Text by Xiaofeng

Source | Bowang Finance

The intensity of the artificial intelligence (AI) rivalry between China and the United States rivals the strategic maneuvering seen in trade wars, international diplomacy, and other competitive arenas.

According to data from a globally recognized internet traffic statistics agency, in early April, the weekly invocation volume of China’s AI large models surged to 12.96 trillion tokens.

This marks the fifth consecutive week that China has outstripped the United States in this crucial metric—a notable departure from the previous trend, where the U.S. had consistently led by a significant margin.

In just a few months, the balance of power has shifted dramatically. Amid the prevailing international narrative that “U.S. AI is pulling ahead,” the global AI landscape has quietly been rewritten.

This signals that the AI competition between China and the U.S. has entered a new phase. It reflects not only differences in market demand between the two nations but also foreshadows a broader competition centered on computing power, innovation, and application.

However, focusing solely on the national-level competition is insufficient. The fact that China’s AI large model invocation volume has surpassed the U.S. for five consecutive weeks also reveals another truth: the AI competition among domestic tech giants has intensified into a fierce, close-quarters battle.

01

Fluctuations in Large Model Invocation Volume

Market performance of commodities is often gauged by trading volume and price increases in the stock market. Similarly, fluctuations in invocation volume are not isolated numerical phenomena; they are widely regarded as the most sensitive barometer for the pace of global AI application deployment.

Looking back over the past year, the trend in invocation volumes between Chinese and U.S. large models has resembled an intense tug-of-war. In the first half of 2025, the global popularity of ChatGPT peaked, with the weekly invocation volume of U.S. large models once surpassing 150 billion, leading the world. At that time, many predicted that China’s AI had been left far behind by the U.S., with the gap only widening.

The reversal came quickly, with the situation beginning to shift in the second half of 2025.

Domestic large models underwent rapid iteration, significantly enhancing their capabilities while prices continued to decline. According to the “2025 White Paper on the Development of Artificial Intelligence Large Models” released by the China Academy of Information and Communications Technology, the inference costs of mainstream domestic large models dropped by over 80% in 2025, with some models priced lower than their U.S. counterparts.

These price reductions and technological innovation efficiency directly drove an explosion in invocation volumes. During the week of February 9–15, China’s AI large models recorded 4.12 trillion tokens in invocation volume, surpassing the 2.94 trillion tokens of U.S. models during the same period for the first time.

The surpassing has been sustained.

According to the latest data from OpenRouter, the total global invocation volume of AI large models for the week in early April was 27 trillion tokens, representing an 18.9% increase from the previous week. Among the top AI large models listed, China’s AI large models saw their weekly invocation volume rise to 12.96 trillion tokens, a 31.48% increase from the previous week; while the weekly invocation volume of U.S. AI large models was 3.03 trillion tokens, showing a mere 0.76% increase. China’s AI large models have seen their weekly invocation volume grow for five consecutive weeks, surpassing the U.S. over the same period.

In terms of platforms, the top six globally in invocation volume are all Chinese AI large models. Among the top three, two are models from Alibaba’s Qianwen 3.6 series. Specifically, Qwen3.6 Plus (free) ranks first with a weekly invocation volume of 4.6 trillion tokens; Qwen3.6 Plus Preview ranks third with a weekly invocation volume of 1.64 trillion tokens.

On the U.S. side, although ChatGPT’s invocation volume previously led globally, its growth rate has significantly slowed.

One easily overlooked detail is the significant difference in user structure between Chinese and U.S. large models. Mobile invocation accounts for a substantial portion of Chinese large models.

This means that China’s large models have truly integrated into the daily lives of ordinary people, while U.S. large models are primarily used by professionals on PCs. From students organizing learning materials to office workers processing work documents, and small vendors designing marketing posters, the adoption rate of large models in China has far exceeded everyone’s expectations.

Objectively speaking, market demand has always been the primary driver of technological development. The surpassing in invocation volume essentially reflects that China’s market demand for AI has exceeded that of the U.S. While the U.S. is still debating the ethical boundaries and security risks of large models, China has already begun transforming them into tangible productivity.

Of course, this difference in demand will continue to grow like a snowball, ultimately translating into technological and industrial advantages.

To support such explosive demand, a robust resource system is essential as a guarantee. This is precisely the second battleground in the AI competition between China and the U.S.

02

New Resource Competition Between China and the U.S.

