Executive Summary
Recent commentary often frames the U.S.-China AI competition as a simple contrast between the American pursuit of frontier artificial general intelligence (AGI) and China’s emphasis on large-scale diffusion of practical AI applications. This paper argues that Chinese AI elites view the competition as a multidimensional contest spanning the entire AI ecosystem, not a choice between frontier innovation and broad deployment.
Facing the constraints of U.S. export controls and restrictions, Beijing is pursuing a strategy analogous to the tai-chi concept of “using four ounces to deflect one thousand pounds” (四兩撥千斤). China’s strategy is not to match U.S. capabilities dollar-for-dollar or chip-for-chip, but to compete across the supply chain in every way it can. Together China’s approach supports a “frugal AI stack”: an ecosystem designed to achieve competitive outcomes without requiring parity with the United States in every technological domain.
This paper analyzes this strategy from the perspective of China’s AI elite, with a particular focus on industry and technology leaders in the private sector. We find Western analysis often overlooks the nuance of internal Chinese debates on AI, particularly in areas these elites acknowledge China faces challenges. Despite the rise in Western pronouncements of China’s impending victory in AI, Chinese industry leaders are clear-eyed about its weaknesses that are often overlooked. Across six layers of the AI stack – market applications, models, compute, infrastructure, talent, and financing – Chinese actors are competing in every avenue they can to offset the constraints of U.S. restrictions.
This does not make China any less of a formidable competitor in AI, and China’s many strengths require allied countries to adopt joint efforts across the full AI stack to counter China’s AI diffusion. To meet this strategy, allied policymakers must also compete with China on all fronts. Doing so requires an economic security strategy that does not merely focus on the most advanced computing chips and the most advanced frontier models, but on all the domains of AI competition.
Figure 1. China’s AI Strategy Across the Stack.
1. Market Applications: China enjoys advantages in consumer and industrial AI but lags in enterprise adoption
China is aggressively expanding in both consumer AI and industrial AI markets, but faces difficulty persuading enterprise users to adopt its enterprise AI solutions. Leveraging the vast user bases of major tech firms such as Alibaba, Tencent, and ByteDance, China continues to diffuse consumer AI globally.
In the enterprise and industrial domains, Chinese firms show a stronger willingness to invest in hardware-based embodied intelligence, while their cost-minimization culture reduces demand for software-driven enterprise AI solutions. China holds advantages in consumer AI where user experience is more important than top-of-the-line compute. But Chinese industry leaders believe that the long-term development of China’s AI sector depends on whether its B2B market can achieve sustained growth .
Table 1. Select Industrial Applications for Chinese AI.
2. Models: Reducing Costs and Building Ecosystems with Open-Weight
China’s approach to models includes two elements: first, a commitment to open models as part of a strategy to speed development and diffusion, and second, using model training optimizations and distillation to quickly achieve leading models under compute constraints. But both approaches have their drawbacks and limitations that Chinese elites acknowledge.
First, pursuing open models has become the consensus in China’s AI community to accelerate development under resource constraints. Open-weighted model allows firms to rapidly gather community feedback and build developer ecosystems that depend on Chinese model platforms. The ability to download and fine-tune open models freely is an advantage for their global diffusion. Lower API prices also help: Bytedance’s enterprise model was priced at approximately USD $0.00011 per 1,000 tokens, nearly 100 percent cheaper than the industry average.
By the end of 2025, Chinese open models achieved 30% global AI market share. However, Robin Li, CEO of Baidu, and Yang Zhilin, founder of Moonshot, both argue that China must pursue both proprietary and open models, as proprietary firms attract the most talent and capital, and that long-term, closed-source technologies are better business models.
The second part of China’s approach to models is “reducing costs and increasing efficiency” (降本增效): innovations in model training to create more efficient, cost-effective models. But firms like DeepSeek and MiniMax still rely on more illicit means including distillation attacks on U.S. firms including OpenAI and Anthropic.
3. High-End Training Compute: Still the Key Bottleneck
Despite limited breakthroughs in domestic chips, China is still highly reliant on Nvidia chips and the CUDA parallel computing framework. China has adopted a “quantity for quality” strategy for its domestic GPUs while trying to quickly develop a domestic AI hardware and software ecosystem for training AI models, to varying degrees of success.
In hardware, China has been restricted by bottlenecks, including domestic foundry limitations and lack of advanced lithography machines. In response, Huawei is focusing on stacking large numbers of chips together to achieve the necessary computing power. Its CloudMatrix 384, scheduled to be launched in 2026, reportedly achieved an aggregate dense compute that rivals or exceeds Nvidia’s GB200 NVL72 in certain metrics.
