Not Just Building Cars, But Also Cloud Chips? Li Auto Reportedly Exploring In-House Inference Chips

Li Auto (LI-US) is extending its in-house chip development from vehicle-side to cloud-side. According to LatePost Auto, citing multiple industry insiders, Li Auto is currently exploring the development of cloud inference chips, with initial plans to adopt the dataflow architecture used in its smart driving chips. However, the entire project remains in its early stages.

If the plan progresses smoothly, Li Auto could eventually use its in-house chips to handle some of the cloud inference tasks currently performed by GPUs. These tasks include data processing, testing, and simulation for smart driving models, as well as request handling for large language models. However, model training involving parameter updates will still be handled by specialized training chips.

Extending Vehicle-Side Chip Design to Reduce R&D Costs

Insiders indicate that extending the dataflow architecture from vehicle-side to cloud-side is technically feasible. One approach involves reusing the computational design of the onboard NPU (Neural Processing Unit), integrating multiple AI compute dies into a single package, and combining them with high-bandwidth memory (HBM) and high-speed interconnect technologies to form a larger-scale cloud inference chip.

This design allows vehicle-side and cloud-side chips to share parts of the hardware architecture, software tools, and development experience, thereby distributing R&D costs. For Li Auto, this also offers the potential to improve integration efficiency from cloud-based development and testing to on-vehicle deployment of smart driving models.

However, cloud chips are not simply scaled-up versions of vehicle chips. Vehicle systems typically handle relatively fixed models and sensor data with stable operational rhythms, focusing design on low power consumption, low latency, and reliable execution.

In contrast, cloud systems must simultaneously support multiple models and respond to frequent updates, variable input lengths, and drastic fluctuations in request volume. This involves complex technologies such as HBM, multi-die interconnects, dynamic batching, and cluster scheduling.

Single-Chip Performance Is Not the Only Key

A chip industry insider noted that creating a single chip that meets specific performance metrics is not the most difficult part. The real challenge lies in whether the entire inference system can achieve cost competitiveness.

Therefore, whether the dataflow architecture can bring actual benefits to Li Auto when transitioning from vehicle-side to cloud-side depends on the chip's ability to support different models, software-hardware co-optimization efficiency, and overall system operating costs.

Overseas, several chip companies are also exploring dataflow architectures, including SambaNova, Groq, and Tenstorrent. Core members of SambaNova and Groq have backgrounds from Stanford University and Google's TPU project, respectively. Tenstorrent is led by Jim Keller, who previously oversaw Tesla's autonomous driving chip development.

However, compared to GPUs, dataflow architectures still require further validation in terms of model generality, software ecosystem maturity, and large-scale deployment capabilities.

Two Chip R&D Leads Depart—Impact on Project Timeline Remains Unclear

As Li Auto explores cloud inference chips, personnel changes have also emerged within its chip team. LatePost Auto confirmed through multiple sources that Jin Yihua, head of chip software development at Li Auto, and Dai Jie, head of the first chip front-end design team, have both left the company.

Both previously reported to Luo Min, head of the computing unit, who in turn reports to Li Auto Group CTO Xie Yan.

The reasons for the departures are unclear, and it remains uncertain whether these personnel changes will impact the development timeline of the cloud inference chip project. Since the project is still in its early stages, it remains uncertain whether Li Auto will formally initiate the project, proceed to mass production, or when it might actually launch.