[China AI Bottleneck] “Still Running on Nvidia Chips”: China’s AI Self-Sufficiency Drive Falters Despite Massive Investment as Hardware, Software and Ecosystem Constraints Persist

China’s AI industry struggles to adopt domestic chips amid software ecosystem deficiencies
“Four Huawei accelerators needed to match one Nvidia chip,” underscoring hardware performance gap
Beijing sustains massive state support, but the industry’s ecosystem foundations remain fragile

Chinese artificial intelligence developers are training cutting-edge models using Nvidia chips. Despite the Chinese government’s aggressive push for semiconductor self-sufficiency, the hardware performance and software ecosystem gaps between domestic chips and Nvidia products have proved stubbornly difficult to close. A shortage of high-quality Chinese-language training data, data distortions caused by internet censorship and rising costs stemming from the proliferation of AI services have also emerged as major constraints on the growth of China’s AI industry.

Chinese AI Industry’s Reliance on Nvidia

According to the Hong Kong-based South China Morning Post (SCMP) on Aug. 12, developers at China’s leading AI laboratories and companies continue to use Nvidia hardware as the standard for training the latest frontier models. They identified the software ecosystem gap as the most decisive obstacle to chip localization. Nvidia’s parallel computing platform, CUDA, has served as a de facto standard across the global AI industry for years. If Chinese companies align with Beijing’s policy agenda and migrate their development environments to Huawei’s proprietary Compute Architecture for Neural Networks (CANN), they must rewrite and optimize their existing Nvidia-based training pipelines and code from scratch.

James Wang, an AI development researcher at a major university laboratory in Shanghai, told SCMP, “Our existing training pipeline is entirely dependent on CUDA and cannot run directly on Huawei Ascend chips.” He added, “Migrating the existing workflow to Huawei chips would increase the time and cost required for a project by at least 50%.” According to reports from The Information and other foreign media outlets, Kimi K3, the latest model developed by Chinese start-up Moonshot AI, which recently stunned the global AI market with rapid performance gains, was also reportedly trained on an Nvidia computing cluster that included the company’s latest Blackwell processors.

Persistently Weak Hardware Efficiency

Chinese chips also lag conspicuously behind Nvidia products in hardware performance. Spanish technology publication Xataka reported last month, citing approximately four hours of a DeepSeek investor meeting released by Tencent Tech, Tencent’s official IT and technology news outlet, that DeepSeek founder and Chief Executive Officer (CEO) Liang Wenfeng identified “available computing resources” as the greatest difference between the Chinese and U.S. AI industries. Liang argued that the two countries showed no fundamental disparity in AI talent, but a widening gulf remained in the scale of resources available for actual model development, including capital and advanced chips. Liang illustrated this reality through the performance gap between Nvidia and Huawei AI accelerators.

According to Liang, training a large AI model with approximately 800 billion active parameters would require 50,000 Nvidia accelerators based on the GB300. Using Huawei’s next-generation Ascend 950 for the same task would require 200,000 units. Huawei accelerators would therefore need to be deployed at roughly four times the volume of Nvidia products to process the same computational workload. Liang also said that, provided the chips could be secured at a reasonable price, purchasing Nvidia accelerators would be a better use of cash than keeping the money in a bank. His assessment appears rooted in the technical difficulties DeepSeek previously encountered while adopting domestic chips. The Financial Times (FT) reported last year that DeepSeek had attempted to train its next model using Huawei Ascend chips but encountered problems involving stability, chip-to-chip connectivity and software, prompting the company to restrict Huawei chips to inference workloads.

Beijing Sustains Aggressive Investment

The Chinese government continues to deploy massive investment to overcome these constraints. The National Integrated Circuit Industry Investment Fund, commonly known as the Big Fund, forms the central financing pillar of China’s semiconductor self-sufficiency policy. Over the past decade, the Big Fund has worked to strengthen the underlying capabilities of China’s semiconductor industry. According to the China Semiconductor Industry Association (CSIA), the first phase of the Big Fund, launched in 2014 with $19.2 billion, financed 75 projects across 23 companies. Major beneficiaries included foundries Semiconductor Manufacturing International Corporation (SMIC) and Hua Hong Group, packaging company JCET, equipment manufacturer Naura Technology and electronic design automation company Empyrean Technology. The second phase, established in 2019 with $29.5 billion, focused on diversifying supply chains in response to U.S. export restrictions. Its largest single investment was a $1.5 billion injection into SMIC. The fund also holds an 8.73% stake in ChangXin Memory Technologies (CXMT) and an 11.38% stake in Yangtze Memory Technologies Corporation (YMTC).

