Alibaba Cloud

Full briefing

Alibaba Cloud (Aliyun), the cloud computing subsidiary of Alibaba Group (NYSE/HKEX: BABA), is the largest cloud provider in China and the third-largest globally by revenue, behind AWS and Azure. Founded in 2009, the service generated approximately $15 billion in annual revenue by fiscal year 2024, with AI-driven workloads accounting for a growing share. The cloud division operates a network of data centers across mainland China, Hong Kong, Singapore, Japan, Germany, Australia, and the United States, giving it one of the most globally distributed footprints among Asian cloud providers. Alibaba's AI chip strategy spans three hardware tiers. First, the Hanguang 800 NPU, revealed in 2019 and fabricated by TSMC on a 12nm process, was designed by Alibaba's DAMO Academy specifically for e-commerce recommendation, image search, and natural language inference workloads. The Hanguang 800 powers Taobao and Alibaba's advertising systems, processing trillions of inference queries daily. It is not a general-purpose training chip but rather a highly specialized inference accelerator — a differentiated approach that reduces Alibaba's dependence on external GPUs for its core workloads. Second, Alibaba Cloud operated NVIDIA A100 and H800 clusters for large-scale AI training; however, following the BIS October 2023 export control expansion, it can no longer receive H100/H200 exports and has transitioned to NVIDIA's China-compliant H20 GPU. Third, Alibaba Cloud has contracted Huawei Ascend 910B clusters as a domestic fallback for workloads that require higher interconnect bandwidth than the H20 provides. Alibaba's Qwen (通义千问) LLM family launched in 2023 and has become one of China's most widely deployed model series. Qwen models span sizes from 0.5B to 72B parameters, with multimodal variants (Qwen-VL) and code-specialized versions (Qwen-Coder). The Qwen2 series released in June 2024 posted competitive benchmark scores against international open-weight models including LLaMA 3 and Mistral. Alibaba has open-sourced most Qwen variants under permissive licenses, making them the most popular Chinese open-weight model series on Hugging Face by download count as of early 2025. The export control impact on Alibaba Cloud is multidimensional. The H20 GPU — NVIDIA's downgraded China-compliant product — delivers significantly reduced NVLink bandwidth and GPU-to-GPU interconnect performance compared to H100/H200, making large-scale training runs less efficient. Alibaba's response has been to invest more heavily in model efficiency techniques, distributed training frameworks optimized for reduced interconnect bandwidth, and longer training runs that compensate through duration. The Qwen training pipeline has been adapted to run across Huawei Ascend clusters, though software toolchain maturity remains lower than CUDA. Alibaba Cloud competes directly with Tencent Cloud, Huawei Cloud, and Baidu AI Cloud in the Chinese market, with AI cloud services — including GPU rental, model API services, and enterprise LLM deployment — becoming the primary growth vector. The Tongyi (通义) brand unifies Alibaba's AI products under a single identity, with Tongyi Qianwen as the consumer-facing chatbot and Tongyi API giving developers access to Qwen models. Alibaba's sprawling consumer ecosystem — Taobao, Tmall, Alipay, Ele.me, DingTalk — provides a large captive deployment surface for AI features, giving Alibaba Cloud an advantage in domestic AI market share that its global cloud peers cannot easily replicate.