In software, Wei Shaojun, Vice Chairman of the China Semiconductor Industry Association, has noted that China’s dependence on NVIDIA’s CUDA architecture is a critical dependence and that China must establish a local chip design ecosystem. For example, Huawei’s CANN and Moore Threads’ MUSA are touted as alternatives to CUDA. Nevertheless, Chinese firms still rely heavily on Nvidia chips. DSET’s previous report also finds that the Chinese AI industry leverages data centers in Singapore and Malaysia to access advanced computing power restricted under U.S. controls.
Table 3. Comparison of Huawei and Nvidia AI Compute Stack.
4. Infrastructure: Challenges Despite Having the World’s Largest Power Grid
If compute is China’s most acute disadvantage, China’s energy output has the potential to be a significant advantage. In 2024, China generated more than twice the electricity of the U.S., which could allow it to pursue the “quantity for quality” compute strategy, which once seemed economically unviable. To coordinate compute resources nationwide, the Ministry of Industry and Information Technology (MIIT) has also established a National Computing Power Internet Service Platform, which enables users to publish demand and find compute, models, data, and applications.
However, Chinese stakeholders in cloud computing and infrastructure note the challenges in translating energy output into usable advantages for actual AI use, citing logistical, technological, and structural problems preventing an efficient allocation of AI infrastructure.
5. Talent: Domestic AI Talent Limited by “Drill-Based Learning” Culture.
The paradigm of talent cultivation has shifted from “recruiting returnees” to “growing our own.” China produces approximately 1.4 million STEM graduates annually, more than six times the U.S. output. At the same time, Beijing has established institutional infrastructure for talent cultivation, such as the Beijing Academy of Artificial Intelligence (BAAI), which binds top universities (Tsinghua, Peking University) to industrial actors (Huawei, Zhipu AI).
However, Chinese AI elites worry that the key problem is not lack in talent but lack in innovation. Yao Shunyu, Chief AI Scientist of Tencent, is concerned that Chinese prefer doing safer things rather than exploring the unknown. Being aware of this constraint, Beijing has reformed evaluation systems to de-emphasize leaderboard rankings and attracts global frontier researchers through the “K visa” program, expecting to promote innovation culture in China.
6. Financing: Patient State Capital and the Pragmatic Market
Figure 4. China’s AI Public Financing Structure
To maximize returns on their investments, Chinese venture capital firms actively help AI companies connect with upstream and downstream industry chains. However, Chinese AI investment still encounters challenges, namely that U.S. private sector funding of AI of $109.1 billion in 2024 dwarfed China’s $9.3 billion by a factor of twelve. To encourage the private sector to invest in strategic AI startups, China is using its state financing playbook of government guidance funds to funnel private funding to priority areas.
The Ministry of Finance established a RMB 100 billion National Venture Capital Guidance Fund emphasizing early-stage, hard-tech enterprises with long-term development at the end of 2025. Private funding is most active downstream in the AI stack, with private funding driving recent chip design and AI model IPOs. Chinese elites argue, however, that further reforms are needed in the venture capital system to more effectively manage funding risks.
Policy Implications
1. Market Applications: Allied policymakers must move beyond frontier AI leadership to secure dominance in the global AI ecosystem. To achieve this goal, they should promote the U.S.-aligned model adoption and strengthen data governance against cross-border risks.
2. Models: Allied governments should restrict using Chinese open-weight models in sensitive sectors while promoting trusted alternatives globally. In addition, allied governments should educate industry on the risks of using Chinese open-source models without imposing outright bans that could hinder innovation.
3. Compute: Policymakers must uphold existing export controls while continuously updating controlled commodities and entities. To prevent China from enhancing chip performance through stacking and other workaround strategies, these controls should focus on critical technological chokepoints, including advanced-process GPUs, high-bandwidth memory (HBM), high-speed interconnect technologies, and related manufacturing equipment. In addition, the U.S. and its allies should strengthen oversight and transparency across global AI infrastructure to prevent China from accessing compute through third-country data centers.
4. Infrastructure: The United States is facing similar energy constraints and inefficiencies in deploying energy and AI infrastructure at scale. Allied policymakers should invest in new energy AI infrastructure and provide tax incentives and subsidies to promote data center development. At the same time, they should deepen integration across allied compute ecosystems to sustain long-term competitiveness.
5. Talent: the U.S. should prioritize retaining global STEM talent through positive incentives, while BIS should restrict U.S. persons engaging in sensitive Chinese AI development and strengthen EAR enforcement.
6. Financing: to maintain leadership in AI competition, the U.S. and its allies should develop analogous capital mechanisms through public funding and partnerships to sustain long-term AI R&D. Moreover, it should prevent capital flows into China’s AI sector via outbound investment restrictions and tools like the non-SDN Chinese Military-Industrial Complex Companies List.
