Aliyun OSS
Alibaba Cloud's S3-compatible Object Storage Service — the dominant object store across mainland China. Standard bucket/key data model with regional clusters spanning the East Data West Computing corridor (coastal hot tiers, western cold/training tiers).
Summary
Alibaba Cloud's S3-compatible Object Storage Service — the dominant object store across mainland China. Standard bucket/key data model with regional clusters spanning the East Data West Computing corridor (coastal hot tiers, western cold/training tiers).
Aliyun OSS is the silent majority of Chinese AI training storage in 2026 — the substrate for DeepSeek, Zhipu AI, Moonshot AI, and Alibaba's own Qwen training runs. From the index's perspective it's the China-side counterpart to AWS S3: same role, different jurisdiction, mutually unreachable.
- "S3-compatible" coverage is good for read/write/multipart/lifecycle but not feature-complete with AWS extensions (S3 Tables, Vectors, Metadata, Files have no Aliyun equivalents at parity).
- Aliyun's international regions exist but are secondary; PRC-domestic regions are the architectural focus and the only ones that meet China Data Localization for PRC-citizen data.
- Cross-cloud replication to/from non-Alibaba providers is regulatorily fraught — a CAC review may be required for export of "important data."
- AWS SDK v2's default
STREAMING-UNSIGNED-PAYLOAD-TRAILERchunked encoding does not work against OSS — Apache Iceberg and Polaris must be configured to fall back toSTREAMING-AWS4-HMAC-SHA256-PAYLOADor to disable chunked encoding entirely. This bites teams porting v2-SDK lakehouse stacks to OSS without changing their default request signing. - The
ossfsFUSE driver looks like a POSIX mount but lacks hard links, extended attributes, and robust file locking. Use CPFS for any metadata-heavy or training-pipeline mount; reserveossfsfor sequential-read or admin scripting. - April 2026 brought a +4.6 to +5.6% price hike on infrequent-access and archive tiers — Standard tier held flat. Storage budgets built on pre-hike prices need a refresh; Standard is now the most cost-stable tier for hot data.
Pricing posture: Five tiers from Standard ~$0.017/GB-month down to Deep Cold Archive ~$0.0011/GB-month — the lowest archival tier on the public market.
AI ecosystem position: OSS is the storage substrate for Panjiu AI Infra 2.0. CPFS layers POSIX semantics on top with sub-1ms metadata; HPN 8.0 RDMA networking removes the bottleneck for distributed training. Qwen 3 was trained on this exact stack, which is part of why Aliyun's hyperscaler position in China is reinforcing rather than declining.
implementsS3 APIenablesEast Data West Computing — the storage substrate of the strategysolvesChina Data Localizationscoped_toSovereign Storage
Definition
Alibaba Cloud's S3-compatible object storage service (Object Storage Service), the dominant object store across mainland China. Provides the bucket/key/object data model and S3-API compatibility, with regional clusters explicitly placed inside the **East Data West Computing** corridor — Beijing/Shanghai/Shenzhen for hot reads, Guizhou and Inner Mongolia for cheap-power compute and bulk training storage. Native integration with Alibaba's PolarDB AI Lakehouse, Hologres, and the Qwen training pipeline.
US export controls and the Cybersecurity Law of the PRC together force Chinese AI workloads onto domestic infrastructure that cannot route through AWS S3. Aliyun OSS is the primary S3 implementation that DeepSeek, Zhipu AI, Moonshot AI, and Alibaba's own Qwen team train against — making it the silent majority of open-weight model training storage in 2026.
S3-compatible storage for Chinese-domiciled AI training, multi-region object storage inside mainland China, Alibaba Cloud-native lakehouse and ML pipelines, regulated workloads requiring data residency under PRC law.
Recent developments
- The strategic reframe — OSS as the AI-native foundation, backed by RMB 380B ($52.7B) investment. Per FutureCIO's coverage of Alibaba Cloud's agentic AI infrastructure unveil and Alibaba's "Reshaping Alibaba Cloud for AI Innovation" piece, OSS is no longer positioned as passive storage at the bottom of the stack — it is the foundation of a full-stack AI-native infrastructure, anchored by the Panjiu AI Infra 2.0 platform and a multi-year RMB 380B investment program. The shift is operational, not marketing: every layer above OSS (Lakehouse, vector retrieval, lakehouse-streaming, Kubernetes-native RDMA) is being co-designed against OSS rather than bolted on top of a passive blob store.
- Vector Bucket + Metaquery — 2 billion rows per table, 85% precision-recall. Per Alibaba's Vector Bucket + Metaquery launch writeup and the multi-index table architecture documentation, a single Vector Bucket scales to 2 billion rows per table via a multi-index table architecture, functioning as the fundamental storage unit for large-scale vector retrieval. Metaquery then layers a content-awareness engine on top: OSS uses multimodal AI models to automatically extract semantic descriptions and generate vectors the moment an object is written, then exposes hybrid retrieval (scalar metadata filters + vector similarity) over the standard object-storage API with natural-language inputs. Production deployments report up to 85% precision-recall on hybrid retrieval, significantly outperforming self-built search solutions. This converts OSS from a "dumb hard drive" into a queryable knowledge graph — the architectural antithesis of AWS S3 Vectors' Bedrock-routed lock-in path.
- OSS Accelerator (EFC Cache) — 100 GB/s burst, single-digit ms latency, no data migration. Per Alibaba's Vector Bucket + Metaquery deep-dive, the OSS Accelerator (EFC Cache) is the compute-proximate caching tier — NVMe cache nodes deployed in the same availability zone as the GPU cluster, reducing standard object-storage retrieval latency to single-digit milliseconds and supporting burst throughputs up to 100 GB/s. Unlike CPFS (which is a separate parallel file system the data has to land in), EFC Cache keeps the data in OSS and accelerates reads in place. When paired with OSSFS V2's POSIX-like mount and negative-metadata cache, this matches PyTorch DataLoader sequential-read patterns directly off the bucket — no migration step, no separate file-system tier.
