The Current State of China’s AI Industry

In recent years, artificial intelligence (AI) has become the core engine of a new round of technological revolution, penetrating deeply into the fabric of human production and life, profoundly reshaping the global economic structure, innovation paradigms, and social governance logic. China has entered the first tier of global AI development and is at a critical opportunity period for transitioning from catching up to leading. In the face of increasingly fierce international competition and the intrinsic demand for high-quality development, we conducted field research to understand the current state of China’s AI industry, its momentum, and its weaknesses.

Current Development Trends in China’s AI Industry

General Secretary Xi Jinping pointed out that “AI is a strategic technology leading this round of technological revolution and industrial transformation, with a strong ’leading goose’ effect.” AI is not merely a linear iteration of a single technology or a partial upgrade of an industry; it represents a comprehensive and disruptive reconstruction of the underlying logic of economic and social operations. To assess its development level and trends, we must break away from traditional technical evaluation and industrial analysis frameworks and conduct a comprehensive analysis from dimensions such as technical capabilities, industrial scale, factor support, and integrated applications.

From a technical capability perspective, AI technology led by open-source has achieved collective breakthroughs, forging new standards within the global developer network. During our research at a laboratory, we observed that the research team introduced a self-criticism mechanism for AI, which significantly improved the accuracy of complex programming problem-solving after multiple rounds of self-competition without human intervention. AI has advanced from “being able to listen and see” to “thinking, reasoning, planning, and mastering how to learn.” Overall, China’s gap with international top levels in key indicators such as model performance, training efficiency, and multi-modal integration continues to narrow, with some fields already achieving parity or leading. By 2025, China’s share of global downloads of open-source models is expected to reach 17.1%. Recent statistics show that eight of the top ten global open-source models come from China. The performance of the DeepSeek-V4 model is on par with international top models, while the application programming interface (API) price is less than 1% of the GPT-5.5 model. This signifies a break from the technological monopoly of a few tech giants, enabling millions of global developers to conduct secondary development based on Chinese open-source models. Open-source not only benefits developers but also accelerates knowledge flow and spillover, continuously forging self-evolution capabilities within China’s AI technology in an open ecosystem.

From an industrial scale perspective, the AI industry has experienced nonlinear explosive growth, with significant value spillover effects behind the trillion-dollar blue ocean. By 2025, the global AI market size is expected to reach $757.58 billion, and China’s core AI industry size will exceed 1.2 trillion yuan. The significance of this 1.2 trillion yuan lies not only in the number itself but also in the growth logic behind it. Traditional industries follow the iron law of linear input and diminishing marginal returns, while AI breaks this curse, with technological breakthroughs and application diffusion mutually reinforcing, forming a positive feedback loop of “the more you use, the stronger it becomes.” Our research found that Beijing, as an innovation source, will reach a core AI industry size of 450 billion yuan by 2025, with a batch of mature algorithm models acting as “digital technology pumps,” continuously delivering intellectual energy to factories in Hebei, ports in Tianjin, and pastures in Inner Mongolia. Shanghai is leveraging the “Mold City” initiative to build an ecological attraction through the “Mold Speed Space.” Shenzhen is focusing on industrial implementation, aiming to create a highly concentrated and precisely serviced enterprise ecosystem for the real economy. Ultimately, the AI industry exhibits a clear multiplier effect, where an investment of one yuan can leverage several yuan; the trillion-dollar scale is supported by a full industrial chain from underlying computing power to upper-level applications, from core algorithms to intelligent terminals, giving rise to new services, new divisions of labor, and new markets.

