China’s Approach to AI: Infrastructural Statecraft, Managed Pluralism
China’s artificial intelligence (AI) trajectory is often characterised as ‘frugal innovation’ (Waters 2025), ‘self-reliance’ (Wang et al. 2025), or a state-controlled approach to innovation and diffusion (Roberts et al. 2021). Many observers have focused on China’s AI policy and regulatory approach, portraying it as ‘vertical’ (as opposed to the ‘horizontal approach’ of the European Union’s AI Act) (O’Shaughnessy and Sheehan 2023), ‘iterative’ (Sheehan 2023), or ‘adaptive’ (Migliorini 2024). In the context of United States–China AI rivalry, geopolitics is always central to such analysis, often based on a close reading of policy documents.
This essay aims to take a different approach that does not rely only on a close reading of policy documents, delving into policymaking processes, or engaging with the geopolitical debate. Instead, it investigates the fundamental features of China’s approach to AI by taking a layered perspective. China’s approach to AI can be summed up in its official slogan: ‘The government builds the stage, enterprises put on the show’ (政府搭台, 企业唱戏). It is a stage built and governed by the state, and the show is managed pluralism.
Setting Up the Stage
China’s AI push is best understood as ‘infrastructural statecraft’—a layered stage for which the state co-designs energy, geography, hardware, talent, and platforms, and then invites firms to compete on it. This is accomplished not only at the central level: governments at provincial and municipal levels also compete to attract talent and invest in high-tech sectors that are identified as the ‘new quality productive force’ (新质生产力) for an innovation-driven growth model (Huld 2024), as specified in the Fifteenth Five-Year Plan (2026–30) (Horn and Tse 2026). These sectors include semiconductors, robotics, electric vehicles (EV), green energy, and—of course—AI.
What we see on the stage, such as the growth of Chinese large language models (LLMs), is sustained by multiple pillars, the first of which is energy. Under the official ‘new infrastructure’ (新基建) initiative (Meinhardt 2020), servers and power plants are treated as a single category, directly linking the development of the AI sector directly to long-term energy planning. Beijing’s buildout of renewables, nuclear, and ultra-high voltage (UHV) transmission is increasingly justified as necessary for ‘green computing’ rather than only in terms of industrial policy or decarbonisation goals. Gigawatt-scale wind and solar bases in the northwest of the country, new nuclear power stations along the coast, and upgraded coal plants in the central provinces are planned, together with large data centre clusters, so that AI-ready power is available as China continues to experience rising demand for energy and computing capability. China has twice the energy capacity of the United States (Chan et al. 2026). This surplus provides a massive, stable foundation for powering the enormous data centres and AI factories needed for LLM training and inference (Gibson 2025).
The second pillar is ‘compute’ (the processing power that runs applications). The ‘East Data, West Computing’ (东数西算) project (also known as the ‘East–West Computing Transfer’) is an ambitious national strategy launched in 2022 to redistribute computing power from the data-heavy, energy-poor eastern regions to the resource-abundant but data-scarce western provinces (Wang 2022). Data centre clusters or ‘computing hubs’ have been built in Guizhou, Gansu, Ningxia, and Inner Mongolia, where land and electricity are cheaper, while UHV lines transmit both electrons and bits across provinces. By turning peripheral places into the digital–urban frontier (Zhao 2025), the stage is set and stretched across the country for state and private companies to compete on and profit from infrastructural development. More recently, Chinese data centres have also been built under the sea to utilise the natural cooling conditions underwater (Tong 2026).
The third pillar is the hardware/microchips. In the face of the US chip bans, China not only prioritises domestic chip production but has also weaponised its dominance in the extraction and processing of rare earth minerals as a counter-bargaining measure (Chen and Jackson 2026). Huawei’s Ascend series and other domestic alternatives (such as Cambricon, Moore Threads, Biren, and Alibaba) are given a protected runway through procurement mandates, tax breaks, and informal guidance for state-linked clouds and data centres to move away from Nvidia products (Jiang 2025). Even though imported chips still dominate the upper stack, the direction is towards a mixed environment in which domestic chips are guaranteed minimum demand and progressively upscaled to replace imported ones (Wang et al. 2025). Huawei provides a full-stack AI infrastructure, ranging from Ascend chips and cloud services to ModelArts (Huawei Cloud’s comprehensive, one-stop AI development platform) and CANN (Compute Architecture for Neural Networks), an alternative to Nvidia’s CUDA (Compute Unified Device Architecture) (Huawei 2026).
