Electricity is China's confidence in the global AI competition
According to a report by Global Times,the 2026 World Robot Conference held in Beijing has become the focus of global industry attention recently. With the increasingly widespread use of artificial intelligence (AI) and related equipment, the focus of AI competition is undergoing fundamental changes. In the past, the competition was about who could make the large model stronger. In the future, what is more crucial is who can make intelligence become an "infrastructure" that is stably supplied, used on demand, and universally accessible, just like water, electricity, and the internet.
China's greatest potential advantage at present lies in its ability to participate in global competition with relatively low unit intelligence costs. China's large AI models adhere to open source and low prices, primarily due to model efficiency, open weights, and sufficient competition. The energy advantage determines whether this low price can be maintained in the long run and whether it can buy time for domestic chips, system software, and models to catch up. From this, we can draw a more important conclusion: whoever can first enable more countries to use, afford, and use AI well will be able to accumulate more real-world tasks, developers, engineering experience, and technical standards, and gain a greater initiative in the next round of competition.
AI competition enters the infrastructure stage
Artificial intelligence is gradually evolving from a high-end technological product into a fundamental capability for economic and social operations. Enterprises embed it into their research and development, manufacturing, marketing, and management processes, governments utilize it to enhance public services and governance efficiency, and individuals use it as a daily tool for acquiring knowledge, processing information, and completing work. As the frequency of usage continues to increase, "intelligence" will be continuously utilized and billed on a per-use basis, just like cloud computing and electricity.
Infrastructure competition follows its own rules. While having the most advanced technology is indeed important, what often determines the speed of popularization is whether the price is affordable, the service is stable, and the deployment is convenient. Once a certain technology system takes the lead in entering government platforms, telecommunications networks, corporate processes, and university curricula, the interfaces, talents, software tools, and compliance systems formed around it will continue to accumulate. Even if later entrants have stronger technology, they will have to pay a high cost to replace it.
This competition is not solely between China and the United States. Nearly 80% of the global population resides outside of these two countries. For numerous governments, operators, and enterprises in Southeast Asia, South Asia, the Middle East, Africa, and Latin America, the most pressing need is not a model that ranks first in all benchmark tests, but one that is affordable, sufficiently capable, can be deployed locally, and adapts to local languages and industrial needs.
Market signals have emerged. According to data released by OpenRouter, an AI model aggregation platform, in 2025, US models accounted for approximately three-quarters of the platform's token traffic; by early June 2026, the token share of Chinese models had surpassed that of US models. Its analysis covers more than 450 trillion tokens in the first half of 2026, with the growth of Chinese models such as DeepSeek mainly coming from real workloads such as programming and intelligent agents, rather than just general chatting. OpenRouter does not represent the entire market, nor does it include a large number of direct enterprise sign-ups and private deployments, but it clearly reflects the direction of change in the price-sensitive market: as model capabilities gradually converge, cost will increasingly directly affect adoption.
What the market needs is affordable intelligence
After the widespread adoption of artificial intelligence, the market's evaluation criteria shift from merely assessing the intelligence level of models to considering the cost incurred in completing a task. This is particularly evident in programming and intelligent agent scenarios, where models require repeated planning, tool invocation, result checking, and error correction. A single task may involve hundreds or even thousands of invocations. The seemingly small price difference for a single invocation can be magnified exponentially in large-scale operations.
The diffusion speed of China's open source models illustrates this point. According to the Spring 2026 report released by Hugging Face, the world's largest artificial intelligence open source community, China's open source models have accounted for 41% of platform downloads, making it the largest country source. Downloads do not equate to commercial revenue, nor do they equate to actual deployment, but they indicate what global developers are choosing, learning, and developing the next generation of applications based on.
The open weights play a pivotal role here. They lower the barriers to model acquisition, secondary development, and local deployment, enabling multiple cloud service providers and model hosting platforms to compete around the same model. For many developing countries, open weights mean that they do not have to rely solely on a few overseas platforms. Instead, they can deploy models on local servers, conduct language adaptation and industry fine-tuning, and keep the data within their own countries. Price, openness, and controllability thus form a synergistic force.