The third phase of the Big Fund, established in 2024, is concentrating its resources on securing competitiveness in advanced semiconductors. Huaxin Dingxin, a private equity fund (PEF) spearheading investment in advanced packaging, invested $44.2 million in Tianshui Xinyuan Technology, which specializes in automotive image-processing technology, to acquire a 31.58% stake. It also secured an 18% stake in Anhui Juhe Microelectronics, which possesses chiplet substrate technology. Equipment investment is handled by Guotou Jixin, a fund in which the third phase of the Big Fund owns a 99.9% stake. A leading example is its $66.3 million investment in Tuojing Jianke, a hybrid-bonding equipment developer spun off from chemical vapor deposition equipment leader Piotech. In the AI computing chip sector, investments are being diversified to spread risk. Major recipients include WinSilicon, which develops ultra-high-speed interconnect technology; QingMicro, which specializes in computing-in-memory technology; and Beijing Aijieke Xin, a developer of generative AI accelerators.

Table 1. Investment Strategies Across the Three Phases of China’s Semiconductor Big Fund

Training Data Supply Approaches Exhaustion

Whether this aggressive government investment will translate into substantive industrial growth remains uncertain. The ecosystem underpinning the AI industry has yet to establish sufficiently robust foundations. SCMP recently cited Chinese AI experts and researchers as warning that a shortage of high-quality training data could emerge as a new bottleneck for China’s technological ambitions, eclipsing even the shortage of hardware. As the supply of high-quality, publicly available text produced by humans steadily dwindles, China’s AI industry faces an especially acute crisis due to the distinctive linguistic structure of Chinese.

According to web analytics company W3Techs, Chinese accounts for just 1.3% of global web text data. The figure falls far short of Spanish at 6%, German at 5.9% and Japanese at 5%. The disparity is also evident in the latest data from Common Crawl, which directly collects webpages. Chinese was the primary language in only 4.43% of the HTML documents collected by Common Crawl last month. The volume of data available to train high-quality Chinese-language AI models therefore remains severely inadequate. Large language models (LLMs) typically improve their performance by learning factual relationships, logical structures, specialized knowledge and cultural context from training data. An insufficient supply of source data inevitably limits efforts to enhance a model’s reasoning capabilities and accuracy. The shortfall can be supplemented with translations of English-language material or synthetic data generated by AI, but such methods carry significant risks of distorting Chinese-specific meanings and contexts or repeatedly training models on errors and biases inherited from existing systems.

Systemic and Cost Constraints Come Into View

Systemic constraints are equally evident. China’s internet censorship apparatus, known as the Great Firewall, stands at the center of these limitations. Since the late 1990s, China has imposed extensive restrictions on foreign internet services and incoming information while controlling the search and distribution of politically sensitive content. Information concerning the Tiananmen Square crackdown, the Cultural Revolution, Xinjiang’s Uyghur population, pro-democracy movements and criticism of the country’s leadership constitutes a primary target of these controls. Such restrictions degrade the quality and composition of the data available for AI training while weakening AI systems’ capacity to understand the real world. One market expert said, “Under the control of the Chinese Communist Party, AI faces consequences that extend far beyond its inability to answer politically sensitive questions.” The expert added, “Analyzing complex social and economic issues requires simultaneous consideration of competing hypotheses, counterarguments and cases of failure. When the training data itself is skewed toward a single perspective, the model becomes accustomed to reproducing established narratives, weakening its ability to analyze reality from multiple angles.”

Cost pressures are also becoming increasingly visible. The AI industry carries a cost profile in which server and electricity expenses rise alongside user growth. Chinese AI companies that built their market positions around price competitiveness have recently begun to falter under the strain of rapid user expansion. Z.ai, one of China’s leading AI start-ups, recorded revenue of $106.7 million last year, up 131.9% from the previous year. Its net loss widened by nearly 60% over the same period, rising from $436 million to $695.3 million. Even cash-rich technology giants are retreating from free and low-cost AI strategies. Alibaba raised prices for some premium AI services by as much as 34%, while ByteDance introduced a three-tier paid subscription plan for its Doubao AI chatbot priced at $10, $29 and $74 per month. Moonshot AI released Kimi K3’s weights, programming code and design information to the public last month while simultaneously imposing separate contractual requirements on model-service providers that generate more than $20 million in revenue over the preceding 12 months and offer Kimi K3 to third parties.