Technology

Huawei OBS

Huawei Cloud's Object Storage Service — S3-compatible, tightly co-engineered with Huawei's domestic AI accelerator (Ascend 910B/910C) and the MindSpore framework. The de facto storage tier for foundation-model training that targets domestic-silicon clusters; notably the storage substrate for Zhipu AI's GLM-5 (744B parameters trained on 100,000 Ascend 910B chips).

9 connections 3 resources 3 posts

Summary

What it is

Huawei Cloud's Object Storage Service — S3-compatible, tightly co-engineered with Huawei's domestic AI accelerator (Ascend 910B/910C) and the MindSpore framework. The de facto storage tier for foundation-model training that targets domestic-silicon clusters; notably the storage substrate for Zhipu AI's GLM-5 (744B parameters trained on 100,000 Ascend 910B chips).

Where it fits

The third pillar of the China S3 trio. Where Aliyun OSS is the broad-market default and Tencent COS is the Tencent-ecosystem default, Huawei OBS is the **vertical-stack** choice — chosen specifically when the training run targets Huawei silicon, integrating with MindSpore and the company's data-fabric stack.

Misconceptions / Traps
  • Huawei OBS is the storage half of a vertically-integrated NVIDIA-substitute stack; using it without Ascend silicon and MindSpore loses most of the integration value.
  • Like the other Chinese clouds, the architectural assumption is in-PRC regions for in-PRC users; international footprint is limited (Singapore, Bangkok, Johannesburg, Mexico City, São Paulo).
  • Huawei is on US Entity List restrictions — Western cloud-portability planning around Huawei OBS must account for sanctions exposure separate from the data-localization story.
  • 11 nines durability vs Aliyun OSS / Tencent COS at 12 nines. Marginal in practice but worth noting for compliance-team comparison sheets.
  • The MLPS Level 3+ certification advantage is real for Chinese government / SOE / regulated industry tenders — Huawei is structurally preferred there. But that advantage doesn't transfer to commercial / consumer-facing workloads, where Tencent or Alibaba ecosystem reach matters more.
  • WORM retention is available but configuration-by-bucket — make the compliance-vs-cost tradeoff at provisioning, not retroactively.

Pricing posture (April 2026): Standard ~$0.017–0.018/GB-mo → IA ~$0.009 → Archive ~$0.0045 → Deep Archive ~$0.002. Outbound traffic ~$0.08–0.12/GB tiered. Free tier of 5 GB Standard.

Why it's the China-AI choice when sovereignty is the constraint: ModelArts (AI platform) + HiLens (edge AI) + IoT Platform integrations come for free with OBS. Zhipu AI's GLM-5 (744B, 100,000 Ascend 910B) trained on OBS as the storage tier — operational proof point for the Ascend-Huawei-OBS-MindSpore vertical stack as an NVIDIA substitute. Object size cap 48.8 TB via parallel multipart upload makes it competitive with S3 at training-corpus scale.

Key Connections
  • implements S3 API
  • enables East Data West Computing
  • enables Training Data Streaming from Object Storage — Ascend + MindSpore + OBS as a stack
  • solves China Data Localization
  • scoped_to Sovereign Storage

Definition

What it is

Huawei Cloud's Object Storage Service — an S3-compatible object store tightly co-engineered with Huawei's domestic AI accelerator stack (**Ascend 910B/910C**) and the **MindSpore** training framework. The de facto storage tier for foundation-model training that uses Huawei silicon as a substitute for restricted NVIDIA hardware, including **Zhipu AI's GLM-5** (744B parameters trained on 100,000 Ascend 910B chips). Provides standard S3 API compatibility plus deep integration with Huawei's data-fabric stack (DLI, MRS, GaussDB). **99.999999999% (11 nines) per-object durability**.

Why it exists

US semiconductor export controls cut Chinese labs off from H100/H200/Blackwell GPUs. Huawei's response — Ascend silicon + MindSpore + OBS — is the vertically integrated Chinese alternative to the NVIDIA + S3 + PyTorch stack. OBS is the storage half of that bet, and the strategic destination for any training run that cannot use US-headquartered cloud or hardware.