From a factor support perspective, China’s core AI resources have achieved a strategic leap, with institutional innovation accelerating the release of factor vitality. The competition in AI not only depends on how fast models can run but also on how solid the computing power foundation is and how smoothly data flows. China has established significant scale advantages in these two core resources. In terms of computing power, 42 intelligent computing clusters have been built, and as of the first quarter of this year, the scale of intelligent computing power reached 1,882 exaFLOPS, ranking among the top globally. In terms of data, there are over 100,000 high-quality datasets nationwide, with a total volume exceeding 890 petabytes, equivalent to 310 times the total digital resources of the National Library of China. Moreover, institutional advantages are gradually becoming apparent. In Beijing’s data foundational system pilot area, a “regulatory sandbox” mechanism has effectively broken the deadlock regarding enterprises’ reluctance to open their resources, allowing them to enter a protected “experimental field” for integrated training without transferring data ownership. A technical leader from a company stated, “Previously, training with our small data led to increasing bias; now, the sandbox gathers real data from over ten industries, significantly improving accuracy, making data more valuable the more it is used.”

From the perspective of integrated applications, China’s AI is accelerating its penetration into various industries, with the breadth of applications and depth of integration building new global competitive advantages. By the end of 2025, the CNC rate of key processes in major industries in China is expected to reach 68.6%, with AI integration applications transitioning from “spot blooming” to “full-chain intelligence.” First, the scope of penetration continues to expand, covering the vast majority of major categories of the national economy, with benchmark applications emerging in manufacturing, healthcare, transportation, finance, and energy. Second, the level of empowerment has significantly improved, advancing from auxiliary roles to core areas such as research and development, production, and operational management. In a heavy equipment manufacturing company in Shandong, we saw an industrial large model system comprehensively taking over the entire process from blueprint analysis, process planning to quality inspection, reducing the time for new process design from several weeks by multiple senior engineers to less than 72 hours, with a 5% increase in yield rate. Third, new business formats and models are emerging rapidly, with intelligent connected vehicles, AI pharmaceuticals, and embodied intelligent robots flourishing, continuously forming new trillion-level industrial tracks. Throughout the research, it was deeply felt that in this global intelligent competition, the party with the richest application scenarios, the closest integration, and the most intensive industrial feedback will master the standards and paradigms of how AI is used, where it is used, and how deeply it is used, thus gaining the initiative in the intelligent era.

Challenges and Problems Facing China’s AI Industry Development

Currently, the global AI technology competition is becoming increasingly fierce, and China’s AI industry development is at a critical juncture of application leadership, foundational catch-up, and ecological breakthrough. Facing external pressures such as computing power blockades and talent competition, we still have many “bottleneck” links and points of blockage, from high-end chips to foundational algorithms, from original innovation to industrial transformation.

International competition is squeezing the development space of the AI industry. Our research found that some Western countries have upgraded their policies towards China from single technology restrictions to systematic ecological blockades. First, the “hard” blockade is intensifying. The U.S. has continuously increased restrictions on the sale of AI chips to China, forcing many domestic innovation teams to slow down the development pace of large models due to “computing power hunger.” Second, “soft” ecological barriers are being constructed. NVIDIA’s graphics processing units (GPUs) occupy over 90% of the global market share, and its unified computing device architecture (CUDA) ecosystem has formed a closed-loop system of “hardware + software + developer community” after more than ten years of accumulation. We learned from a domestic chip company in Shanghai that although its hardware computing power indicators are close to international mainstream levels, customers are most concerned about whether it can be compatible with CUDA. The crux is that chip replacement is not a simple hardware swap but involves a complete technical stack migration, including development frameworks, operator libraries, debugging tools, and development habits. Millions of developers are deeply bound to the CUDA ecosystem, and the high migration costs and long adaptation cycles pose obstacles to large-scale applications, even if domestic alternatives meet performance standards. Third, the competition for rule-making and discourse power is intense. Global AI technology standards, governance norms, and cross-border data rules are largely dominated by Western countries. At the beginning of 2025, the DeepSeek large model caused a stir in the global market due to its technological breakthroughs, prompting several Western countries to issue bans or initiate strict reviews. The reality warns us that technological leadership does not guarantee market access; lacking discourse power can restrict the international expansion of the industry.

