July 22, 2026, Beijing - The 18th International High-Performance Computing and Artificial Intelligence Advisory Council (HPC AI Advisory Council) China Annual Meeting was held grandly. With the theme of "Toward AI Factory through AI Network", this annual conference brings together the world's top scholars, industry leaders and technical experts to discuss the road to building high-performance and efficient AI factories around AI network, computing, storage and mainstream software frameworks.

A profound industrial change

Currently, data centers are undergoing a comprehensive transformation from traditional general-purpose computing to AI factories. The business focus is gradually focusing on the training, reasoning and recommendation scenarios of AI models. With the exponential growth of the parameter scale of large models, distributed computing has extended from the training stage to the inference stage. The popularization of technologies such as PD separation and KVCache has further increased the demand for distributed computing and storage.

As an innovative pioneer in the field of AI storage and computing infrastructure, LUISUAN Technology was invited to attend this event. Mr. Kong Weihai, General Manager of LUISUAN Technology Products and R&D Center, delivered a keynote speech titled "When GPU Meets G3.5: Full-stack Collaboration and Future Vision of AI Infrastructure", sharing the cutting-edge thinking and practical results of LUISUAN Technology in full-stack collaborative innovation of AI infrastructure.

LUISUAN Technology was launched as the first domestic product to adapt to NVIDIA A manufacturer of commercial full-stack solutions based on the G3.5/CMX architecture reconstructs AI infrastructure through "computing and storage separation": using the GP7000 all-flash storage platform (JBOF) as the hardware cornerstone, relying on DPU+RDMA to build a direct connection channel for computing power and storage, achieving the ultimate performance of 20μs level latency, 300GB/s bandwidth and 70 million IOPS; combining Dynamo intelligent scheduling and SDS software-defined storage, KV Cache performs hot, warm and cold hierarchical management, while bypassing CPU/PCIe bottlenecks, replacing expensive HBM with NVMe SSD, increasing resource utilization from 30% to more than 70%, reducing TCO by 30%-50%, and using video memory virtualization to break through the capacity limit of long text reasoning for large models. Facing the future, this solution has laid out technologies such as CXL memory pooling, software hardening, and GPU direct-read flash memory. It is committed to realizing the ultimate architecture of "storage as memory" and promoting the intelligent computing center to move from the integration of storage and computing to the era of minimalist and efficient storage and computing separated computing power pools.

Explore the road to AI factory construction

The evolution of AI infrastructure is no longer a breakthrough in a single technology, but a full-stack collaboration. LUISUAN Technology will take this annual meeting as an opportunity to continue to delve into the field of AI storage and computing infrastructure, and work with industry partners to promote full-stack infrastructure innovation from data, storage, network to computing power, and contribute "LUISUAN power" to building an autonomous, controllable, extremely efficient next-generation AI factory.