China AI 2026: Large Model Market Exceeds RMB 70 Billion, Agent Market RMB 44.9 Billion, 72% Adoption

In 1956, a symposium at Dartmouth College in the United States saw scholars such as John McCarthy and Marvin Minsky formally propose the concept of artificial intelligence. The core idea is to simulate human perception, cognition, decision-making, and execution through computer programs or hardware, enabling machines to think and act like humans. This direction has evolved over 70 years, but the basic definition has never departed from the scope of simulation ability and taking action. AI development has undergone three waves. The first wave, from the 1950s to the 1970s, was represented by expert systems, which relied on manually written rules for reasoning but were constrained by rule complexity and contradictions, leading to funding exhaustion in the late 1970s. In 2006, Geoffrey Hinton proposed the concept of deep learning, laying the foundation for the third wave. In 2012, deep learning significantly reduced error rates in the ImageNet image recognition challenge, validating the scalability of neural networks. Subsequently, AlphaGo defeated Lee Sedol in 2016, the Transformer architecture appeared in 2017, GPT-3 was released in 2020, ChatGPT launched in 2022, open-source models saw large-scale application in 2025, and in 2026 AI is moving toward autonomous task execution. AI can be classified by intelligence level into weak AI, strong AI, and superintelligence. Currently, all practical applications are weak AI, which excels in specific domains such as language processing or image recognition but lacks cross-domain general capabilities. Strong AI and superintelligence remain theoretical. By function, AI is divided into discriminative AI and generative AI; the former is used for recognition and classification, the latter for creating new content. By capability level, researchers propose a four-stage classification: reactive, limited memory, theory of mind, and self-awareness. Current technology is at the limited memory stage, moving toward exploration of theory of mind. AI's operation relies on three elements: algorithms, computing power, and data. Algorithms have evolved from traditional statistical learning to deep learning, with the Transformer architecture's self-attention mechanism solving long-range dependency problems, becoming the basis of large language models. Computing power primarily relies on GPUs and specialized chips; training a model with hundreds of billions of parameters requires tens of thousands of GPUs running for months, while the human brain consumes only about 20 watts, highlighting the energy efficiency gap. Data requires large-scale, high-quality corpora; large model training often needs trillions of tokens, and data quality directly affects model performance. The AI technology stack is divided into five layers from bottom to top: infrastructure layer, framework layer, model layer, application layer, and agent layer. The infrastructure layer provides computing chips and cloud computing platforms; the framework layer provides development tools such as TensorFlow and PyTorch; the model layer includes pre-trained large models like GPT and DeepSeek; the application layer supports specific scenarios such as intelligent customer service and AI programming; and the agent layer enables autonomous task decomposition and execution. AI applications have covered multiple domains. In content generation, generative AI is used for text, image, and video creation. In intelligent customer service, data shows that an AI assistant on a payment platform handled two-thirds of customer service conversations, reducing response time from 11 minutes to 2 minutes, saving about USD 60 million in 2025. In enterprise knowledge management, large models combined with retrieval-augmented generation transform internal documents into queryable systems. In vertical industry applications, the proportion of industrial enterprises using large models rose from 9.6% in 2024 to 47.5% in 2025. For agent autonomous execution, in 2026, 72% of enterprises have completed Agent pilots, deploying an average of 3.5 scenarios. In embodied intelligence, AI executes tasks in physical environments such as robots and autonomous driving. In 2026, the AI field shows multiple data changes. China's large model market has exceeded RMB 70 billion, and the enterprise agent market has grown to RMB 44.9 billion. Token consumption has increased 300 times in a year and a half, with China's daily token calls exceeding 140 trillion. Open-source models have accumulated over 10 billion global downloads, and China's large model weekly calls have continuously surpassed the United States. Multimodal processing has become standard, small models have achieved cost reductions of 95% in specific scenarios, and edge AI applications are accelerating. Meanwhile, 47% of organizations have reported AI-related security incidents, making AI governance a prerequisite for deployment.