Large models face reliability crises in specialized scenarios. While large models perform impressively in general dialogue, their capability deficiencies become apparent when entering fields such as industrial inspection, medical diagnosis, and financial risk control, where precision and reliability are critical. A manufacturing company reported that its AI visual inspection system misjudged good products as defective due to slight changes in lighting, leading to defective products being released and requiring manual re-inspection. “Stunning during demonstrations but failing on production lines” has become a true reflection of AI deployment in many enterprises. The issue lies in the generalization ability exhibited by large models in open-domain tasks, which does not naturally transfer to specialized scenarios with near-zero tolerance for errors. The gap between “being able to talk” and “being able to use reliably” is a significant engineering hurdle. The “hallucination” problem is also concerning. In general scenarios, such errors may be minor flaws, but in medical dosages, legal judgments, and financial risk control, every instance of “seriously talking nonsense” could trigger irreparable risks. This exposes a fundamental flaw in large models: they are essentially pattern matchers rather than logical reasoners. Transitioning from “being able to speak” to “speaking the truth,” and from “guessing answers” to “understanding causality” is a threshold that the industry must cross for deeper development.

High-quality datasets still struggle to meet model development needs. Our research found that a common issue is that while there is an abundance of “raw oil” data, the “refining” capability is insufficient. The scale of globally available private data far exceeds that of public data, but due to institutional barriers such as non-unified data standards, inadequate authorization mechanisms, and unclear compliance boundaries, a large amount of high-value data is trapped in “data islands.” Although China possesses massive data resources, the data truly usable for large model training is severely lacking. In globally common training datasets of 5 billion, the proportion of Chinese corpus is only 1.3%. Furthermore, the bottlenecks in data circulation hinder China’s data scale advantage from being fully transformed into core competitiveness. Copyright and legal risks are also on the rise. A company expanding overseas reported that its video generation model was accused of unauthorized scraping of overseas platform videos for training, resulting in a collective lawsuit abroad. If data sovereignty and copyright barriers evolve into new trade weapons, they could cut off domestic enterprises’ legal access to high-quality international data resources.

The commercial closed-loop for AI applications has yet to be established. The AI industry application is at a crossroads of transitioning from policy-driven to market-driven, with sustainable business models still under exploration. First, the “gears of the industrial chain are misaligned.” The computing power layer is expensive and insufficiently compatible with models, the model layer is general but weak in industry customization, and the application layer consists mainly of single-point tool-type products that do not communicate with each other, lacking effective engagement mechanisms among the three links of computing power, models, and applications. Second, the profit model for enterprises is unclear. Domestic users have not yet formed a payment habit, and many application companies can only rely on project-based contracts to sustain themselves or depend on government subsidies for “blood transfusions.” The transition from “policy blood transfusion” to “market blood production” is key to whether the industry can emerge from its nurturing phase. Third, scaling products for replication is challenging. An industrial AI founder admitted, “Three factory pilot projects were successful, but when the client asked to change the production line, the solution became useless. Without standardization, there can be no scaling; without scaling, there will always be cash burn.” The difference between a “model room” and a “commercial building” is not just in single technologies but in a standardized product system that is configurable, replicable, and maintainable, with the prerequisite being standardized interfaces formed among all links of the industrial chain.

Accelerating the Development of China’s AI Industry Requires a Systematic Collaborative Battle

