Alibaba Cloud (阿里云)

#1 in China's AI public cloud market share; both the captive compute for first-party Qwen models, and China's exportable GPU cloud — one of the largest relative beneficiaries under H20 export controls.

1. Core Product / Service

Alibaba Cloud (Aliyun) is Alibaba Group's cloud business; AI-related core includes GPU ECS instances + Lingjun AI compute clusters + Bailian (DashScope) model service + first-party Tongyi (Qwen) models.

AI-related SKUs:

• GPU ECS gn7e/gn7i — 8× NVIDIA A100 80GB; early workhorse, still on sale.
• gn8v / gn8x — 8× NVIDIA H800 / H20 (export-control-compliant version), the main offering from 2024–2025 [1]. Structurally tight inventory and new orders due to the 2025 H20 suspension [7].
• Lingjun AI compute clusters — 10K-card-scale RDMA networks, serving Qwen training internally, and sold externally via contracts to finance/government/AI startups.
• Domestic GPU instances — instances compatible with multiple domestic chips like Cambricon MLU, Huawei Ascend, MetaX (specific SKU details not uniformly public).
• Bailian / Tongyi Lingma — model API platform hosting the full Qwen series + third-party models (DeepSeek, Llama, Stable Diffusion, etc.).
• Tongyi (Qwen) — Qwen2.5, Qwen3, QwQ and other first-party model families, with both open weights and closed-weight flagships.

Capacity strategy: NVIDIA H20/A100 + domestic GPU + in-house Yitian ARM CPU — three tracks; from 2025, capex was sharply revised up, with a 3-year ¥380B (~$53B) AI infrastructure budget [3].

2. Target Users & Pain Points

• Large Chinese enterprises / government / financial institutions — need on-premises LLM deployment + data localization compliance; Alibaba Cloud is the preferred public cloud option.
• Chinese AI startups / model companies — external Qwen ecosystem users; Moonshot Kimi (early days), Zhipu, 01.AI etc. have all trained models on Alibaba Cloud.
• Overseas customers targeting China / Southeast Asia customers — Alibaba Cloud's international version (alibabacloud.com) has the deepest overseas penetration among Chinese clouds.
• Internal Alibaba Group — Taobao, Cainiao, DingTalk, Quark AI Search etc. are captive workloads.

Pain points solved: top-tier Chinese-context model quality, mixed domestic/H20 supply under export controls, ecosystem lock-in with Alibaba Group, and full regulatory licensing.

3. Competitive Landscape

Differentiation: #1 public cloud share in China (per IDC, Alibaba ~33–37% IaaS share 2024–2025, also #1 in AI PaaS) [6]; Tongyi model quality benchmarks GPT-4o/Claude 3.5 in Chinese scenarios; relatively ample GPU reserves (stockpiled heavily before export controls).

