NVIDIA will resume sales of its H20 computing chip in China, according to an announcement on the company's official website. CITIC Securities highlights that despite performance adjustments to comply with U.S. export controls, the H20 retains competitive advantages in software ecosystems and cluster interconnect technology.

The chip was developed to address restrictions imposed by the U.S. Bureau of Industry and Security (BIS), which since 2022 has prohibited exports of semiconductors exceeding specific computational thresholds: either total processing performance (TPP) ≥4,800 TOPS or transmission rates ≥600 GB/s. The 2023 regulations further tightened controls by adding performance density (PD) parameters.

While the H20's peak computing performance is capped at 148 Tflops @ FP16, it maintains a robust 900 GB/s NVLink interconnect bandwidth. This preserves its cluster networking advantages, combined with NVIDIA's dominant CUDA ecosystem, making it a viable solution for Chinese internet companies.

The resumed shipments could accelerate capital expenditure among AI clients. Major internet firms have already signaled aggressive AI investments during recent earnings calls. CITIC expects downstream AI companies to significantly expand infrastructure spending in H2 2025, driving opportunities across the supply chain.

Server manufacturers stand to benefit directly from product shipments, while cloud service providers could enhance offerings through improved chip availability. Market observers should monitor developments in: 1) Server OEMs 2) Cloud computing providers

Potential risks include slower-than-expected AI industry growth, constrained capital expenditure from internet giants, ongoing regulatory constraints on advanced chips, and potential supply chain bottlenecks.