Review of the AIGC Industry in 2025: User Explosion, Ecosystem Restructuring, and the Computing Power Competition

When the number of domestic AIGC users surpassed the 500 million mark, the industry quietly shifted from a technological frenzy to deep application cultivation. Behind this intelligent revolution lies an unprecedented competition for computing power resources.

In 2025, the AIGC (Artificial Intelligence Generated Content) industry moved beyond its initial hype phase and entered a stage of large-scale application and implementation. As of June 2025, the number of generative AI users in China reached 515 million, with a penetration rate of 36.5%, representing an increase of 266 million users and a growth rate of 106.6% compared to the end of the previous year.

Alongside the explosive growth in user numbers, AIGC is transitioning from a "creation tool" to a "creation partner," evolving from vertical specialization to cross-industry collaboration and building a new social ecosystem of human-machine symbiosis. At the same time, the widespread adoption of AI agents (Agents) has spurred an exponential increase in computing power demand, intensifying the supply-demand tension in the GPU market.

01 Current Development: Explosive Growth and Ecosystem Restructuring in the AIGC Industry

In 2025, the AIGC industry achieved comprehensive breakthroughs in user scale, application scenarios, and technological penetration. This transformation is not only reflected in numbers but also profoundly restructures industrial logic and organizational forms.

As of August 2025, a total of 538 generative AI services had completed filing procedures in China. The technology system centered on large models has gradually permeated various industries. Application scenarios now cover multiple areas, including answering questions, daily office work, entertainment, and content creation, expanding from consumer-facing to production-facing domains and from general scenarios to core industry-specific scenarios.

On the user side, in March 2025, the total new downloads of AI apps exceeded 160 million, with an average of about 200 million people using AI apps daily. Six products recorded over 10 million new downloads that month. For web-based AI products, the total number of visits exceeded 600 million, with a total monthly active user base of over 120 million, and 11 products recorded more than 10 million total visits.

02 Path of Transformation: Triple Transition from Tool Assistance to Ecosystem Restructuring

The widespread adoption of AIGC is driving a systematic reshaping of industrial logic, primarily reflected in three dimensions:

Production logic shifts from "process-driven" to "human-machine collaboration." Traditional industrial production and services rely on standardized processes, focusing on "tools assisting humans." In contrast, AIGC has given rise to a "generative production" model. In fields such as programming, design, copywriting, legal documentation, and market analysis, artificial intelligence can generate drafts, provide solutions, and identify errors. In areas like drug development, materials science, and engineering architecture, AIGC can uncover patterns difficult for human experts to detect by mining vast amounts of data and simulating complex scenarios, significantly shortening R&D cycles and costs.

Industrial organization evolves from a "chain-like ecosystem" to a "networked ecosystem." The penetration of AIGC breaks down inherent industry boundaries, fostering more flexible network-like ecosystems. For example, "decentralized" production and creation in content industries such as media, entertainment, and education have significantly lowered the barriers to professional content production. Small teams or even individuals, aided by AI tools, can now produce animations, music, and courses that approach professional standards.

Corporate competition logic shifts from "scale" to "ecosystem." In the age of intelligence, model performance heavily depends on high-quality data and user feedback. Companies with rich scenarios that can form a closed loop of "user usage—data generation—model optimization—improved experience—attracting more users" will establish dynamic moats.

03 Application Deepening: The Rise of Agents and Cost Optimization Drive Accessibility

2025 has been dubbed the "Year of Agents" by the industry. AI agents, characterized by two major features—"from passive response to proactive action" and "from single function to ecosystem integration"—have realized a new interaction paradigm of intent-driven direct access and service distribution networks.

The widespread adoption of agents has led to exponential growth in computing power demand. The number of tokens invoked by a single agent is hundreds or even thousands of times greater than that of a text-based conversational application. Meanwhile, new model architectures and training strategy optimizations are significantly advancing the democratization of large models. Benefiting from innovations in model architectures (such as Mixture of Experts, MoE), optimizations in training strategies (such as post-training combining Supervised Fine-Tuning, SFT, and Reinforcement Learning, RL), and software-hardware co-design, the democratization of large models is unfolding comprehensively.

In terms of cost, industry competition continues to drive prices downward. For example, among three models focused on inference, Alibaba Cloud's QwQ-Plus (Inference) is priced at 1.6 yuan per million tokens for input and 4 yuan per million tokens for output; Tencent Cloud's hunyuan-T1 (Inference) is priced at 1 yuan per million tokens for input and 4 yuan per million tokens for output; and DeepSeek-reasoner (Inference) is priced at 4 yuan per million tokens for input and 16 yuan per million tokens for output.

