The supply model for AI computing power is at a turning point of redefinition. Over the past decade, the discrete architecture of CPUs paired with independent accelerator cards has been almost the only solution for data center inference. However, the repeated data movement between memory and VRAM, the compounded power and cooling costs during deployment, and the exponential growth of failure points during operations and maintenance have exposed the undeniable fatigue of this architecture in the era of AI large models. A computing power utilization rate of less than 40% is not due to insufficient computing power, but rather too many "walls" — the memory wall, the interconnect wall, and the cost wall.
Figure | The Truth Behind Computing Power Utilization Rate of Less Than 40%; Source: Lanxin Computing
At the 5th Dishui Lake China RISC-V Industry Forum held today, Xu Qingwei, Head of Ecosystem Marketing at Lanxin Computing, introduced a mass-produced chip — the LX500.
Figure | Xu Qingwei, Head of Ecosystem Marketing at Lanxin Computing
Xu Qingwei believes that the uniqueness of this chip lies not in merely stacking 48 cores or 75 TOPS of INT8 computing power, but in its heterogeneous design featuring the convergence of AI and general-purpose computing, which natively integrates general-purpose computing and AI inference onto a single chip. Unified memory management, along with support for DDR5 and CXL 2.0 to expand the addressing space to 2TB, means that large-parameter models like Kimi and Qwen no longer need to transfer data across cards; a single chip can handle hundred-billion-level inference workloads. More importantly, it addresses the long-questioned shortcomings of stability and security in the RISC-V architecture by incorporating server-grade RAS (Reliability, Availability, and Serviceability) features, support for Chinese national cryptographic algorithms, and DVFS (Dynamic Voltage and Frequency Scaling) power management.
Xu Qingwei stated that the LX500 has made significant simplifications in its architecture, with the goal of translating them into efficiency gains during actual deployment. Therefore, we see that the AI agent all-in-one machine based on the LX500 directly bypasses the procurement and adaptation of discrete inference cards, achieving "out-of-the-box" usability. For enterprises with limited budgets, weak O&M capabilities, yet facing data privacy pressures, this represents a practical and feasible path for AI implementation.
Figure | LX500 Included in Lenovo Wentian WR5219 G5 Product Line and Officially on Sale; Source: Lanxin Computing
On the industrial level, the commercialization progress of the LX500 has already stepped out of the laboratory or sample stage. In July 2026, in collaboration with Lenovo, the world's first mass-produced RISC-V server rolled off the production line at the Hefei LCFC factory, and was included in the Lenovo Wentian WR5219 G5 product line for official sale. It has already been deployed in core business adaptation for telecom operators, hosted deployment in data centers of central and state-owned enterprises, intelligent risk control and quantitative trading scenarios in the financial industry, and secure simulation platforms for university research.