To understand why the Chinese model is price-competitive, we can analyze it from three levels: energy, computing power, and model. Energy is at the bottom level, providing power for data centers; chips are at the middle level, converting power into computation; and models are at the top level, converting computation into usable intelligence. The three together determine the cost per unit of intelligence, in addition to server, network, cooling, construction, and capital expenses.
It should be noted that China is not leading in all three dimensions. The United States still holds some advantages in terms of computing power and models. The current low price of Chinese models is mainly formed at the model and business layers. Opening weights reduces licensing costs, and fierce hosting competition drives down service prices. The price advantage of Chinese models primarily stems from model efficiency, engineering optimization, open ecosystem, and market competition. The role of energy is even more profound: it determines whether this advantage can be maintained as the invocation volume continues to grow.
Transforming energy advantages into computational cost advantages
At the current stage, electricity cost is not the highest proportion in the model inference cost. According to relevant calculations, the median power consumption for one text request in large model applications is approximately 0.24 watt-hours. However, as artificial intelligence transitions from occasional question-answering to intelligent agents, industrial processes, and social infrastructure, energy constraints will become increasingly prominent. In the past, the industry focused on whether there were enough chips; in the future, with billions of users and massive enterprises continuously invoking models, the question will gradually shift to whether gigawatt-scale power can be provided stably over the long term. Chips determine computing efficiency, and electricity determines computing scale; both are indispensable.
China's advantage is primarily reflected in the speed of energy facility construction and infrastructure commissioning. In 2025, the country's newly installed wind and solar power capacity exceeded 430 million kilowatts, with solar power accounting for 318 million kilowatts of this increase. The United States also possesses a vast pipeline of proposed power sources. What China needs to address is how to allocate and consume energy more efficiently, while the United States faces more prominent bottlenecks in grid connection and expansion.
At the hub of the "east-to-west data transmission and computing power allocation" initiative, energy advantages have begun to translate into cost advantages in computing power. Relying on the green electricity park mechanism, the Zhongwei data center cluster in Ningxia has maintained a stable household electricity price of around 0.36 yuan per kilowatt-hour, which is about 45% of the price in the eastern region. At the national level, efforts are also being made to promote the synergy between data transmission and computing power, direct supply of green electricity, and an increase in the proportion of renewable energy consumption in data centers. The aim is to more effectively match energy resources, computing power facilities, and practical tasks.
Low-cost electricity cannot alter chip performance, but it can partially alleviate the operational cost pressure brought about by insufficient chip energy efficiency. This enables Chinese enterprises to expand their inference scale even when they are not yet fully technologically advanced, and buys time for continuous iteration of chips, system software, and models. More precisely, the actual cost of a country's global supply of intelligence is determined by a combination of electricity prices, chip energy efficiency, and model efficiency. The lower the electricity price, the more competitive the cost per unit of intelligence generated by large models becomes.
Today, our advantages lie in the low prices formed by model efficiency, an open ecosystem, and full competition. In the future, a more sustainable advantage will be the scale space provided by energy and infrastructure for such low prices. In the next stage, our focus should not only be on providing cheap electricity to data centers, but also on producing more effective intelligence at lower costs. Firstly, we should accelerate the coordination between computing and electricity, enabling the scheduling of training, batch processing, and delayable inference tasks in time and space with low-cost green electricity. Secondly, we should establish a long-term, stable, and predictable green electricity supply mechanism to provide long-term investment and global pricing basis for high-utilization computing facilities;Third, we should shift away from evaluating solely based on the number of racks, peak computing power, and power consumption scale, and place greater emphasis on how many real tasks are completed per kilowatt-hour and how much effective intelligence is generated. Fourth, we should link policy support to energy efficiency improvement, compatibility of domestic software and hardware, actual utilization rate, and overseas service capabilities, to prevent low-level redundant construction and idle computing power.
It must be clearly understood that for some time to come, the United States will still possess the most advanced chips, the strongest cutting-edge models, mature basic software, and high-value enterprise markets. The opportunity for China's artificial intelligence lies in making sufficiently advanced large models and intelligent agents more reliable, open, and universally accessible, and in turn promoting technological progress through global deployment.
