NVIDIA H200 shipments Q3 2026 remain constrained by export rules amid enterprise AI training demand

NVIDIA maintained an inventory of approximately 500,000 H200 GPUs primarily staged for Chinese buyers through October 2026, even as actual shipments remained limited by licensing requirements. Enterprise AI training workloads at hyperscalers and AI labs continued to generate strong orders for data center GPUs worldwide. Production of the China-specific H200 variant stopped in March 2026, with capacity redirected to other markets.

NVIDIA H200 shipments Q3 2026 face export caps and domestic competition

Under US rules updated in early 2026, licensed H200 exports to China carried a 25 percent fee, third-party testing, and firm-level volume caps near 75,000 units. ByteDance and Tencent each received roughly 10,000 units by mid-August, representing about 13 percent of their individual caps. Beijing’s NDRC approvals added further hurdles, and earlier reports described overall shipments to China as trivial in volume.

Chinese companies placed orders exceeding 2 million H200 units in the first quarter of 2026, far above available licensed stock. Domestic alternatives such as Huawei Ascend accelerators captured growing share, with IDC data showing NVIDIA at 34 percent of China AI accelerator shipments in the first half of 2026, down 44 percent year over year. Total China shipments reached about 2 million units, and non-GPU accelerators overtook GPUs for the first time.

Inventory that had reached estimates as high as 700,000 units earlier in the year, including some GH200 systems, sat largely unsold to Chinese buyers after production shifted. Capacity previously allocated to China-specific H200 chips moved to serve other regions where demand for enterprise AI training remained elevated.

Enterprise AI training sustains NVIDIA data center revenue growth

NVIDIA reported data center revenue of $51.2 billion for the quarter ending late October 2025, up 66 percent year over year, with hyperscalers and AI labs driving purchases for both training and inference. Cloud GPU capacity stayed sold out through 2026. Guidance emphasized compounding demand across workloads, while H20 sales in China contributed negligibly to results.

Google Cloud, Microsoft Azure, AWS, OpenAI, and Anthropic expanded infrastructure partnerships that relied on NVIDIA hardware for large-scale model training. These investments occurred alongside power and data center constraints that limited how quickly new capacity could come online. The H200, based on the Hopper architecture, served as an upgrade path for existing installations but competed with newer Blackwell systems in unrestricted markets.

Policy developments continued to shape availability. The US administration permitted limited H200 exports under case-by-case licenses, yet Beijing promoted local chips and imposed its own approval layers. Future shipment volumes will depend on ongoing regulatory decisions, production transitions toward Rubin architecture, and infrastructure buildouts outside China.