On September 17, Xinzhansu officially announced the rollout of its "Cloud-Edge-Terminal" full-spectrum AI strategic layout. During the event, Chairman and CEO Sun Dan pointed out that while GPU computing power has surged nearly 60,000-fold over the past two decades, memory bandwidth and capacity growth has remained almost linear. Trapped by this "memory wall," the mainstream solutions appear to be limited to KV Cache compression and HBM stacking. In response, Xinzhansu has chosen to forge a third, innovative path—letting storage participate in computation to strengthen computing power through storage capacity.
Sun Dan elaborated on the strategy of "empowering compute with storage for full-spectrum coverage," describing the construction of a three-tier data foundation spanning cloud, edge, and terminal. This includes cloud-based AI SSDs to enhance overall cluster throughput and stability, an edge "Token Factory" for real-time inference, and localized AI capabilities at the terminal level. Xinzhansu Vice President Li Zhen, alongside VAST Data's China Pre-Sales Director Liu Chenghong, jointly unveiled Xinzhansu's full-media stack solution for cloud AI storage, which has already received official certification from VAST Data.
Liu Chenghong noted that a key direction in AI storage evolution is moving toward a non-layered, unified data platform, where upper-layer applications see one unified system and one unified dataset rather than multiple isolated tiers. He emphasized that the ultra-high endurance of the Xinzhansu PT200Z AI SSD and the massive 122.88TB capacity of the MQ series are purpose-built for cloud storage demands that favor a non-layered architecture. Li Zhen added that the real need of cloud storage customers such as VAST Data is precisely this layered-free architecture, which aligns with Xinzhansu's core strategic positioning: a combination of SCM, high-density QLC, and high-performance TLC.
Xinzhansu Vice President Xu Wei and Senior Director Yang Zehong also presented the AI90 training and inference acceleration engine. Xu Wei detailed the technical implementation of the AI90 Token Factory, explaining that through intelligent KV Cache scheduling and smart P2P interconnection, the engine offloads KV Cache from HBM to high-performance AI SSDs. This establishes a three-tier HBM+DRAM+SSD storage system, enabling limited video memory to handle unlimited context and allowing consumer-grade hardware to achieve enterprise-level performance. Yang Zehong demonstrated with a Live Demo, showing the AI90 raising inference throughput from roughly 120k tokens per second to 610k tokens per second, while first-token latency dropped from 27.99 seconds to 0.56 seconds—a five-fold improvement in throughput and a 50-fold faster response time.
From a partner’s perspective, Haiwei Zhisuan's CTO Liu Chao affirmed the AI90's practical results, noting that on an AMD-AI90 integrated system running the Qwen 72B model, the per-token cost is only about one-67th of comparable cloud API pricing. Flowy CEO Qi Guoliang shared insights into AI PC technology, stressing that the AI30 turns terminal devices into truly independent AI compute nodes, effectively paving the final mile for edge-side AI. On the same day, Xinzhansu Partner and COO Qian Ya'ou signed strategic cooperation agreements with Guangdong Summit, Haiwei Zhisuan, Yun Gongchang, and Flowy. These collaborations span three key scenarios—intelligent computing centers, cloud services, and terminal AI—officially completing the cloud-edge-terminal ecosystem loop.