AI is an extremely special general-purpose technology, distinctly different from any frontier technology in historical technological revolutions. First, it has a strong path dependence and ecological lock-in effect. The underlying chips define the form of computing power, the intermediate frameworks determine the development paradigm, and the upper applications deeply rely on the interface standards of the former two—this highly coupled technical architecture means that once a first mover gains dominance at any level, it can penetrate upwards and downwards, locking the entire industrial chain into its ecological system. Second, competition has evolved into a systematic game where every link is interconnected. Traditional technological competition focuses on single technologies, where overcoming one can lead to breakthroughs; however, AI competition encompasses a full-dimensional race covering chips, frameworks, data, applications, and rules, where any shortcoming in one dimension can become the “Achilles’ heel” of the entire system. Third, the diffusion cycle has been drastically compressed. The electrical revolution took a century to fully penetrate society, and information technology took half a century to reshape business forms; however, AI is rewriting the underlying logic of industries at a speed of instantaneous emergence, penetration, and transformation, significantly accelerating the conversion of first-mover advantages into lock-in advantages, leaving little reaction time for followers. In this global competition that determines the future, we face not just a “bottleneck” at a specific technology point but a full-stack competition from underlying hardware to upper-level ecosystems, from technical standards to governance rules. To break the deadlock and seize the initiative, we must engage in a “full-factor + full-ecosystem” systematic collaborative battle. We need to ensure that various factors such as computing power, data, algorithms, and scenarios flow freely, stimulate the innovative vitality of diverse entities such as enterprises, universities, research institutions, and developer communities, and align all forces under a national strategy to form a collective effort.

In recent years, Hebei has become an important node in the national computing power industry layout, accelerating the construction of a leading computing power industry ecosystem with policies as guidance, infrastructure as a foundation, integrated development as a goal, and regional collaboration as a path. The “2025 Comprehensive Computing Power Index” shows that Hebei maintains the top position in the country. The image shows the Qinhuai Big Data Industrial Park in Huailai County, Zhangjiakou City, Hebei Province, taken on September 7, 2025. Photo by Chen Xiaodong / People’s Pictures.

Strengthening Core Technology R&D to Build a Solid Foundation for Independent and Controllable Development

Core technology R&D must upgrade its goals from chasing single indicators to a systematic battle driven by ecosystem building. First, it must root itself in fundamental principles. If source innovation only focuses on the application and engineering levels, it will always be limited to patching up others’ theoretical frameworks. More resources must be directed towards foundational research areas such as algorithm interpretability, causal reasoning, and brain-like computing to master the underlying logic that defines technical routes, thus fundamentally breaking away from path dependence. Second, targeted breakthroughs and large-scale iterations must be balanced. Focus on core links of the AI industry chain such as AI chips, development frameworks, and foundational software, implementing mechanisms like “challenge-based competitions” and “horse racing” to concentrate efforts on overcoming key bottlenecks. More importantly, technological breakthroughs must form a closed loop with market applications; only by investing domestic software and hardware on a large scale in real training scenarios and continuously iterating and optimizing through large-scale trial and error can we use market feedback to enhance technological maturity and gradually form an ecosystem that can compete with first movers.

Optimizing Data Supply to Unblock High-Quality Supply Bottlenecks

China has obvious advantages in data resources, but it must address the two major bottlenecks of “refinability” and “circulation.” First, build high-quality “data oil fields.” Relying on national-level data labeling bases, prioritize establishing standardized dataset systems in mature fields such as industry, healthcare, and finance, while increasing R&D investment in data synthesis and intelligent enhancement technologies. Only by processing raw data into high-quality data that can be directly used for model training can data elements truly enter the production function. Second, use institutional innovation to unblock circulation bottlenecks. Accelerate the provision of foundational systems around property rights definition, revenue distribution, and safety compliance, promoting innovative models such as “data sandboxes” and “regulatory sandboxes” to achieve multi-source data fusion training while ensuring ownership remains unchanged and is safe and controllable, allowing data to truly realize value multiplication through flow.

Accelerating the Promotion of Scaled Applications to Build a Sustainable Commercial Closed Loop

Application scenarios are the ultimate battlefield for testing the quality of AI. The core issue facing the current development of the AI industry is not the absence of good pilot projects but the inability to replicate good pilots in bulk. It is essential to implement the “AI+” initiative deeply. First, deeply embed AI into core business processes, pushing it from auxiliary scenarios into high-value areas such as R&D design, production scheduling, and risk control, thereby significantly reducing costs and increasing efficiency to stimulate enterprises’ willingness to pay. Second, construct a collaborative engagement mechanism for the industrial chain. Promote deep coupling among computing power providers, model vendors, and industry users, forming a collaborative network where computing power is supplied on demand, models are adapted on demand, and scenarios are rapidly implemented, breaking the “each managing their own” situation through standardized interfaces. Third, firmly promote the transformation towards productization. Transition from customized project-based solutions to standardized solutions that are configurable, replicable, and maintainable, allowing for the dilution of R&D and computing power costs through scaling, driving the industry from a cash-burning cycle into a profit cycle.

