Li Auto showcased a motherboard equipped with its in-house developed chip at the World Artificial Intelligence Conference in July 2026. Credit: Li Auto
- Li Auto is exploring an in-house cloud inference chip, and the project is still at an early stage.
- Cloud inference chips can take on some of the inference workloads that currently run on GPUs.
Li Auto (NASDAQ: LI) is exploring an in-house cloud inference chip, according to a report on Friday by local media outlet LatePost.
The chip will use the same dataflow architecture as its assisted driving chip, and the project is still at an early stage, the report said, citing multiple sources.
AI (artificial intelligence) data centers generally use GPUs for model training and inference, though some also deploy dedicated cloud inference chips, the report noted.
Cloud inference chips can take on some of the inference workloads that currently run on GPUs, a person in the chip industry told LatePost.
These include data processing, testing and simulation for assisted driving models, as well as request handling for large language models. Training steps that involve parameter updates still require training chips.
Taking the dataflow architecture from the vehicle to the cloud is technically feasible, industry insiders told the outlet.
One path is to reuse the compute design of the vehicle-side NPU (AI processor), packaging multiple AI compute dies together. Adding high-bandwidth memory and high-speed interconnects then makes it possible to build a larger inference chip.
That would allow the vehicle and cloud sides to share some designs and software tools, spreading out R&D costs.
But a cloud chip is not simply a scaled-up version of a vehicle chip. On the vehicle side, models, sensor inputs and operating rhythms are relatively fixed, with the main goals being low power consumption, low latency and stable execution.
The cloud, by contrast, needs to handle more models at the same time, with faster model updates and greater fluctuations in input length and concurrent requests.
It also involves high-bandwidth memory, multi-chip interconnects, dynamic batching and cluster scheduling.
Hitting a certain performance target with a single chip is not the hardest part — the real difficulty is making the overall inference cost competitive, the person said.
Whether extending the dataflow architecture from the vehicle to the cloud can create a cost advantage depends on model adaptation capabilities and system operating efficiency.
Companies including SambaNova, Groq and Tenstorrent are also exploring dataflow architectures, LatePost said.
The core teams at SambaNova and Groq come from Stanford and Google’s TPU (tensor processing unit) project respectively, while Tenstorrent is led by Jim Keller, who previously headed Tesla’s self-driving chip work.
Compared with GPUs, dataflow architectures still need to prove themselves in model versatility, software ecosystems and large-scale deployment.
As the cloud project moves forward, Li Auto’s head of chip software R&D, Jin Yihua, and Dai Jie, head of one of its chip front-end design groups, have left the company, LatePost confirmed through multiple channels.
Both previously reported to Luo Min, head of the computing power unit, who in turn reports to group CTO Xie Yan.
It is unclear whether the personnel changes will affect the progress of the cloud inference chip project.
Li Auto officially unveiled its in-house Mach M100 assisted driving chip on May 12, built on a 5nm automotive-grade process with 1,280 TOPS of computing power per chip.
The chip is now in mass production in the all-new Li L9, L8 and L6, with the dual-chip version offering a combined 2,560 TOPS.
The company’s founder, chairman and CEO Li Xiang had said that the in-house chip was not about proving technical capability, but about making AI actually work in the physical world.
Li Auto registered a new company, Xinchuang Zhihe (Shanghai) Technology Co Ltd, on July 13, with a business scope that includes integrated circuit chip design.
That could signal the company is considering running its semiconductor business independently, following the path of peer Nio Inc (NYSE: NIO).
Nio set up chip subsidiary GeniTech Co Ltd (Shenji) in June 2025, which has since raised nearly 3 billion yuan ($442 million) at a post-money valuation of about 8.27 billion yuan.
Xpeng (NYSE: XPEV) has also put its in-house Turing assisted driving chip into its vehicles.
Morgan Stanley says GeniTech’s debut at WAIC 2026 takes Nio’s stock narrative one step further toward a vertically integrated AI chip platform.
($1 = 6.7878 yuan)