- CPFS for Lingjun + OSS Connector for AI/ML + ossfs 2.0 — the tiered data path. Per Alibaba's CPFS product page, the PAI OSS docs, and the open-source OSS Connector, the AI data path now tiers: CPFS for Lingjun (2 TB/s, 30M IOPS) for synchronous-checkpoint and MoE-synchronization workloads, standard CPFS (2.8M IOPS) for fine-tuning and analytics, OSS Accelerator for read-heavy workloads that should stay on object storage, and the OSS Connector for AI/ML (PyTorch-native, no POSIX mount, optimized for small-file streaming) for direct DataLoader feeds. Teams pick the right tier per pipeline instead of forcing every workload through one driver.
- HPN8.0 + eRDMA — 800 Gbps per node, 15,000 GPUs per pod, TCP/IP bypass. Per the Alibaba HPN paper (SIGCOMM 2024) and Alibaba Cloud's eRDMA documentation, HPN8.0 is a 2-tier dual-plane data-center network engineered specifically for LLM-training traffic shapes (small number of periodic highly-bursty flows that would otherwise overwhelm standard ECMP/Clos hash polarization), delivering 800 Gbps per node and supporting seamless cluster scaling to 15,000 GPUs in a single pod. Coupled with elastic RDMA (eRDMA) — kernel-bypass zero-copy memory transfers over standard VPC underlays — and the
aliyun/erdmaKubernetes extended resource for direct ACK pod access. This is the connective tissue that makes Vector Bucket + EFC Cache + CPFS physically viable at scale — without HPN8.0 the storage layer's throughput numbers are unreachable. - Apache Paimon streaming-batch bridge — unifying analytics + AI on OSS. Per Alibaba's Paimon-Iceberg-compatibility writeup and vivo's lakehouse case study, Apache Paimon sits directly on OSS as the streaming-batch unified table format — LSM-tree write path delivers minute-level ingestion latency with primary-key deduplication, every write creates an immutable snapshot enabling time travel + ACID, and Flink CDC propagates schema changes from source systems automatically. The practical effect: analytics queries (StarRocks/Trino) and multimodal AI training run against the same OSS-backed Paimon tables with no separate ETL hop. This is the architectural mechanism by which OSS becomes "an active intelligence platform rather than a passive data graveyard."
- Saudi Cloud Computing Company (SCCC) + ASEAN BSSN Level 5 — sovereign expansion via joint ventures. Per Capacity's SCCC launch coverage and the SCCC capabilities portal, SCCC is the Alibaba Cloud joint venture with STC Group, eWTP Arabia Capital, the Saudi Company for Artificial Intelligence (SCAI), and the Saudi Information Technology Company (SITE), operating data centers physically located in Riyadh — Saudi enterprises train AI models on NVIDIA L20/H20 GPU clusters inside the Kingdom against locally-encrypted OSS buckets, satisfying PDPL data-residency requirements absolutely. In Southeast Asia, Alibaba operates 91 availability zones across 29 global regions, with Alibaba Cloud achieving the highest Level 5 (Innovative) IKAS maturity ranking from Indonesia's BSSN — the regulatory groundwork for an interoperable non-Western technological sphere of influence.
- MIIT 2025-2027 codification — parallel cloud standards by 2027. Per China's State Council announcement and SESEC's analysis, the Ministry of Industry and Information Technology (MIIT) issued the "Cloud Computing Comprehensive Standardization System Construction Guideline (2025 Edition)," mandating the development of over 30 new national and industry cloud-computing standards by 2027 across six architectural components (basic terminology, technical frameworks, service delivery including AIaaS/DaaS, application integration, management, security). Codifies domestically that "object storage" is now defined parallel to the Amazon S3 standard. As Chinese providers export their infrastructure to ASEAN and the Middle East, these standards travel with them — the bifurcation is not just commercial drift, it is regulatory.
- S3 compatibility drift — concrete divergences, not just framing. Per Alibaba's own S3 compatibility documentation and the Chernov migration notes, the divergences from AWS S3 are explicit and operational: OSS supports only virtual-hosted-style URLs (no path-style — legacy AWS SDK code using path-style fails instantly); ETag generation for standard PUT uses uppercase letters (AWS uses lowercase — client-side hash validations trigger false failures from case sensitivity); multipart-upload ETags use a proprietary algorithm (file-integrity checks across hybrid-cloud sync mismatch); access-control semantics diverge with heavily restricted canned ACLs (Terraform / Pulumi modules must be rewritten to map permissions). "Cross-region OSS↔S3 mirroring" is now its own engineering line item — middleware like dify-cloud-kit or heavily-customized rclone, not a flag flip.
Connections 11
Outbound 10
scoped_to3implements1constrained_by1Inbound 1
depends_on1Resources 5
Alibaba Cloud's official OSS product page covering pricing tiers, storage classes (Standard/IA/Archive/Cold Archive), and the regional availability footprint.
OSS API reference and developer guide — the ground truth for S3-compatibility scope and the OSS-specific authentication signing scheme.
Operation-level documentation showing which AWS S3 API operations are supported, including multipart upload behavior and lifecycle rule semantics.
The canonical workaround for the AWS SDK v2 `STREAMING-UNSIGNED-PAYLOAD-TRAILER` incompatibility — required reading for anyone porting Iceberg/Polaris/Spark v2-SDK stacks onto OSS.
CPFS (Cloud Parallel File System) overview — the POSIX layer over OSS that Qwen 3 was trained against, and the architectural template for "object storage with sub-1ms metadata for training pipelines."