Enhancing Safety Governance Capabilities to Establish a Solid Safety Bottom Line for Industry Development

The black box nature of AI, its self-evolution capabilities, and generalization abilities extend risk sources from external attacks to the “genetic defects” of the models themselves. Safety governance must upgrade from static compliance checks to dynamic protection throughout the entire lifecycle. First, establish an agile governance structure that is layered and categorized. Emphasize transparency and traceability for general foundational models, while implementing differentiated regulation based on risk levels for vertical application scenarios, such as strict certification and robustness evaluation for high-risk fields like healthcare and finance, and lighter regulation for other lower-risk scenarios, achieving a precise balance between safety and development. Second, strengthen internal safety barriers in technology. Increase R&D investment in safety technologies such as algorithm interpretability, privacy computing, and adversarial training, and establish a routine model safety inspection mechanism to preemptively address risks with a “technical firewall,” making safety capabilities the “factory settings” of models rather than an afterthought. Third, proactively lead the construction of global rules. Promote the transformation of China’s practical experiences in areas such as data classification, algorithm filing, and safety assessment into international governance solutions, seizing the initiative in rule-making within multilateral frameworks to avoid being locked in from behind.

Strengthening Multi-Party Collaborative Support to Build a Full-Element, Full-Ecosystem Support System

Systematic breakthroughs require matching institutional provisions and factor support. On the funding side, it is essential to cultivate genuinely patient capital that adapts to innovation. Leverage national funds to lead and link local areas to form a tiered matrix of patient capital, ensuring long-term investments in foundational breakthroughs and infrastructure development. Simultaneously promote inclusive tools like “computing power vouchers” to lower the barriers for small and medium enterprises to participate in innovation. On the talent side, focus on cultivating “dual-use talents” who understand both algorithmic logic and industry pain points. These composite talents cannot be mass-produced in classrooms; they must be nurtured through collaborative platforms built by leading enterprises and universities in real industrial scenarios. Accelerate the establishment of a composite talent training system with scale effects, forming a tiered supply from top scientists to large-scale application talents. In terms of open cooperation, it is necessary to root in China while connecting globally. Relying on mechanisms like the “Belt and Road Initiative,” support enterprises in deeply embedding themselves in global innovation networks through open-source collaboration and joint R&D, breaking non-commercial barriers under compliance, and enhancing competitiveness in open competition, thereby mastering strategic initiatives in the new round of technological revolution and industrial transformation.

Research Notes

From the perspective of the historical evolution of human civilization, the profound significance of AI may far exceed our current cognitive boundaries. It is not only an iteration of technology and an upgrade of industries but also a systematic reshaping of human cognition and social organization forms. As machines begin to learn, reason, and create, we face not only a technical competition but also a re-evaluation of humanity’s own position. Throughout the research journey, from the computing power artery woven by “East Data, West Computing” to the data vitality activated by “regulatory sandboxes,” from the ecological wave sparked by open-source large models among global developers to humanoid robots working alongside humans on production lines, the sights and experiences have deeply conveyed a vigorous upward force. This indicates that in this wave of technological innovation, we are no longer latecomers, followers, or remedial learners, but competitors on the same stage, and even leaders in certain fields. As global AI development and governance remain in a state of chaos, our path choices are opening up new possibilities—replacing closure with openness, replacing monopoly with collaboration, and replacing control with empowerment in a new paradigm of intelligent civilization. Years later, when people look back at the starting moment of this AI revolution, they may evaluate it this way: at the historical juncture of the new era, China did not hesitate or miss out but instead stepped forward boldly.