We believe that the resource competition in the AI era has changed the rules. In the past, major powers competed for oil, coal, and iron ore; now, we compete for computing power, electricity, and data infrastructure.

Many believe that the AI competition is about algorithms and models, but in reality, when model architectures gradually converge, what truly determines the outcome is the underlying resource capabilities. Training a large model with 100 billion parameters requires consuming tens of millions of kilowatt-hours of electricity, while the energy consumption during the inference stage is dozens of times that of the training stage. Without strong computing power and electricity support, even the most advanced models would remain mere castles in the air.

In terms of computing power construction, China has demonstrated astonishing speed. Data from the Ministry of Industry and Information Technology shows that as of the end of June last year, the total number of computing center racks in use in China reached 10.85 million standard racks, with intelligent computing power reaching 78.8 exaflops per second and storage capacity exceeding 1,680 exabytes. Additionally, 1,509 large models have been released, ranking among the highest globally.

At this rate, China’s total computing power is expected to soon surpass that of the U.S.

More important than computing power is electricity. The essence of AI computing power is the digital conversion of electricity. Whoever can obtain cheaper and more stable electricity will gain a cost advantage in the AI competition.

China possesses unique advantages in this regard.

In 2025, data released by the National Energy Administration at the beginning of the year showed that the cumulative electricity consumption for the whole society in 2025 was 10,368.2 billion kilowatt-hours, a year-on-year increase of 5.0%, marking a historic breakthrough of 10 trillion kilowatt-hours. This figure is the first of its kind for a single country globally, equivalent to more than twice the annual electricity consumption of the U.S. and higher than the combined annual electricity consumption of the EU, Russia, India, and Japan. Electricity consumption serves as a barometer for socio-economic and industrial operations.

It signifies the accelerated upgrading of traditional industries and the continuous implementation of new quality productive forces...

Meanwhile, data infrastructure is also a strength of China. By 2025, the total number of 5G base stations will reach 4.84 million, accounting for approximately 60% of the global total. The total length of optical cable lines nationwide will reach 74.99 million kilometers, with fiber networks covering all administrative villages and the number of gigabit broadband users exceeding 300 million. This globally most comprehensive digital infrastructure provides a solid network foundation for the popularization of large models. Whether in remote rural areas or bustling cities, users can smoothly access large model services.

Another easily overlooked resource is data. China has the world’s largest internet user base, with over 1.2 billion netizens generating vast amounts of data daily. This data covers all aspects of life, work, and production, serving as the most valuable raw material for training large models.

Resource competition is a protracted battle. It does not bring overnight surprises like algorithmic breakthroughs, but it determines a country’s long-term competitiveness in the AI industry. The complete industrial system and infrastructure network that China has built over decades are now unleashing new vitality in the AI era. When these resources are fully mobilized, the energy they release is astonishing.

However, resources are merely the foundation. Ultimately, what determines the height of AI development is original innovation capability. In this regard, China and the U.S. have taken distinct paths.

03

Close-Quarters Battle Among Tech Giants in Original Innovation and Application Competition

In fact, the AI competition between China and the U.S. has entered a new phase of differentiated development. The U.S. maintains its lead in original innovation, while China has taken the lead in application implementation globally. We believe that this dual-powerhouse pattern will persist for a long time to come.

There is no denying that the U.S. remains the global leader in AI foundational research. From the proposal of the Transformer architecture to the iteration of the GPT series models and breakthroughs in multimodal technology, nearly all significant original innovations have originated from the U.S. The foundational models of companies like OpenAI, Anthropic, and Google still lead domestic models by 3–12 months in terms of general capabilities, reasoning abilities, and multimodal capabilities.

However, China’s strength lies in its ability to rapidly transform the latest technologies into practical applications. China possesses the world’s most complete industrial system and the richest application scenarios, providing a vast testing ground for large models.

From intelligent quality inspection and predictive maintenance in manufacturing to intelligent customer service and personalized recommendations in the service industry, and from auxiliary diagnosis in healthcare to personalized tutoring in education, large models are taking root in various industries across China.

Furthermore, looking deeper, the fact that China’s AI large model invocation volume has surpassed the U.S. for five consecutive weeks also reflects the ongoing competition among domestic tech giants. Whether it’s Alibaba, ByteDance, Tencent, Baidu, or other small and medium-sized players,

All parties are continuously innovating in model iteration, releasing new models to the market every month. Failing to do so would likely result in becoming a laggard in the market competition.

The AI competition has only just begun. Among nations and among tech giants, the players are already at the table.