As export restrictions deepen, China is mobilising state capital, domestic innovation, and supply chain workarounds to secure its AI future, challenging US efforts at technological containment
On 17 September 2025, China’s Cyberspace Administration instructed major technology firms such as Alibaba and Tencent to cease purchasing Nvidia chips, including the RTX Pro 6000D and H20, citing alleged legal violations. This move places Chinese technology companies—including Huawei, Semiconductor Manufacturing International Corporation (SMIC), and Cambricon—at the forefront of domestic GPU development, positioning China as a direct competitor to the United States (US) in the graphics processing unit (GPU) market.
Beijing’s countermeasures began with cancelling orders for the RTX Pro chip designed specifically for the Chinese market and discouraging the procurement of the H20. In November 2025, authorities further barred the use of foreign chips in state-led data centre projects and critical infrastructure, particularly in projects that were less than 30 percent complete, while subjecting other cases to review on a case-by-case basis.
These measures signal a strategic push towards long-term self-reliance in artificial intelligence (AI) through domestic firms. The approach underscores the importance of assessing its implications for both US and Chinese GPU supply chains. China’s systematic efforts to build domestic capabilities are exemplified by the Yizhuang AI Zone in Beijing, which has pledged 80 billion yuan (approximately US$11 billion) to develop Nvidia-free AI infrastructure by 2025. This ambitious project focuses on the RISC-V open instruction set processor architecture, aiming to replace foreign-controlled instruction set architectures and mitigate vulnerability to foreign sanctions.
China’s broader technological initiatives are not solely state-corporate centric but are also supported by academic and industrial sectors. Comparable to deep-learning programmes that rely on frameworks such as TensorFlow or PyTorch running on Nvidia GPUs, China is developing domestic alternatives across the AI stack. Tsinghua University has developed the Chitu (赤兔) AI inference framework, which has reportedly reduced GPU dependency by 50 percent while increasing processing speeds by over 315 percent (more than threefold) when tested on the DeepSeek-R1 model. Crucially, Tsinghua’s Chitu is compatible with domestic alternatives such as Huawei’s Ascend 910B, reflecting a systematic strategy to replace foreign components across the AI supply chain.
SMIC and local suppliers like Shanghai Micro Electronics Equipment Group (SMEE) have been compelled to push deep ultraviolet (DUV) technology to its technical limits. Although this has allowed them to continue competing with Nvidia, it has resulted in much higher costs and lower yields.
State-backed programmes further reinforce these efforts. China is advancing “AI Plus” integration across six key sectors by 2027, positioning AI as a central pillar of future growth. Realising this vision depends on the procurement and development of advanced semiconductors. In this context, a significant milestone was achieved by Huawei and SMIC in 2023 with the production of 7 nm chips, demonstrating that China can reach critical technological thresholds despite restrictions; however, manufacturing efficiency yield rates continue to lag behind industry leader TSMC.
For Beijing, achieving true self-sufficiency remains a formidable challenge due to the complex nature of the GPU supply chain. Export restrictions imposed by Washington have specifically targeted critical bottlenecks, particularly the prohibition on ASML’s extreme ultraviolet (EUV) lithography machines, which are crucial for manufacturing cutting-edge chips (7 nm or smaller). Consequently, SMIC and local suppliers like Shanghai Micro Electronics Equipment Group (SMEE) have been compelled to push deep ultraviolet (DUV) technology to its technical limits. Although this has allowed them to continue competing with Nvidia, it has resulted in much higher costs and lower yields.
The Trump administration 2.0 further tightened restrictions in April 2025 by prohibiting Nvidia from exporting the Hopper H20 series to China, following earlier curbs on advanced chips such as the H100, H800, and the downgraded H20 tailored for the Chinese market. Nvidia’s CEO, Jensen Huang, cautioned that escalating sanctions have inadvertently spurred domestic innovation in China, prompting companies to bolster local alternatives. Demand once met by Nvidia is increasingly being redirected to Huawei, whose Ascend 910C has partially filled the gap. However, the restriction was reversed in July 2025, after which a licence was issued permitting exports to resume.
