Huawei Cloud (华为云)

The only fully non-NVIDIA AI cloud stack in China: in-house Ascend (昇腾) 910B/910C chips + in-house ModelArts platform + in-house Pangu (盘古) large model — the representative "sovereign cloud" forged by sanctions.

1. Core Product / Service

Huawei Cloud is the cloud business of Huawei; the fundamental difference vs. other major US-China clouds is that Huawei is the only player on this value chain with "hardware + compiler + cluster networking + model platform + first-party (1P) model" all developed in-house.

AI-related core:

• Ascend 910B / 910C — in-house AI chip; 910C mass production in 2025 [3]. 910C single-card BF16 ~800 TFLOPS (close to H100's ~990 TFLOPS), smaller HBM capacity but compensated through larger cluster sizes [2].
• CloudMatrix 384 / supernodes — Huawei's 2025 "supernode" with 384 910C chips fully optically interconnected; total system compute targets NVIDIA GB200 NVL72, with higher deployment density (lower power efficiency but greater chip density) [4].
• AI accelerator clusters + Atlas 800/900 server series — 10k-card-scale Ascend clusters, sold externally to government, finance, telecom, Saudi Arabia, and Middle East markets.
• ModelArts — integrated model training + inference + MaaS platform; CANN (Compute Architecture for Neural Networks) is its compiler/software stack, the analog to NVIDIA CUDA.
• Pangu large model — the 5.0 series launched 2024, focusing on industry large models (healthcare, mining, weather, government). The foundation model quality is lower than Qwen3 / Hunyuan / DeepSeek, but industry deployment depth is high [5].
• GPU instances (NVIDIA route) — Huawei Cloud still offers a small number of H20 / legacy A100 instances for customers needing CUDA compatibility; the strategic main line is Ascend.

2. Target Users & Pain Points

• Chinese government / state-owned enterprises / SOEs — top choice under the "Xinchuang" (信创) catalog; highest demands for security review and self-controllable autonomy.
• Telecom carriers / finance / energy / manufacturing — private deployment + industry large models; Huawei's traditional ICT customer base.
• Overseas sovereign cloud market — Middle East (Saudi PIF data center 2025 [7]), Africa, Southeast Asia, Latin America — governments needing "non-US" cloud options.
• DeepSeek and the Ascend ecosystem — after Spring Festival 2025, DeepSeek-R1's inference optimization on Ascend 910B made Huawei Cloud the primary choice for inferring R1 on a non-NVIDIA path [3].

Pain points solved: sustainability under sanctions/export controls (not dependent on NVIDIA), government compliance, industry-model + private deployment + integrated hardware delivery.

3. Competitive Landscape

Differentiation: the only "fully non-US" route — top choice for government / sovereign / sanction-sensitive customers. Weaknesses: CUDA ecosystem moat is still enormous; CANN software stack maturity lags behind CUDA; developer ecosystem is thin; Pangu model quality is not in the top tier.