Next is the human/talent layer—the ultimate trump card that China holds. In 2025, half of the world’s top-tier AI researchers were China-born. They dominate the world’s AI publications and global AI patents (Normile 2026; WIPO 2024). ‘China is winning the AI talent race’, according to The Economist (2026). Talent is treated as human infrastructure—a pipeline that must be planned, subsidised, and geographically steered. China has taken a multipronged strategy to grow its talent pool in science and technology, from education policy reforms to attracting talent (Zwetsloot 2020). Starting by establishing AI majors in the early 2000s, Chinese universities have focused on AI talent under the Double First-Class (双一流) initiative (2015–50), aiming to build globally first-class universities and academic disciplines, with quotas and funding tilted towards computer, quantum, and materials sciences, and engineering (Fu and Ji 2024). AI has become part of the curriculum in schools (Ruwitch 2026). China has mandated AI education in primary and secondary schools, to cultivate future tech talent (Department of Education 2025). At the same time, hundreds of policies, incentives, and initiatives have been rolled out to attract the world’s top scientific talent (Mallapaty 2025), ranging from financial and research incentives and targeted visa programs to specialised efforts to reverse the brain drain (to bring back overseas-educated Chinese professionals) (Groenewegen-Lau and Hmaidi 2024). There is intense competition—sometimes fragmented but at other times through coordinated industry–university–government partnerships—between key cities, zones, and industries to attract and create a steady flow of engineers, data scientists, and product managers (CCG 2017).
Finally, governance is embedded through the pillars supporting the stage. It underpins infrastructure building, chip policy, cloud procurement, content-moderation law, talent cultivation and recruitment, and the institutional capacity to deploy models in areas where needed. A suite of AI-relevant laws, regulations, and guidelines—from data security and personal information protection to algorithmic recommendation rules and generative AI provisions (CMS 2026)—defines the outer boundaries of acceptable model behaviour, deployment, and applications. Chinese AI governance is characterised by strategic pilots (Hu and Au 2025) and early alignment of values and systems to prevent undesirable outputs from the source (inputs), rather than applying penalties for the outputs; it uses licensing and registration as chokepoints, particularly for public-facing generative AI services (Zou and Zhang 2025), including algorithms, models, and products. It sets guardrails against misuse and overuse and stipulates responsibility to protect vulnerable people (including workers) (Viaplana 2026; Wilkins 2026). Chinese AI governance is integrated with national strategies and international cooperation: economic integration, national security, geopolitical supremacy, and international ethical guidelines are all treated as a holistic design problem, not separate domains.
Such a governance framework—not a single comprehensive law but a suite of incremental, adaptable, actionable, controllable, and multilevel laws, regulations, initiatives, funding channels, and standards—is commissioning an AI‑ready landscape: comprehensive manufacturing capability, renewable and conventional power plants feeding UHV lines; inland data centre clusters absorbing compute workloads; a protected runway for domestic accelerators from Huawei to newer AI chip firms; a talent pool; and a permissive space for intense competition among firms.
On this engineered stage, enterprises ‘perform’—racing to build better models, platforms, and applications, but always within the scripted boundaries of political and security acceptability.
Layered Performance
On the stage, both state-owned enterprises (SOEs) and private enterprises perform at and across different layers. Here I focus on three layers: the foundational model, the platform, and the application.
The layered architecture—from models to platforms to applications—is routed around the holistic governance framework. It has a temporal dimension—that is, a phased strategy. For example, we first witness the ‘hundred-models war’ (百模大战) in which Chinese tech giants and startups race to develop the best value-for-money LLMs (Colville 2024). This competition leads to consolidation, with the emergence of a few successful platforms that not only develop and support base models but also provide toolkits for third-party developers and applications. Finally, the leading models and platforms embed their products in the industrial, administrative, and everyday fabric—all on the stage supported by the four pillars (energy, infrastructure, chips, and talent) and under a governance regime that treats political alignment, safety, economic upgrading, and social stability as a single integrated design problem.
The Foundational Layer
First, let us look at the foundational layer, which is exemplified by the ‘hundred-models war’. The term became a buzz phrase in Chinese tech and media sectors and has since been taken as shorthand for a mass movement among tech companies and startups to train LLMs. Starting in 2023, after the release of ChatGPT in late 2022, dozens upon dozens of LLMs emerged from tech giants, SOEs, and startups. Supported by government subsidies, they competed for talent and users in both upstream (base model) and downstream (application) domains. The ‘hundred-models war’ allowed the industry to test multiple technical paths, observe market responses, and ensure a controlled competitive environment from which a smaller group of national champions would emerge.