Huawei's Ascend series, while effective for inference computing, still lags behind Nvidia in pre-training large AI models. Its main shortcoming is the absence of a mature software ecosystem akin to Nvidia’s CUDA, which has been honed for more than a decade. This gap leads to persistent bugs and optimisation challenges. Nevertheless, China is compensating by concentrating on large-scale systems that integrate a vast number of GPUs. Huawei's CloudMatrix 384 (CM384), powered by the Ascend 910C, exemplifies this strategy, utilising enhanced inter-GPU connectivity to mitigate hardware limitations.
China has responded with a comprehensive state-driven approach that integrates subsidies, innovations, and strategies to circumvent sanctions.
Yet critical dependencies persist. Domestic GPU production still heavily relies on imported high-bandwidth memory (HBM), as local supplier ChangXin Memory Technologies (CXMT) has not yet met the necessary global standards. Before US sanctions were expanded under President Biden, Huawei secured substantial HBM imports from Samsung, approximately 13 million units, sufficient to support 1.6 million Ascend chips. However, the future remains uncertain as supply routes become increasingly constrained.
China has responded with a comprehensive state-driven approach that integrates subsidies, innovations, and strategies to circumvent sanctions. The government has unveiled a US$47 billion subsidy initiative aimed at accelerating GPU research and strengthening companies like SMIC, which serves as an essential connector or “wafer bridge” between restricted and unrestricted fabrication facilities. Meanwhile, Huawei has relied on a “shadow fab network” and trailing-edge innovation to sustain production capabilities.
China employs layered strategies, including shell companies, offshore diversion routes through hubs in Malaysia and Singapore, and regulatory loopholes such as renaming nodes (for example, rebranding 18nm DRAM as 19nm). Together, these measures mitigate sanctions pressure and sustain Huawei’s and SMIC’s capacity to produce advanced chips like the Kirin 9000.
China’s stronghold in legacy semiconductor nodes, such as 90nm chips that power most everyday electronics and IoT devices, provides both revenue and strategic leverage. Profits from these trailing-edge markets subsidise higher-end chip development, while multi-patterning DUV lithography helps navigate EUV restrictions. Concurrently, China employs layered strategies, including shell companies, offshore diversion routes through hubs in Malaysia and Singapore, and regulatory loopholes such as renaming nodes (for example, rebranding 18nm DRAM as 19nm). Together, these measures mitigate sanctions pressure and sustain Huawei’s and SMIC’s capacity to produce advanced chips like the Kirin 9000.
The evolving policies in Washington and Beijing show that control over GPUs and advanced computing has become central to national security threat perceptions, shaping the future of the US-China chip rivalry. China’s technological strategy has moved beyond exploiting technological gaps towards more strategic and systematic responses, including supply chain balancing and stronger regulatory oversight. US policies have struggled to fully contain Beijing’s innovation momentum, prompting Chinese countermeasures such as restricting RTX chips and discouraging the use of Nvidia processors in certain data centres. Coupled with stronger coordination with allies, China’s dominance in trailing-edge production is likely to support progress towards advanced nodes by 2027.
This situation is reshaping global innovation pathways, straining supply chains, and heightening security risks. US intelligence assessments, such as Microsoft’s Digital Defense 2023, which documents Chinese espionage targeting AI models, underscore the broader strategic stakes. Ultimately, the semiconductor race reflects how mastery of advanced computing has become central to strategic dominance in the 21st century. China aims to secure its AI future through a range of state-led initiatives and industrial policies. However, the US and major technology firms are likely to respond cautiously to Beijing’s strategic moves. Taken together, China’s countermeasures and domestic efforts suggest a determined push towards self-reliance rather than continued dependence on Western technology. If China continues to expand its computing capabilities at scale, defining the limits of its technological ascent will become increasingly difficult.
Nistha Kumari Singh is a PhD scholar and Dr TMA Pai Fellow in the Department of Geopolitics and International Relations at MAHE.
Vanni Filloramo is a PhD candidate in Social Sciences (Politics, Society, and Technology) at the University of Macerata, Italy.
The views expressed above belong to the author(s). ORF research and analyses now available on Telegram! Click here to access our curated content — blogs, longforms and interviews.
Nistha Kumari Singh is a PhD scholar and Dr TMA Pai Fellow in the Department of Geopolitics and International Relations at the Manipal Institute of ...
Read More +
Vanni Filloramo is a PhD candidate in Social Sciences (Politics, Society, and Technology) at the University of Macerata, Italy. His research focuses on the intersection ...
Read More +
PREV
