NVIDIA verbally required Chinese cloud providers on November 4 to reduce the proportion of high-end computing power and mandate that new cloud contracts allocate over half of sellable computing capacity to non-Chinese customers. On the same day, Alibaba (9988.HK), Tencent (0700.HK), and ByteDance collectively placed orders for approximately 6 million RTX Pro units valued at an estimated $70 billion, but sources indicated only about 1% — tens of thousands of units — may arrive by February next year. The restrictions and delivery delays have created an immediate computing power shortage in China's cloud infrastructure market, where GPU rental prices have risen to approximately $5.72 per hour for certain models as providers turn to leasing and secondary markets to bridge supply gaps.
Chinese Cloud Providers Place $70 Billion GPU Orders Amid Delivery Constraints
Chinese cloud providers responded to uncertainty by simultaneously placing large-scale orders and accelerating equipment leasing and secondary market procurement. Alibaba, Tencent, and ByteDance's combined order of approximately 6 million RTX Pro units represents an attempt to maintain training capacity and expansion plans, but the estimated 1% delivery rate by February next year leaves a substantial near-term gap. GPU rental prices in the secondary market have increased, with certain models reaching approximately $5.72 per hour. While leasing and secondary market purchases provide short-term relief for some workloads, they cannot replace new equipment arrivals for long-term capacity expansion. NVIDIA's verbal requirements remain in a gray area regarding contract design and customer allocation; formalization into written rules would create more explicit constraints on medium- to long-term procurement and deployment.
HBM and Advanced Packaging Supply Limits Domestic GPU Substitution
The industry is advancing domestic accelerator and integrated system solutions, including Huawei and Ascend platforms, to reduce long-term dependence on imported GPUs, but critical component and process capabilities remain primary bottlenecks. High-bandwidth memory (HBM) supply is highly concentrated, and advanced packaging technologies such as CoWoS face delivery timeline and yield pressures, limiting both per-unit performance and large-scale production timelines for domestic accelerators. Training workloads transitioning to domestic hardware require software compatibility verification, framework support, and system integration, preventing immediate "plug-and-play" equivalent performance in the near term. The immediate supply constraints stem from delayed high-end GPU deliveries and concentrated supply of HBM, advanced packaging, and advanced process resources.
Investment Banks Lower AI Supply Chain Valuations on Shipment Concerns
Market impact has already appeared in capital markets, with some investment banks reducing price-to-earnings multiples and target prices for Chinese AI supply chain companies, reflecting concerns about shipment timing and price competition risks. NVIDIA stated in its financial reports and outlook that older chips permitted for sale to China represent less than 1% of data center revenue, and the company excluded Chinese data center revenue from its Q3 outlook. This creates visible tension between company statements and market expectations. Near-term computing power gap resolution depends on two factors: whether NVIDIA converts verbal requirements into actionable written rules with specific shipment schedules, and whether domestic accelerators and related supply chains (HBM, advanced packaging, processes) can improve yield and delivery timelines within months to make equivalent computing power genuinely usable.
FAQ
What restrictions did NVIDIA place on Chinese cloud providers on November 4?
NVIDIA verbally required Chinese cloud providers to reduce the proportion of high-end computing power and mandate that new cloud contracts allocate over half of sellable computing capacity to non-Chinese customers.
How many RTX Pro units did Alibaba, Tencent, and ByteDance order, and what is the expected delivery rate?
The three companies collectively ordered approximately 6 million RTX Pro units valued at an estimated $70 billion, but sources indicated only about 1% — tens of thousands of units — may arrive by February next year.
What bottlenecks limit domestic GPU substitution in China?
High-bandwidth memory (HBM) supply is highly concentrated, and advanced packaging technologies such as CoWoS face delivery timeline and yield pressures, limiting both per-unit performance and large-scale production timelines for domestic accelerators.
