Guosen Securities indicates NVIDIA's Q1 report revealed $4.5 billion in inventory write-downs and procurement obligation losses, suggesting H20 inventory revenue likely approached $10 billion. Resumption of H20 procurement is expected to trigger aggressive stockpiling by cloud providers, substantially boosting capital expenditure.
Despite a modest dip in AI client numbers during Q2, scenario penetration rates continued climbing across domestic cloud providers, maintaining the sector's vigorous expansion momentum. Post-validation of business use cases, heightened usage by mid-to-large clients and private solution deployments should accelerate cloud growth in H2 2025.
While domestic GPU purchases gained share over rentals in Q2, yield limitations and delivery bottlenecks constrained computing capacity. Consequently, providers consumed inventory chips. Capital expenditure plateaued quarterly as cloud firms temporarily moderated AI server acquisitions amid H20 supply constraints. The dual resumption of H20 and Blackwell-based specialized GPUs should catalyze a quarterly inflection point in demand.
China's AI model segment saw further consolidation among leaders during Q2. Alibaba concentrated on versatile high-performance open-source models, with Qwen3 claiming global open-source supremacy in April before launching Qwen3-Embedding and 3D digital human projects. ByteDance debuted its Doubao 1.5 reasoning model and video generation tools, while Tencent advanced multi-domain innovations—its Hunyuan TurboS entered ChatbotArena's global top eight. Baidu pioneered foundational models and development utilities, releasing Ernie 4.5 Turbo before introducing interactive digital avatars.
H20's restored availability alleviates China's AI GPU scarcity, slashing deployment timelines and experimentation costs. This accelerates enterprise cloud adoption and fuels downstream demand. Investment recommendations include NVIDIA (NVDA.US), Alibaba (09988.HK), Tencent Holdings (cloud and social ecosystem), and Xiaomi Group (01810.HK) for its device hardware network. Risks encompass delayed AI model advancements, sluggish commercialization, and safety concerns like algorithmic hallucinations.