The reference to ‘one hundred’ invokes history. For some, it brings memories of Mao Zedong’s ‘Hundred Flowers Campaign’ (百花齐放) of 1956, when Chinese intellectuals were invited to speak freely about the political situation—just in time for the most outspoken critical voices to be persecuted in an anti-Rightist backlash the following year. For others, it conjures up the ‘Hundred Regiments Offensive’ (百团大战) of 1940, when the Communist army joined Nationalist forces against the invading Japanese. One is remembered as a trap and a betrayal; the other for acts of courage and solidarity. Today’s ‘hundred-models’ competition reprises memory in a very different register: where the Mao-era ‘hundred’ movements managed ideas through ideological and military campaigns, the contemporary ‘hundred-model war’ manages code through infrastructure, licensing, and standards.
Policy and capital have helped sustain an initially messy war of LLMs, which narrowed into a smaller cluster of base-model champions, known as the ‘LLM Six Tigers’ (大模型六小虎; see Table 1). They are leading generative-AI startups, each valued at more than US$1 billion. They are widely recognised as the key private ventures competing alongside tech giants such as Alibaba and ByteDance.
Table 1: LLM Six Tigers (compiled by the author from multiple publicly available sources, including official websites, and industry, media, and research reports)
The state did not choose these firms directly, but it shaped their environment: setting computing priorities, encouraging domestic alternatives to foreign models, supporting the open-source movement, and signalling which models are ‘backbone’ (主干网络) models. This logic has seen further consolidation in the LLM arena. The small number of startups is seen alongside existing tech giants and platform companies as strategically important national assets (Table 2). Together, they form the foundational model layer: Chinese general purpose and open-source models dominate the open-source market (Keary 2026), which other actors are expected to finetune, extend, or build on top of.
Table 2: Leading Foundation Models from Major AI Companies (compiled by the author from multiple publicly available sources)
The Platform Layer
On top of these foundation models sits a technical layer of major AI platforms, from general cloud-based services to specialised LLM applications and AI development tools. Here, the goal is not simply to train raw models, but to build controllable, auditable, and interoperable infrastructure into which enterprises and governments can plug.
At this layer, a handful of platforms turn abstract ‘large models’ into enterprise tools and governable infrastructure (Table 3). Let us start with Baidu, ByteDance, Alibaba, and Tencent, Chinese internet giants known for their cross-industry super-integration capability as their added value (Yin 2025). Baidu’s ERNIE models (Toh 2025), distributed via Baidu AI Cloud and integrated into enterprise platforms such as Inspur’s EPAI (also connected to DeepSeek models), functions as a state-aligned base model: it is open enough to be customised, but wrapped in deployment frameworks that handle computing, privacy, and compliance for key sectors such as finance and transport. ByteDance’s Volcano Engine (Cheng 2026), built on the Doubao models, plays a similar role from a different lineage: it exposes large-model application programming interfaces (APIs), real-time conversational AI, and an agent development stack (Coze Studio, Coze Loop, HiAgent) into which enterprises can plug for everything from advertising and e-commerce to internal ‘digital employees’. Alibaba Cloud’s PAI (platform for AI) offers an end-to-end toolchain—from data labelling through distributed training to online serving—so that enterprises can build on Qwen models without ever leaving a managed environment (Alibaba Cloud Community 2026). Tencent’s Hunyuan model family, delivered through Tencent Cloud, provides general and vertical LLMs that plug directly into Tencent’s wider ecosystem (WeChat, games, advertising, fintech), turning foundation models into readymade services for social, content, and enterprise scenarios (Tencent 2023).
Table 3: ‘Model Plus’ Platform: Cloud and Mobility as a Service Platforms—Key Examples (compiled by the author from multiple publicly available sources)
These are enterprise tooling and integration platforms for vertical developer environments. They are embedded in everyday work environments and collective decision-making. Non-traditional platform companies such as SenseTime and Huawei provide Model as a Service integrated toolchains, sector-specific packages, and safety/surveillance modules to cities, manufacturers, and media firms. These AI platforms sit atop the energy, infrastructure, chip, and talent stacks and translate LLMs into everyday choices.
The Application Layer
China’s approach to AI ultimately rises or falls on the application layer—not on whether LLM chatbots impress benchmark watchers or rival their US counterparts, but on whether models embed themselves in everyday life and work, from factories, hospitals, schools, banks, and logistics networks to city management and social governance systems. This application-centric approach is not an afterthought. From the 2017 AI Development Plan through successive AI-plus action plans at national and provincial levels (Webster et al. 2017), from President Xi Jinping’s instruction to China’s AI industry in 2025 to the Five-Year Plans, policymakers and industry players have treated AI as a productivity driver and economic policy (Xinhua 2025b).
