Alibaba Cloud Launches Agentic Storage Matrix for AI & Agent Workloads

Alibaba Cloud unveiled its full-stack Agentic Storage product matrix at the 2026 Yunqi Conference, targeting both AI Native and Agent Native workloads. The matrix covers the entire data path from underlying infrastructure to application-layer agents.

AI Native Storage: From Million-GPU Training to Millisecond Inference

Training: Next-Generation CPFS

The new fully self-developed CPFS is built on the Feitian Pangu foundation and adopts a metadata-data separation scale-out architecture designed for next-generation AI training clusters. It delivers up to 100 TB/s throughput and 100 million IOPS , with a single file system scaling to 100 PiB (5× increase). In real model training, it reduces average model startup time by 50% , boosts peak GPU utilization by 30% , cuts AI storage costs by 69% , and doubles business capacity — shifting GPUs from "waiting for data" to continuous efficient compute.

Customers already running this CPFS in production include XPeng Motors , Black Sesame Technologies , and Aishi Technology , supporting autonomous driving model training, chip R&D with large-scale simulation verification, and multimodal model training.

Inference: KVCacheStore

To handle exploding KV cache capacity and multi-turn tasks, Alibaba Cloud launched KVCacheStore — the world's first public-cloud G3.5-layer KV cache storage product. It supports trillion-scale KV storage and in customer production environments achieves a 900% expansion of cache coverage time window , 20% throughput increase , and 54% reduction in first-token latency .

Multimodal Data: Unified OSS Foundation

OSS now provides a unified multimodal data storage foundation covering unstructured, structured, and vector data. Following OSS Vectors , the new OSS Tables (based on open Iceberg and Lance formats) delivers serverless data lake capabilities for AI training data and traditional OLAP analytics, enabling unified permissions, access, and management across compute engines.

Agent Native Storage: Data Infrastructure for Million-Agent Concurrency

AgenticFS: Industry-First File Storage for Agents

AgenticFS is the industry's first file storage product purpose-built for agent scenarios. It supports a single file system of 100 PiB and trillion files , dynamic creation of million AgenticSpaces , and 100k QPS mount/unmount (1000× traditional file storage). It provides triple isolation — capacity, performance, and secure access — delivering a large-scale, secure, elastically scalable storage environment for sandboxes. Qwen Office and MiniMax are already running on it.

Agentic DiskPool & Agentic Bucket

Agentic DiskPool upgrades individual cloud disks into pooled resources, solving the "boot storm" problem in million-sandbox concurrent scenarios and guaranteeing second-level sandbox start/stop with instant state resume.

Agentic Bucket offers unlimited scaling for massive video and image data, supporting 100k concurrent agent creations per service account with 100k QPS and second-level elasticity. A leading domestic large-model company has adopted it for agent services.

Memory & Knowledge Management

OSS Vectors provides low-cost, large-scale vector storage serving hundreds of customers.

Tablestore Agentic Memory is upgraded to a native context service fusing knowledge and memory, adding capabilities like memory consolidation, reflection, and Wiki knowledge compilation to help agents continuously accumulate and self-evolve.

Data Services, Protection & Agentic Interfaces

OSS AI VibeStudio and OSS AI VibeFlow handle data processing and vectorization orchestration, turning raw data into training- and inference-ready datasets.

Cloud Backup and Data Disaster Recovery Center provide unified data protection and disaster recovery for cloud and on-premises data.

Agentic Enterprise Cloud Drive supports enterprise file collaboration and knowledge management.

OSS Agent , CDE Agent , and Data Protection Agent encapsulate storage, data, and protection capabilities as conversable, orchestratable intelligent agents, lowering the barrier to using data infrastructure.

Conclusion

This full-chain matrix marks Alibaba Cloud Storage's completion of an end-to-end layout spanning training to inference, and from Agent Data Infra to Data Agent — establishing a unified data storage foundation for the Agentic era where compute efficiently reaches data and intelligence continuously grows within data. As of August 2026, Alibaba Cloud Storage has served millions of enterprise customers across frontier scenarios including LLM training, autonomous driving, embodied intelligence, and scientific simulation.