On the evening of January 15, XSKY Xingchen Tianhe announced the launch of its full-stack AI data solution specifically designed for AI scenarios. The company stated that through its three core products—MeshFS, MeshSpace, and MeshFusion—it aims to break down the IO wall, data gravity wall, and memory wall that constrain AI efficiency, thereby helping enterprises build highly efficient and controllable AI factories.
In the era of large models, the trend of algorithm homogenization is becoming increasingly evident. The genuine differentiated competitive advantage for enterprises lies hidden within their unique "proprietary data." However, these digital assets, which carry a company's core experience and value, cannot be externalized due to security and compliance requirements. The critical challenge for enterprise AI transformation is how to realize value from this data within a private environment.
Furthermore, traditional storage architectures struggle to support the efficient transformation of private, high-value data. This directly leads to GPU utilization rates being dragged down to 30-50% by I/O wait times in numerous AI training and data engineering scenarios, and even lower in extreme cases. The resulting idle computing power also means storage read/write speeds cannot keep up with computational throughput (the IO wall), the explosive growth of model parameters is constrained (the memory wall), and the cost of cross-domain data movement remains prohibitively high (the gravity wall), creating new data silos.
How to maintain data security while simultaneously breaking through these efficiency bottlenecks has become a core demand for the practical implementation of enterprise AI. Recognizing this trend, XSKY is leveraging its full-stack AI data solution to fortify the baseline of data security and establish a private, controllable AI data foundation, enabling enterprises to safely convert their proprietary data into intelligence within their own environments.
Reportedly, XSKY achieves this through an architectural innovation—proposing the AIMesh data and memory network. This innovative architecture unifies three networks: the training data network (MeshFS), the global object network (MeshSpace), and the inference memory network (MeshFusion). This "three-in-one" approach precisely targets and breaks through the three major pain points—the IO wall, gravity wall, and memory wall—delivering core value characterized by "speed, connectivity, and cost-efficiency."
To address the bottleneck where computational throughput far exceeds storage read/write speeds, leading to low GPU utilization, MeshFS integrates the mature POSIX semantics of XGFS with the extreme performance of the XSEA all-flash array base. It supports full-protocol interoperability (POSIX/S3/HDFS) and can run without modifying Python or TensorFlow training code. Its fully distributed architecture allows performance to scale linearly with the number of nodes, with metadata processing latency as low as microseconds. Tests show a 30% improvement in sequential read bandwidth and a 50% improvement in sequential write bandwidth compared to similar products, effectively solving the problem of lagging data supply.
Addressing the issue where ultra-long contexts in AI inference lead to heavy occupation of the KV Cache and the continuously rising cost of GPU memory (HBM), MeshFusion transforms server-local NVMe SSDs into L3-level external memory. This achieves a near-infinite context window at just 1% of the hardware cost. Tests indicate that its performance gap compared to pure DRAM is controlled within 10%, with throughput scaling linearly in high-concurrency scenarios. Under resource-constrained conditions, it can even achieve a 20% performance advantage.
Comments