Chinese Tech Giants Train AI Models Offshore to Bypass US Chip Restrictions

Offshore AI Training Intensifies Amid US Sanctions

Major Chinese technology firms, notably Alibaba and ByteDance, are reportedly expanding their operations to train advanced artificial intelligence models in data centers located outside of China. This move is a direct response to the escalating US export controls that limit their access to high-performance Nvidia AI chips. The strategy involves leasing computing power from non-Chinese entities in foreign data centers, predominantly across Southeast Asia.

The shift has seen a steady increase since April, following further US restrictions on the sale of Nvidia's H20 chip, which was designed for the Chinese market. By utilizing offshore facilities, Chinese companies can access crucial hardware like Nvidia's A100 and H100 GPUs, which are essential for developing sophisticated large language models (LLMs).

US Export Controls and China's Response

The United States implemented comprehensive export controls on advanced computing and semiconductor manufacturing items to China starting in October 2022, citing national security concerns and aiming to curb China's technological advancements, particularly in military applications. These restrictions have significantly impacted Nvidia's ability to supply its most advanced GPUs directly to the Chinese market.

In response, Chinese tech giants have adopted a multi-pronged approach:

• Offshore Training: Leasing compute capacity in data centers in regions like Singapore and Malaysia, which are equipped with high-end Nvidia products. This practice is currently considered legally compliant under US regulations, especially after the rollback of the proposed 'diffusion rule' that aimed to block such overseas leasing arrangements.
• Stockpiling Chips: Some companies, like AI developer DeepSeek, managed to acquire and stockpile a significant quantity of Nvidia chips before the bans took full effect, allowing them to continue domestic training.
• Domestic Development: There is an intensified focus on developing indigenous AI chips, with companies like Huawei collaborating with firms such as DeepSeek to optimize and create next-generation Chinese AI chips. China has also reportedly discouraged local tech companies from purchasing advanced Nvidia chips and banned foreign AI chips in state-funded data centers.

Geographical Focus and Operational Details

Southeast Asia has emerged as a primary hub for this offshore AI training. Data center clusters in Singapore and Malaysia have experienced a boom in demand, fueled by Chinese companies. These data centers are typically owned and operated by non-Chinese entities, with Chinese firms entering into lease agreements to access the necessary computing infrastructure. Beyond Southeast Asia, some Chinese companies are also exploring data center access in other regions, including the Middle East.

While this strategy provides access to advanced hardware, it presents certain limitations. Chinese tech giants are generally restricted from moving private data out of China. This means that for AI models requiring customization based on specific local client data, training must still occur within China.

Implications for Global AI Competition

The move by Chinese tech firms underscores their determination to maintain competitiveness in the global AI landscape despite geopolitical challenges. Companies like Alibaba's Qwen and ByteDance's Doubao large language models have achieved top-tier performance globally, with their development partly attributed to this offshore training strategy. This dynamic highlights the complex interplay between national security interests, technological advancement, and global supply chains in the rapidly evolving field of artificial intelligence.