High Demand for Video Content Drives China’s AI Compute Consumption

A significant portion of China’s artificial intelligence computing resources is currently being directed toward the production of short dramas and video content. According to unverified reports circulating on social media platforms like Reddit, approximately 55% of China’s AI computing power is utilized for generating these digital media formats.

This surge in video-centric compute usage has prompted speculation regarding potential regulatory shifts. Some observers suggest that Chinese authorities may eventually implement crackdowns on high-intensity video generation to reallocate computational resources toward large-scale model training. This tension between consumer-facing generative applications and foundational model development mirrors discussions in the West, where the high compute requirements of video models were cited as a factor in OpenAI’s decision to pause the rollout of its Sora model.

Advancements in Video Generation and Model Capabilities

The Chinese AI ecosystem is seeing a rapid expansion in specialized video models. Recent developments include Alibaba’s WAN 3.0 and MiniMax’s H3, which feature capabilities such as seamless body-swapping in video content using a reference image. While these tools offer creative utility, they have also raised concerns regarding the potential for generating non-consensual adult content.

In the competitive landscape of video generation, Kuaishou Technology has unveiled a video generator intended to rival OpenAI’s Sora. Additionally, ByteDance has released Seedream 4.0, an image model designed to compete with Google’s “nano banana” technology.

Hardware and Infrastructure Developments

Despite ongoing U.S. export restrictions and bans on technical assistance, China continues to advance its domestic AI infrastructure. Baidu has reported that its cluster of 30,000 self-developed P800 Kunlun chips is capable of training models similar to DeepSeek. Furthermore, Alibaba’s RISC-V IP has reportedly powered over 6 billion shipped chips, targeting the AI and automotive sectors.

To mitigate the impact of chip restrictions, Chinese firms are forming alliances to build a domestic ecosystem. Notable infrastructure developments include:

  • Huawei: Has launched the CloudMatrix 384 server as an alternative to Nvidia’s infrastructure stack and is developing Ascend chip-based training technology.
  • Green Energy: China has launched its first AI data center powered entirely by green electricity.
  • Quantum Integration: Researchers have utilized the 72-qubit Origin Wukong quantum computer system to fine-tune billion-parameter AI models.

Scaling and Economic Efficiency

Data suggests a massive scale in Chinese AI operations, with reports indicating that Chinese AI models are processing 4.69 trillion weekly tokens, surpassing U.S. totals. Efficiency in training is also becoming a focal point; MiniMax has claimed its M1 model costs approximately 200 times less to train than OpenAI’s GPT-4.

The sector is seeing diverse applications ranging from scientific research to industrial processes. This includes the use of AI in solar observation, pharmaceutical drug screening via Tsinghua University’s DrugCLIP, and the deployment of 10 AI models in orbit for space infrastructure. While some developments, such as the autonomous agent Manus, represent shifts in agentic AI, others, like the use of AI in industrial fermentation, focus on specialized process control.