04 Computing Power Demand: Dual Drivers of Inference Demand Surge and Agent Proliferation

In 2025, AI large models are developing towards greater strength, efficiency, and reliability, featuring deepened inference models and an explosion of agent models. Industry analysis reports indicate that 2025 marks the "Year of Agents," with inference demand driving an explosion in computing power requirements and highlighting the trend of domestic computing power autonomy and controllability.

The speed at which AI large models penetrate industries exceeds that of the previous internet revolution. ChatGPT achieved a penetration rate in two years and one quarter that roughly corresponds to the ten-year penetration progress of the PC internet. This rapid penetration has brought unprecedented demand for computing power.

According to market research reports, large models and generative artificial intelligence are driving up computing power demand, with China's intelligent computing power growing faster than expected. In 2025, China's intelligent computing power scale is projected to reach 1,037.3 EFLOPS, with expectations to reach 2,781.9 EFLOPS by 2028. In 2025, China's general computing power scale is estimated to be 85.8 EFLOPS, a growth of 20%, while intelligent computing power growth will reach 1,037.3 EFLOPS, an increase of 43%. The growth rate of intelligent computing power far exceeds that of general computing power.

05 GPU Competition: Capital Race Among Tech Giants and Supply Chain Tensions

Facing the explosive growth in computing power demand, global tech giants are engaged in an unprecedented competition for GPUs. In 2025, major technology companies' capital expenditures reached historic highs.

Estimates suggest that the total capital expenditure of Microsoft, Amazon, Google, and Meta in 2024 reached $246 billion, up from $151 billion in 2023. Expenditures in 2025 may exceed $320 billion. Within such high capital expenditures from tech giants, GPU purchases will constitute a significant portion, with clusters of 100,000 cards gradually becoming the standard for AI computing.

Specifically, OpenAI announced the "Stargate Project," a plan to invest $500 billion over the next four years to build AI infrastructure, with an initial $100 billion investment in 2025. Amazon plans to invest $100 billion in infrastructure, up from $77 billion in 2024; Microsoft plans to invest $80 billion in the 2025 fiscal year to build data centers capable of handling AI workloads; and Google plans to invest $75 billion, a 42% increase from last year's $53 billion.

06 Supply Chain Challenges: Advanced Packaging and HBM Memory as Production Bottlenecks

The primary issue with GPU shortages lies on the production side. CoWoS packaging and HBM memory, as two pillars of GPUs, also constrain GPU production capacity. Driven by strong demand for cloud AI accelerators, global demand for CoWoS and similar packaging capacity is expected to increase by 113% in 2025.

TSMC, as a key supplier, has undergone multiple adjustments to its advanced packaging roadmap in 2024. According to TSMC's latest plans, CoWoS monthly production capacity is expected to reach 35,000 wafers in 2024, double to 75,000 wafers by 2025, and further increase to 135,000 wafers by 2026.

Regarding HBM (High Bandwidth Memory), major supplier SK Hynix's HBM production capacity for 2025 is already fully sold out. Consequently, it is continuously expanding HBM production capacity, aiming to achieve a monthly capacity of 140,000 wafers by next year. Simultaneously, it is accelerating product iteration, closely collaborating with TSMC on 16-layer HBM4, with mass production expected to begin in the second half of 2026.

07 Domestic Alternatives: China's Computing Power Development Under the Trend of Autonomy and Controllability

Against the backdrop of the U.S. expanding restrictions on high-end GPU exports, domestic companies are striving for autonomy and controllability. From 2023 to 2025, the U.S. has revised export rules multiple times to expand control scopes. Since GPUs are foundational to the development of artificial intelligence in China, relevant policies continue to intensify support for the domestic computing power industry.

The trend of domestic computing power autonomy and controllability is becoming increasingly prominent, with various manufacturers accelerating the research, development, and application of domestic AI chips. Despite facing technical challenges, China's pace of independent innovation in the AI computing power field has noticeably accelerated, forming a complete industry chain that includes AI chips, server systems, liquid cooling, and optical modules.

When Meta plans to deploy over 1.3 million GPUs by the end of 2025, and Oracle builds Zettascale-level superclusters supporting up to 131,072 Blackwell GPUs, this computing power race has transcended mere technological competition, becoming a contest of technological prowess among nations.

In the AIGC industry of 2025, surpassing 500 million users is just the beginning; the "Year of Agents" initiates an even more complex ecosystem competition. Behind this intelligent revolution, computing power—particularly GPUs—has become a strategic resource scarcer than data. In the future, whoever masters the advantage in computing power will hold the initiative in the AI era, a consensus that has become widely accepted in the industry. In this silent war, technological innovation, industrial collaboration, and autonomy and controllability will be the key factors determining victory.