Electricity Without Leaving the Country, Value Has Gone Global: China's Large-Scale Model is Frequently Used by Users Worldwide

Source: People's Daily Date: 2026-03-31 11:25

3 PM, New York, USA. Faced with a newly received engineering task, Mike, an engineer from an American startup, skillfully used an artificial intelligence (AI) model assistant from the Chinese company Large-Scale Model. After inputting key requirements for a few seconds, a clearly structured and comprehensive project plan appeared on the screen. With this "intelligent helper," a task that would normally take half a day was completed in less than half an hour.

Human-Computer Dialogue—

AI Understands Human Language by "Breaking Down" Sentences and Paragraphs

Overseas, more and more people are starting to use China's Large-Scale Model to improve work efficiency. Each dialogue of the model consumes a digital resource called a "token." A token is the smallest unit for AI to understand human language. Just as humans don't read the entire article at first when learning a language, but break it down into sentences, and then sentences into words, AI also needs to "break down" sentences and paragraphs to understand human language. Generally speaking, one Chinese character is approximately equal to one word unit; one English word is approximately equal to one to two word units; punctuation marks also count as word units. For example, the sentence "今天天气很好。" can be broken down by AI into "今"/"天"/"天"/"气"/"很"/"好"/".", totaling approximately seven word units.

Current large-scale models have a limited capacity for processing text at a time. For example, if a user inputs a sentence and the model generates a response, the combined number of word units cannot exceed a certain limit. AI tools charge based on the number of word units used, much like phone calls are charged by the minute; the more text input and output, the higher the cost.

Data shows that in March of this year, China's daily word unit usage exceeded 140 trillion, a more than 1,000-fold increase in two years. According to the latest statistics, China's large-scale AI models lead the world in word unit usage.

Why do global users choose China's large-scale AI models? Yan Yijun, Vice President of MiniMax, explained: “Users’ selection criteria are very straightforward. First, is the model smart enough, easy to use, and responsive enough to truly solve complex problems? Second, is the price reasonable and sustainable?” The essence of word usage volume is real-world usage, reflecting the depth of penetration, frequency, and breadth of application of large-scale AI models in real-world scenarios. “Chinese large-scale AI models are frequently used by users worldwide. This is recognition from global users,” said Yan Yijun.

Model Optimization—

Rich application scenarios continuously provide a testing ground for technological iteration

For large-scale models, simultaneously achieving greater intelligence and lower cost is not easy—improving model performance often involves an exponential increase in the number of parameters, meaning more words are needed; and the more words are consumed, the higher the operating costs for enterprises and the usage costs for users.

How to strike a balance? Yan Yijun explained, taking the "MiniMax M2.5" model as an example, "On the one hand, through algorithmic innovation, we enable the model to approach the answer with a more efficient and streamlined reasoning path, reducing word consumption at the source; on the other hand, we are committed to improving the quality of each word." Under the condition of a high throughput of 100 words per second, the model only costs $1 to operate continuously for one hour. Some organizations have calculated that, under the same performance, the cost of calling Chinese models is only about 1/10 of that of American models.

"The decrease in cost is not only a result of the technological progress of Chinese large-scale model companies, but also a concentrated manifestation of China's advantages in electricity and supply chains," said Li Zhiqing, professor at the School of Economics, Fudan University. my country's rich application scenarios continue to provide a testing ground for the iteration of AI large-scale model technology. As of December 2025, my country's generative artificial intelligence users reached 602 million, an increase of 141.7% compared to December 2024. The rapid expansion of the user base has driven AI to extend from the internet sector to deeper scenarios such as office collaboration and industrial design, transforming AI from a novel technology into an everyday tool. The increasing application penetration rate also provides continuous data feedback for model iteration, improving the ability to handle complex tasks and opening up space for future development.

The ultimate goal of AI is electricity. "An AI server consumes 5 to 8 times more power than a traditional server. Training a large model requires hundreds of millions of kilowatt-hours of electricity, and daily operating power consumption exceeds 500,000 kilowatt-hours. Electricity costs are crucial to the global layout of the computing power industry," explains Li Zhiqing. "This is precisely China's advantage. China has the world's largest power supply system, coupled with ultra-high-voltage power grids and green energy consumption mechanisms, and fully independent and controllable power technology, giving our computing power stability and a cost advantage."

East-West Collaboration—

Intelligent Scheduling Achieves Precise Matching of Redundant Green Energy with Computing Power Shortages

NARI Technology Co., Ltd., located in Nanjing, Jiangsu Province, is a crucial node in this chain. On its technical support platform serving power grid scheduling, real-time power consumption is clearly visible. Inexpensive and abundant green electricity from the Gobi Desert in western China and photovoltaic power from the Qinghai-Tibet Plateau is continuously being fed into the grid and efficiently transmitted to computing hubs in the east through a self-developed multi-level intelligent dispatch system.

"The west has a large amount of green electricity, but storage is difficult; the eastern computing centers have high electricity demand, but costs are high. Intelligent dispatch can achieve precise matching between redundant green electricity and computing power gaps," explained a representative from NARI Group. During periods of high renewable energy generation, the system enhances the transmission and consumption capacity of green electricity, supporting full-load operation of computing centers; during peak electricity demand or when renewable energy output is insufficient, it can intelligently adjust computing load, prioritizing the operation of core AI businesses and implementing off-peak electricity consumption.

"Furthermore, the complete AI supply chain further reduces industry costs," said Li Zhiqing. Strong collaboration among AI chips, servers, computing infrastructure, cross-border networks, edge computing, and cross-border settlement creates a complete supply chain advantage.

How significant are the benefits of this collaboration? Li Zhiqing calculated that for traditional products, one kilowatt-hour of electricity typically generates 1-2 times the value of electricity, but for word units, this value increases by tens or even hundreds of times. "Now we are transforming our energy and manufacturing advantages into digital value that reaches the global market, achieving the goal of exporting value without leaving the country."

"Word unit export" also faces some challenges. Li Zhiqing believes that, on the one hand, computing power faces some "bottleneck" technologies, and its ceiling needs to be broken; on the other hand, it is necessary to guard against the rise of trade protectionism. He suggests accelerating the top-level design of relevant governance mechanisms to address potential risks related to data, computing power, and market expansion.

Editor: Liu Chuqi