How XSKY AIMesh 2026 Redefines AI Data Infrastructure for End‑to‑End Flow

Enterprises adopting large‑model AI now find that competitive advantage stems from proprietary data rather than raw compute, yet most struggle to move massive private datasets into training and inference pipelines; traditional monolithic storage cannot meet the varied access patterns of AI workloads.

AI Data Layering Concept

Data is classified by lifecycle and access frequency into five tiers: G1/G2 (runtime data in GPU/NPU HBM/DRAM), G3/G3.5 (near‑edge hot data on local NVMe SSDs), G4 (shared production data for training and inference), and G5 (enterprise‑wide data assets such as documents, media, logs, and RAG corpora). Each tier has distinct storage responsibilities.

G3/G3.5 – MeshFusion Breaks the Memory Wall

MeshFusion pools local NVMe SSDs into a shared KV‑Cache pool, offloading expensive HBM and dramatically lowering inference cost. Two deployment modes are offered: an integrated mode that reuses existing GPU‑server NVMe without extra storage nodes, and a separated mode that decouples compute and storage for massive clusters, matching NVIDIA CMX architectures. The native KV‑Cache interface lets applications access the pool without modifying inference frameworks, enabling cross‑node cache sharing and higher concurrency.

G4 – MeshFS 7.4 Opens the I/O Wall

MeshFS is a high‑performance parallel file system targeting the G4 layer. Version 7.4 improves read bandwidth to remain industry‑leading and raises write bandwidth by 24 % over version 7.3, stabilising checkpoint writes under heavy training loads. A distributed parallel gateway spreads NFS traffic across the whole cluster, boosting client read/write throughput by 3.8× and further increasing NFS read performance by 7.1× through intelligent tiering and prefetching. MeshFS also supports Intel, AMD, and Kunpeng CPUs, Infiniband/RoCE networks, and complies with domestic‑innovation (信创) environments.

G5 – MeshSpace Dissolves the Gravity Wall

MeshSpace provides a unified logical namespace that aggregates data across buckets, clusters, and regions into an EB‑scale object store. A two‑layer metadata design enables a single bucket to hold exabytes; a real‑world deployment shows a 100 PB bucket delivering 2.5 Tbps write, 3 Tbps read, and 4 KB read OPS of 4 million objects per second. Low‑cost long‑term archiving is achieved by IceLake integration, which aggregates small files and writes them sequentially to cheap media while keeping them searchable and retrievable for audit or re‑analysis.

MeshSpace also offers lake‑warehouse integration via Tables: raw objects are stored via S3, while datasets are organised as Iceberg tables, allowing Spark, Flink, Trino, and other engines to consume data without migration. Cross‑domain data scheduling automatically moves data between hot, warm, and cold tiers based on workload needs, eliminating manual data movement.

Three‑Product Synergy

MeshFusion, MeshFS, and MeshSpace are not isolated systems; they align along the AI data usage chain to form a closed‑loop pipeline. During inference, MeshFusion supplies near‑edge hot data cache; during model training, MeshFS delivers high‑throughput shared file access; and throughout the lifecycle, MeshSpace governs historical and raw assets, feeding valuable data back to training or inference as needed.

This layered architecture simultaneously addresses the memory wall, I/O wall, and gravity wall, unlocking the value of proprietary data for both small AI pilots and large‑scale intelligent‑computing clusters.