4. Unique Observations

• GPU pricing: H20 GPU instance (gn8x.4xlarge per-card-level ~ 8-card chassis) public on-demand list price ¥80–¥120/card·hour ($11–$17/H20·hour), annual reserved 1-year ¥45–¥65/card·hour ($6.30–$9.10), 3-year deep commitment ¥30/card·hour ($4.20) [1]. H800 / A100 are no longer receiving new supply domestically; inventory prices exceed H20. H20 was briefly placed under new U.S. controls in April 2025 then released again, and Alibaba Cloud as one of the largest downstream buyers experienced significant short-term volatility [7]. B200 is not sold to China, so Alibaba Cloud has no B200 SKU.
• Qwen captive load: Tongyi Qwen3 flagships (235B MoE / 30B dense, etc.) are trained entirely on in-house Lingjun clusters; the full Qwen series open weights + closed-weight max versions are sold via Alibaba Cloud's first-party API [5]. This is similar to google-cloud Gemini and microsoft-azure OpenAI's first-party captive model — the difference is that Qwen also open-sources weights, allowing customers outside Alibaba Cloud PaaS to also run Qwen in their own stacks, thereby expanding ecosystem spillover.
• China AI cloud share: In 2024 IDC China AI public cloud market share, Alibaba ~33%, #1; Volcengine ~14%, fastest-rising; Baidu ~12%, Tencent ~10%, Huawei ~7% (numbers vary slightly by methodology) [6]. AI PaaS (model service) share Alibaba > 30%, more pronounced.
• GPU sourcing under export controls: (a) H20 is NVIDIA's H100 cut-down version designed for China compliance, with heavy 2024–2025 inflows; in 2025, Alibaba Cloud is widely considered one of the single largest H20 downstream buyers; (b) Huawei Ascend 910B/910C purchased from Huawei as supplement; (c) Cambricon MLU, MetaX, Enflame and other domestic chips as diversification; (d) historical H100/A100 inventory still in use [7]. The group announced a 3-year ¥380B AI/cloud capex plan in February 2025 [3] — the highest explicit numerical commitment among domestic clouds.
• AI revenue share: Alibaba Cloud FY2025 Q3 (Oct–Dec 2024) revenue ¥31.7B (+13% YoY); AI-related revenue has grown triple digits YoY for 6 consecutive quarters (management call language) [2][3]. The group does not disclose AI standalone absolute values; external estimates put AI-related revenue at ¥10–15B quarterly run-rate (1/3+ of Alibaba Cloud), expected to ramp quickly as capex translates.
• Customer concentration: No major single-customer disclosure. Internal group captive (Taobao/DingTalk/Quark AI Search) is estimated at ~25–35%; external finance/government long-tail contracts are relatively dispersed; unlike microsoft-azure's OpenAI or oracle-cloud's Stargate, there is no super anchor. This is actually a structural advantage: single-customer risk is lower than U.S. hyperscalers.

5. Financials / Funding

• Parent: Alibaba Group (NYSE: BABA / HK: 9988); Alibaba Cloud is a publicly disclosed segment.
• FY2025 Q3 (Oct–Dec 2024): Alibaba Cloud revenue ¥31.74B (+13% YoY), adjusted EBITA ¥3.13B [2].
• FY2025 Q4 / full year: full-year cloud revenue ~¥119B; FY2026 (starting Apr 2026) growth guidance accelerates.
• AI Capex: Group committed to ¥380B (~$53B) AI / cloud infrastructure over 3 years (FY2025–FY2027) [3] — the highest single item among Chinese private tech companies.
• AI-related revenue: triple-digit YoY growth (6+ consecutive quarters), absolute value not separately disclosed.
• Equity structure: Alibaba Group 100% owns Alibaba Cloud.

6. People & Relationships

• Alibaba Cloud CEO: Eddie Wu (吴泳铭), concurrently Alibaba Group CEO directly responsible for cloud and AI; from 2023, restructuring returned the cloud business to the group.
• CTO / Tongyi lead: Jingren Zhou (周靖人).
• Group Chairman: Joe Tsai (蔡崇信).
• Key partnerships / investments: invested in Moonshot kimi (leading/following multiple rounds, 2024 ~$10B+ valuation round); cloud resource partnerships with 01.AI, Zhipu AI; 2025 rumored investment in deepseek unconfirmed.
• GPU supply: NVIDIA (H20 / historical H100/A100), Huawei Ascend, Cambricon, MetaX.
• Competitors: tencent-cloud, huawei-cloud, aws, ByteDance Volcengine, Baidu AI Cloud.

Sources

[1] https://www.alibabacloud.com/help/en/ecs/product-overview/gpu-accelerated-instance-families (2026-05-10) [2] https://www.alibabagroup.com/en-US/document-1745597821080158208 (2026-05-10) [3] https://www.reuters.com/technology/artificial-intelligence/alibaba-53-billion-ai-capex-2025-02-24/ (2026-05-10) [4] https://www.scmp.com/tech/big-tech/article/3296715/alibaba-cloud-china-ai-leader-qwen-2025 (2026-05-10) [5] https://qwenlm.github.io/blog/qwen3/ (2026-05-10) [6] https://idc.com/getdoc.jsp?containerId=prCHC52850225 (2026-05-10) [7] https://www.bloomberg.com/news/articles/2025-04-15/alibaba-nvidia-h20-chip-export-controls (2026-05-10)