Key players in the foundation and platform layers are active contenders in the application race. Companies not only develop models but also compete on vertical integration—that is, application to the real world. Success is measured by how AI products are embedded in factories, warehouses, enterprise back offices, customs clearance systems, classrooms, aged care facilities, traffic lights, or power grids. Their race to be included in urban governance and social management is an example of how AI mediates the everyday operations of the state (Cugurullo 2021; Arcesati 2022), from smart cities to social credit systems (see Table 4).
Table 4: Platform Race in Urban Governance (compiled by the author from multiple publicly available sources)
While much of the infrastructure underlying the social credit system is still rule-based and database-driven rather than pure machine learning, the logic—using data fusion and algorithmic scoring to categorise and steer behaviour—sits within the broader AI application agenda, even when the specific models are not advanced LLMs. Urban governance becomes a key application arena for China’s AI strategy because it institutionalises these tools in durable arrangements via integrated ‘city brain’ platforms and cross-bureau data exchanges (Xu et al. 2025). Model and platform contenders plug into this environment by offering end-to-end governance AI solutions. AI-driven systems are integrated with algorithmic, statistical, and machine-learning tools to form the everyday grammar of contemporary social governance.
Conclusion
China’s approach to AI, seen through the ‘stage show’ metaphor, is ultimately about authoritarian governance via managed pluralism. The ‘stage’ is infrastructural and layered, including: an energy system branded as ‘green computing’, a spatial rewiring of data and servers through data centre buildouts in western provinces and under the sea, a protected but increasingly capable domestic chip stack from Huawei’s Ascend to other AI accelerators, and a human infrastructure of AI-oriented education, labs, and innovation hubs that feed talent into the burgeoning industry. The governance mechanism pre-empts risks and bakes political red lines and data controls into the technical system and licensing regimes from the start.
The result is a state-orchestrated stage co-designed by the public and private sectors, on which existing and emerging technology and platform companies and startups race to turn energy, compute, chips, and people into open-source LLMs, enterprise tooling, governable infrastructure, and scalable applications. It is not a winner-takes-all scenario but a managed pluralism wherein competition is encouraged and benefits are shared. What ties these sectors together is not a single frontier model, advanced chip, or killer app, but a policy mindset: AI is not a consumer toy; it is an enabling technology for upgrading the real economy and advancing China’s interests in the world. AI is regarded as a pivot into advanced manufacturing (such as robotics) and ‘embodied intelligence’ (具身智能) such as humanoid robots (Xinhua 2025a). Success is measured less by viral products than by how these systems support state-led high-quality development, boost productivity, and enable more effective and responsive forms of governance.
This model of ‘managed pluralism on a state-built stage’ is markedly different from the situation in the United States. In China, the state designs the stage in collaboration with firms and then invites ‘a hundred’ models and firms to compete on it, expecting eventual consolidation into a curated core that serves national development, security, and ideological goals. In the United States, AI has developed far more from firm-led experimentation, venture capital, and open research cultures (Contreras 2025), with infrastructure and governance often going after commercial breakthroughs rather than pre-structuring them or pre-empting risks. The US Government has increasingly engaged with the tech sector through export controls, antitrust actions, and nascent safety regulations (O’Callaghan and Shueard 2025), but it does not co-plan the layers (energy, compute, chips, platform architecture, and talent pipelines) as a long-term and full-stack design the way China does (Chan et al. 2025).
The rhetorical shift from Mao’s ‘let a hundred flowers bloom’ to today’s ‘hundred-models war’ is not accidental. Both are centrally framed campaigns that treat pluralism as a tool, not a value. Today’s ‘a hundred models’ bloom on the stage that the Party-State can manage and reconfigure. How far this managed pluralism will carry China in the evolving AI landscape remains an open question, but it has become a distinctive feature of the country’s bid to shape the next technological frontier.
Finally, it is important to recognise that this seemingly coherent governance logic—infrastructural statecraft plus managed pluralism—is not entirely new, and its outcomes are not linear or fully planned. Other experiments in digital governance, such as the fragmented and locally driven social credit initiatives, show how centrally framed campaigns can generate heterogeneous, contingent, and sometimes unexpected assemblages of indicators and practices beyond the engineered assumptions (Liu and Róna-Tas 2025; Trauth-Goik 2023). In the same way, today’s ‘AI-ready’ stage invites a diversity of actors, models, and urban applications whose trajectories may realise some official goals such as productivity gains or tighter control while generating heterogeneous discourses and even shifting the aims of experimentation over time. Locally varied practices and disparate pathways among diverse actors can produce hybrid infrastructure rather than a single pre-designed system that the centre then must manage. The ‘stage’ is therefore not a fixed scaffold, but a dynamic, evolving infrastructure assemblage that is continually reshaped by experimentation, contestation, and feedback from the performing actors.
References