Li Auto is expanding chip design ambitions from assisted driving to on-vehicle and data center AI inference, signaling Chinese EV makers' vertical integration into compute.
Li Auto is exploring an in-house cloud inference chip, according to a report by Chinese media outlet LatePost. The chip will use the same dataflow architecture as Li Auto's assisted driving chip, though the project remains in an early stage.
Cloud inference chips can handle some of the inference workloads that currently run on GPUs in AI data centers. These tasks include data processing, testing and simulation for assisted driving models, as well as request handling for large language models. Training steps requiring parameter updates still need dedicated training chips.
Industry insiders confirmed that reusing the dataflow architecture from vehicles to the cloud is technically feasible. One approach involves packaging multiple AI compute dies from the vehicle-side NPU (AI processor) together, then adding high-bandwidth memory and high-speed interconnects to build a larger inference chip. This would allow vehicles and cloud infrastructure to share design and software tools, spreading R&D costs.
However, a cloud chip is not simply a scaled-up version of a vehicle chip. On the vehicle side, models and sensor inputs are relatively fixed, with priorities on low power consumption, low latency and stable execution. The cloud must handle multiple models simultaneously, accommodate frequent model updates, and manage variable input lengths and concurrent requests. This requires high-bandwidth memory, multi-chip interconnects, dynamic batching and cluster scheduling. The core challenge is not achieving a certain performance target with a single chip, but making overall inference cost competitive through effective model adaptation and system efficiency.
Companies including SambaNova, Groq and Tenstorrent are also exploring dataflow architectures. SambaNova and Groq's core teams originated from Stanford and Google's TPU project respectively, while Tenstorrent is led by Jim Keller, who previously headed Tesla's self-driving chip development. Compared with GPUs, dataflow architectures still need to prove themselves in model versatility, software ecosystems and large-scale deployment.
During the cloud project's advancement, Jin Yihua, Li Auto's head of chip software R&D, and Dai Jie, head of one of its chip front-end design groups, have left the company. Both previously reported to Luo Min, head of the computing power unit, who reports to group CTO Xie Yan. The impact of these personnel changes on the cloud inference chip project remains unclear.
Li Auto officially unveiled its in-house Mach M100 assisted driving chip in May 2026, built on a 5nm automotive-grade process with 1,280 TOPS of computing power per chip. The chip is now in mass production across the all-new Li L9, L8 and L6 models, with the dual-chip version providing 2,560 TOPS combined. Founder, chairman and CEO Li Xiang emphasized that the in-house chip was not about proving technical capability, but about making AI work in the physical world.
Li Auto registered a new company, Xinchuang Zhihe (Shanghai) Technology Co Ltd, on July 13 with a business scope including integrated circuit chip design. This could indicate the company is considering operating its semiconductor business independently, following the path of peer Nio. Nio established chip subsidiary GeniTech Co Ltd in June 2025, which has since raised nearly 3 billion yuan ($442 million) at a post-money valuation of approximately 8.27 billion yuan. Xpeng has also deployed its in-house Turing assisted driving chip in its vehicles.