, cluster efficiency, and per-unit inference cost.