China to Approve NVIDIA H200 AI Chip Imports

NVIDIA H200 chips enter mainland China as Beijing balances local chipmaking goals with AI acceleration. Credit: Liu Liqun/Getty Images

China is easing limitations on NVIDIA H200 AI, permitting these shipments to help domestic technology firms train advanced models alongside local chips

Beijing is advancing its AI ambitions with a little help from NVIDIA by allowing controlled shipments of the US tech giant’s H200 chips to enter the mainland, relaxing regulatory restrictions to keep pace with international competitors.

The move is aimed at helping domestic technology firms train advanced models alongside local chips and to prevent local technology companies from falling behind US rivals in the race for advanced AI.

Major enterprise players, including ByteDance and Tencent, have reportedly secured initial consignments of approximately 10,000 NVIDIA H200 units each. Additional Chinese tech firms are expected to secure regulatory clearance for similar quantities shortly, as noted by the Financial Times.

While Washington permits Chinese firms to purchase up to 100,000 H200 units each under specific licensing agreements, the hardware remains two generations behind NVIDIA’s premier silicon offerings, which remain strictly off-limits under US trade restrictions.

US President Donald Trump said on 29 July that his administration is considering asserting more control over AI tools after the unprecedented incident. Credit: The White House

Impeding processing plans

To maintain momentum for local semiconductor manufacturers, Chinese authorities prefer that technology organisations deploy imported components beyond the mainland border.

Officials have advised firms that they can route and operate these processors through Hong Kong. However, utilising the region as a hardware hub comes with severe operational hurdles.

Hong Kong faces significant power constraints and lacks the necessary data centre infrastructure to accommodate high-density computing loads.

Industry stakeholders remain optimistic that regulatory conditions will ease over time as local infrastructure adapts to compute demands.

On the other hand, NVIDIA maintains an estimated 500,000 H200 units in reserve according to the FT, intended primarily for customers based in China.

Jensen Huang, Chairman and CEO of NVIDIA. Credit: Getty Images

Previous distribution efforts have stalled following regulatory scrutiny from Beijing aimed at bolstering local chip manufacturers such as Huawei.

Server integration partners, which were held up due to restrictions, are beginning to resume sales as well.

Technology firm Lenovo, which is a NVIDIA partner, informed clients last week that they can place new orders for systems powered by NVIDIA H200 components. This aggregates central processing units and networking gear into enterprise AI servers.

However, procuring these integrated systems is not automatic as prospective buyers must navigate a dedicated submission process managed by the National Development and Reform Commission (NDRC).

Beijing continues balancing this approval process against its long-term ambition to cultivate a self-sustaining semiconductor ecosystem capable of meeting national demands.

Bridging gaps in complex model training

On one hand, limited access to high-end semiconductor manufacturing tools prevents domestic foundries from satisfying total market demand. Consequently, on the other side of the pole, foreign hardware remains essential for resource-intensive AI workflows.

Regulators are therefore loosening limitations on US chips, granting access to international silicon so local developers can build foundational systems capable of competing with frontier global architectures like Anthropic Mythos 5.

As a result, recent releases from Chinese AI laboratories demonstrated a rapidly shrinking capability gap between domestic and global models.

In July, Moonshot unveiled Kimi K3, which demonstrated performance metrics exceptionally close to top American models.

Shortly after this development, technology firms Alibaba, DeepSeek and Z.ai introduced comparable iterations of their own.

Comparison data on coding performance that was released by the company. Credit: Moonshot

To optimise available resources, Chinese AI developers employ a hybrid infrastructure model, deploying domestic chips for real-time inference and response generation, while reserving NVIDIA processors for complex model training and pattern identification.

The hardware logistics are also unfolding alongside broader geopolitical pressures. According to an internal draft reviewed by Reuters, the US is planning to inform dozens of countries that they must choose sides in the global AI race. This came along with a warning that signing up for Beijing’s competing frameworks could lead to exclusion from a US-led coalition.

China responded by urging international respect for digital sovereignty, stating that it opposes division in the AI development.

“Each country has the right to choose its partners based on its national conditions and development needs,” said Lin Jian, Foreign Ministry Spokesperson of China Foreign Ministry at a regular briefing.