Primary use cases

Training storage for foundation models running on Ascend silicon, S3-compatible object storage inside Huawei Cloud regions, sovereign-cloud deployments for regulated industries (Chinese banks, telcos, government), GLM-5 / Pangu / DeepSeek-class training corpora that must stay on domestic hardware.

Recent developments

Latest signals
  • OBS 3.0 documentation actively maintained through April 2026. Huawei published OBS 3.0 24.3.0s-HCS (Huawei Cloud Stack) documentation in February 2026, with subsequent updates through April 2026 — the service is in active development, not maintenance mode. Per Huawei Support — OBS 3.0 24.3.0s-HCS docs.
  • S3-compatible API surface — drop-in for standard S3 tooling. Familiar S3-compatible interface, high durability, multi-region availability at often-lower costs vs hyperscaler incumbents. Per Huawei Cloud — Object Storage Service product page.
  • Migration tooling: Amazon-S3 → OBS via Object Storage Migration Service (OMS). Huawei publishes a best-practices guide for migrating data from AWS S3 to Huawei Cloud OBS using its Object Storage Migration Service — covering the inventory + transfer + verification pipeline. Per Huawei Support — Migrating from S3 to OBS.
  • Laravel storage adapter for OBS available. Community-developed Laravel storage adapter for Huawei OBS, signaling adoption beyond enterprise into the wider PHP/Laravel ecosystem. Per Medium — Laravel + Huawei OBS adapter.
  • Comprehensive API reference + multi-language SDK support. OBS API reference covers the full S3-equivalent surface plus Huawei-specific extensions (cross-region replication, custom domain bindings, advanced lifecycle rules). Per Huawei Cloud — OBS API Overview.

Connections 9

Outbound 8
scoped_to3
S3Object StorageSovereign Storage
implements1
S3 API
enables2
East Data West ComputingTraining Data Streaming from Object Storage
solves2
Vendor Lock-InChina Data Localization
Inbound 1
depends_on1
East Data West Computing

Resources 3

DocsHigh
www.huaweicloud.com/intl/en-us/product/obs.html

Huawei Cloud Object Storage Service product page covering storage classes, regional deployment, and integration with Huawei's data fabric (DLI, MRS, ModelArts).

DocsHigh
www.huaweicloud.com/intl/en-us/product/obs.html

OBS technical documentation and API reference — establishes the S3-compatibility scope and the integration path with Ascend/MindSpore for foundation-model training.

BlogMedium
letsdatascience.com/blog/china-trained-frontier-ai-model-glm...

Documents GLM-5's training architecture using 100,000 Huawei Ascend 910B chips with OBS as the storage tier — the canonical case study of the OBS + Ascend + MindSpore vertical stack at frontier-model scale.

Featured in

2026-05-10 The POSIX Gap is Closing: How S3 Quietly Became a File System The May 7 post was about why storage suddenly matters again for AI workloads. This one is about how the access model itself evolved. After a decade of failed FUSE clients trying to bolt POSIX semantics onto S3, the storage service finally absorbed the operations the filesystem world expected — and a string of 2020-2026 changes (strong consistency, Express One Zone, directory buckets, S3 Tables, GPUDirect Storage) made S3 Files possible. Plus the parallel story in China: Aliyun CPFS+OSS, Huawei OBS+MindSpore, the same shape drawn three different ways. 2026-05-08 When the Index Gained 13 Nodes: A Field Guide to the May 2026 Wave On May 7 the index added 13 new nodes — five technologies, six pain points, one architecture, one standard. Each answers a specific question engineers building on object storage are asking right now: what is China's S3 stack, why are GPUs starving, what replaced MinIO, what are the three data gravity wells, and a few more. This post walks through them by question rather than by node type, because that's how the questions actually arrive in production. 2026-05-07 When DeepSeek Made Storage Strategic: The China Direction Reshaping the S3 Ecosystem DeepSeek's $5.6M training run didn't just reset model architecture. It made object storage performance-critical, made silicon export controls a storage decision, and pushed Aliyun OSS, Tencent COS, and Huawei OBS from regional curiosities into the global S3 vocabulary. This index added 13 nodes to track what changed.