4. Unique Observations

• GPU pricing (Ascend): Huawei Cloud doesn't publicly disclose Ascend per-card-hour pricing; industry intel data — 910B monthly commitment roughly ¥30k-¥50k/month (equivalent to ~$60-$100/card·day, ~$2.50-$4.20/card·hour), 910C slightly higher [3]. No H100/H200/B200-equivalent SKU (due to a combination of export controls + strategic choice). The NVIDIA route only offers limited H20 supply, priced similarly to Alibaba/Tencent at ~¥80-¥120/card·hour.
• Capacity source is Ascend-only (strategically): aside from limited H20 instances retained for compliance needs, the main line bets entirely on Ascend. 910C entered mass production in 2025 Q2, with design foundry originally relying on SMIC's 7nm-equivalent process; sanctions slowed the capacity ramp below expectations, but internal 2025 target was annualized 500k-800k 910C chips [3]. CloudMatrix 384 supernodes compensate for single-card performance gap with more chips + fully optical interconnect; according to Huawei's public data, single-system training throughput approaches GB200 NVL72 [4]. This is the critical bet for China's AI compute path.
• Sovereign Cloud play: 2025 signing with Saudi PIF to build an Ascend data center [7], with subsequent negotiations in Malaysia, Indonesia, Mexico. Logic: in the US-China decoupling landscape, third-world countries don't want to bet entirely on AWS/Azure/GCP (data + sanctions risk); Ascend + ModelArts provides a "Chinese/Huawei version" sovereign AI stack, similar to Huawei's historical telecom equipment expansion path. This is a niche aws / microsoft-azure / google-cloud won't pursue.
• ModelArts' role: equivalent to a combination of AWS SageMaker + Bedrock — private deployment + model marketplace + Agent platform. From 2024, MaaS revenue + model subscription growth has been rapid, but absolute scale is still smaller than Alibaba Bailian. Strengths are co-design with Ascend hardware (CANN compiler + ModelArts upper-level scheduling); training MFU on 910B reaches 40%+, approaching NVIDIA H100's 50%+ level [3][4].
• AI revenue share: Huawei doesn't break out cloud revenue. Group FY2024 revenue ¥862.1B (+22% YoY), with "ICT Infrastructure + Digital Energy" as the main axis; Huawei Cloud is generally estimated at the ¥80-100B/year range, with the AI-related share rising but specific values undisclosed. Pangu + ModelArts + Ascend full-system sales combined could be in the ¥30-50B range.
• Customer concentration: undisclosed. China captive (internal Huawei + affiliated customers: SOE telecom, power grid, national pipeline, etc.) estimated at ~25-30%, with the remainder dispersed. Sovereign cloud overseas contracts (Saudi + Middle East) are becoming the anchor incremental growth, but absolute value hasn't reached oracle-cloud-OpenAI scale.

5. Financials / Funding

• Parent: Huawei Investment & Holding Co. (private, employee-owned via Union Trade Association); Huawei Cloud is an internal BG, not separately disclosed.
• Group FY2024 (reported Mar 2025): total revenue ¥862.1B (~$118B, +22% YoY); net profit ¥62.6B.
• Huawei Cloud revenue estimate: external estimate ~¥80-100B/year (not public).
• Ascend business: not separately disclosed; 910C mass production 2025 Q2, annual capacity target 500k-800k chips [3].
• R&D investment: FY2024 ¥179.7B (20.8% of revenue), long above industry average.
• Not listed: Huawei is a private company.

6. People & Relationships

• Group founder / CEO: Ren Zhengfei (任正非).
• Rotating Chairman / CEO system: currently rotating among Meng Wanzhou, Xu Zhijun, Hu Houkun.
• Huawei Cloud head: Zhang Ping'an (张平安), Executive Director + Huawei Cloud CEO.
• Ascend head: Zhang Dixuan (张迪煊, President of Ascend Computing Business).
• Pangu head: Tian Qi (田奇, Chief Scientist, Huawei Cloud BU).
• Key partnerships: DeepSeek (R1 inference optimization on Ascend); China Telecom, China Mobile, China Unicom; Saudi PIF; multiple Middle East sovereign funds.
• GPU route: in-house Ascend 910B/910C; NVIDIA H20 retained in limited quantity as compatibility path.
• Competitors: alibaba-cloud, tencent-cloud, ByteDance Volcano Engine, aws, microsoft-azure.

Sources

[1] https://www.huaweicloud.com/intl/en-us/product/modelarts.html (2026-05-10) [2] https://e.huawei.com/en/products/computing/ascend (2026-05-10) [3] https://www.reuters.com/technology/artificial-intelligence/huawei-ascend-910c-mass-production-2025-04-21/ (2026-05-10) [4] https://www.scmp.com/tech/big-tech/article/3303871/huawei-cloudmatrix-384-launch-2025 (2026-05-10) [5] https://www.huaweicloud.com/intl/en-us/news/2024-pangu-large-model-5-0.html (2026-05-10) [6] https://idc.com/getdoc.jsp?containerId=prCHC52850225 (2026-05-10) [7] https://www.bloomberg.com/news/articles/2025-09-15/huawei-saudi-arabia-data-center-deal (2026-05-10)