Chinese tech firms accelerate offshore AI training to bypass U.S. export controls and boost domestic chip development

Chinese technology giants are increasingly shifting their artificial intelligence model training overseas to navigate U.S. restrictions, while investing heavily in indigenous chip manufacturing to achieve technological self-reliance amidst geopolitical tensions.

Chinese technology giants such as Alibaba and ByteDance have increasingly shifted their artificial intelligence (AI) model training operations offshore to circumvent stringent U.S. export controls on advanced Nvidia chips. According to reports from the Financial Times, these companies predominantly lease data centres in Southeast Asian countries like Singapore and Malaysia, enabling them to access high-end Nvidia products while adhering to U.S. legal restrictions. This approach, facilitated by the Trump administration’s suspension of the “diffusion rule” previously enforced under President Biden, allows Chinese firms to exploit a regulatory loophole to continue utilising Nvidia chips despite the export bans. The crackdown on chip sales to China notably tightened in April, when U.S. restrictions specifically targeted Nvidia’s H20 chips, vital for training large AI language models. Chinese companies continue to use domestic resources to customise AI models in compliance with data sovereignty regulations, but offshore training has become their primary means to leverage cutting-edge hardware.^[1][2]

DeepSeek is a notable outlier in this landscape, as it reportedly manages to sustain its AI training domestically using a substantial stockpile of Nvidia chips procured before the imposition of export limits. The company also collaborates with Huawei engineers to optimise and develop China’s next generation of AI chips, reflecting a dual strategy: reliance on existing Nvidia technology for immediate needs and aggressive investment in indigenous chip development for future autonomy.^[1][2][7]

The Chinese government’s regulatory stance further complicates Nvidia’s role within China. Authorities have barred ByteDance from deploying Nvidia chips in new data centres, a measure designed to bolster domestic chip manufacturers like Huawei and others. This move aligns with broader national objectives to cut dependence on foreign technology amidst ongoing geopolitical tensions. Furthermore, the government mandates that state-funded data centre projects utilise only domestically produced AI chips, holding companies to stringent localisation requirements. In response, Nvidia has introduced China-specific versions such as the RTX6000D chip, although demand for these variants remains limited amid growing regulatory pressures and the government’s push for tech sovereignty.^[3]

Huawei itself is at the centre of China’s effort to expand its chip manufacturing capabilities. Despite U.S. export controls capping Huawei’s production of advanced AI chips at no more than 200,000 units in 2025, the firm has been promoting its Ascend 910C chips as alternatives to Nvidia’s offerings. Though Chinese AI chips currently lag behind leading U.S. counterparts by one to two years, massive investments exceeding $25 billion annually indicate a strong domestic commitment to closing this technological gap. Commerce Department officials warn that China’s progress in AI chip design and production should not be underestimated, underscoring the evolving complexities of global tech competition.^[5]

Complementing these industrial efforts, Chinese AI company DeepSeek recently released DeepSeek-V3.2-Exp, its latest large language model explicitly designed to support domestic hardware including Huawei’s Ascend Neural Processing Units (NPUs) and CANN software stack. This initiative marks a deliberate pivot away from Nvidia’s CUDA ecosystem, showcasing a significant stride towards AI hardware independence. The model incorporates innovations such as a sparse attention mechanism to optimise performance and cost-efficiency. Rapid industry-wide adoption and integration of this model by Chinese hardware vendors reflect a strong, coordinated push towards a diversified and self-sufficient AI ecosystem in China.^[6]

While Nvidia remains central for many Chinese tech firms, including DeepSeek’s continued use of Nvidia chipsets for its most advanced AI reasoning models, the broader trend signals a gradual divestment from U.S.-based technology. DeepSeek is reportedly evaluating AI GPU accelerators from several domestic manufacturers, including Huawei, Baidu, and Cambricon, for training smaller AI model variants as it seeks to reduce dependence on foreign processors. The company faces developmental challenges with its upcoming R2 AI model, which has delayed its launch, yet it intends to blend Nvidia and domestic chips strategically across its product lines.^[7]

Amid these complexities, there is some indication of potential loosening in U.S. export policy. The Trump administration reportedly is reconsidering whether to allow Nvidia to sell its advanced H200 AI chips to China, although no final decision has been made. Such a move would represent a significant shift and could influence how Chinese companies access and develop AI technologies, although it continues to evoke concerns in Washington about the military implications of advanced chip exports to China. Nvidia also faces pressure to remain competitive in China despite geopolitical headwinds.^[4]

In sum, China’s AI industry is navigating a complex terrain shaped by U.S. export controls, domestic regulations, and a determined push for technological self-reliance. While offshore training using leased data centres allows Chinese firms limited access to Nvidia chip technology, the government’s backing of domestic alternatives and collaborative efforts between AI companies and local chipmakers underscore a strategic priority to build an indigenous AI hardware ecosystem resilient to external pressures.

📌 Reference Map:

• ^[1] Silicon.co.uk – Paragraphs 1, 2, 4, 6
• ^[2] Reuters – Paragraphs 1, 2
• ^[3] Reuters – Paragraph 3
• ^[5] Reuters – Paragraph 4
• ^[6] Tom’s Hardware – Paragraph 5
• ^[7] Tom’s Hardware – Paragraphs 2, 6
• ^[4] Reuters – Paragraph 5

Source: Noah Wire Services