Recent data from OpenRouter, the world's largest AI model API aggregation platform, shows that as of March 15, the weekly Token volume for Chinese large AI models reached 4.69 trillion, surpassing that of the United States for the second consecutive week. Chinese models have secured the top three positions globally for Token usage. J.P. Morgan predicts that China's AI inference Token consumption will grow from approximately 10 quadrillion in 2025 to about 3,900 quadrillion by 2030, representing a 370-fold increase over five years.

At the recent China Development Forum 2026 annual meeting, Liu Liehong, Director of the National Data Administration, stated that the Token is not only a value anchor in the intelligent era but also a "settlement unit" connecting technological supply and commercial demand, providing a quantifiable basis for business model implementation. He also provided the official Chinese translation for "Token" as "词元" (Cíyuán). Director Liu pointed out that at the beginning of 2024, China's daily Token volume was 100 billion; by the end of 2025, it had surged to 100 trillion; and this March, it exceeded 140 trillion, marking a growth of over a thousand times in two years. Since the end of January this year, some model companies have achieved revenue records where 20 days of income surpassed their total revenue for the entire year of 2025. These figures indicate the rapid evolution of a new business logic based on Token billing.

According to Volcano Engine data, as of December 2025, the daily Token usage of Doubao, the large model under ByteDance, exceeded 50 trillion, a more than tenfold increase compared to the same period last year and a 417-fold increase from its launch in May 2024. Google disclosed that as of October 2025, the monthly Token volume processed across its platforms had reached 1,300 trillion, equivalent to a daily average of 43.3 trillion, compared to a monthly average of just 97 trillion a year earlier. Many industry experts believe the global large model industry is transitioning towards a "Token economy," where the value of models is no longer measured solely by advanced reasoning capabilities but also incorporates commercial metrics priced in Tokens. Tokens are becoming a new valuation anchor for AI companies.

MiniMax, leveraging the explosive growth of its M2 series text models, saw its daily Token consumption surge more than sixfold compared to December 2025. This drove its full-year 2025 revenue to skyrocket 158.9% year-on-year to $79.04 million, with its gross margin significantly jumping from 12.2% to 25.4%, demonstrating exceptional commercialization efficiency. Meanwhile, Zhipu AI, relying on the rapid commercialization of its flagship GLM-5 model API services, has seen its Token consumption continuously climb—reaching a daily average of 4.2 trillion Tokens by November 2025, which propelled its 2025 revenue to achieve multiple-fold growth. Its MaaS platform has attracted over 3 million enterprise and application developers, and its API prices were cumulatively raised by 83% in the first quarter of 2026, showing a clear trend of growth in both volume and price.

A recent Goldman Sachs research report noted that Chinese AI models excel in multimodal fields such as text, video, and audio, further expanding Token application scenarios. Goldman Sachs analysts believe the surge in Token demand will lead to revenue breakthroughs for Chinese AI model companies while also boosting cloud service revenue. Kaiyuan Securities stated that "Token = AI chips (domestic computing power + computing power leasing) = AIDC," suggesting that with the continuous growth of Tokens, demand for CDN may also increase significantly. They emphasize three core investment themes: AIDC, computing power leasing, and CDN.

**Related Concept Stocks:**

**SENSETIME-W (00020):** The company's Lingang AIDC computing-power-electricity coordination platform integrates an energy large model and CATL's energy storage system, creating an intelligent scheduling system covering the entire chain from "data-algorithm-strategy-execution-verification." Through a bidirectional application loop of "computing follows electricity" and "electricity follows computing," it achieves precise matching of computing power and energy. In 2025, SENSETIME-W achieved total revenue of RMB 5.015 billion, a 32.9% year-on-year increase from RMB 3.772 billion in 2024. Its annual loss was RMB 1.782 billion, narrowing by 58.6% year-on-year. Its EBITDA for the second half of 2025 was RMB 376 million, turning positive for the first time since its listing.

**GDS-SW (09698):** Huatai Securities believes that as a leading domestic third-party data center operator, GDS-SW has abundant reserves of cabinet resources in core regions. The firm is optimistic about the potential for valuation increases and performance improvement opportunities in the company's domestic IDC business.

**XUNCE (03317):** The company positions itself as a provider of AI real-time data infrastructure. Its business model has evolved from traditional subscriptions to pricing based on Token consumption, directly benefiting from the exponential growth in enterprise-level AI calls. In 2025, the company's ARPU doubled year-on-year to RMB 5.59 million, and revenue grew by 103%. Its adjusted net profit turned positive in the second half of the year, with performance growth highly synchronized with the increase in Token volume. Its business currently covers diverse scenarios such as finance, robotics, energy, and the low-altitude economy. As a "miner and refiner" of high-quality industry data, its development direction is highly aligned with the national strategy of promoting the construction of high-quality datasets.

**ILUVATAR COREX (09903):** ILUVATAR COREX provides general-purpose GPU products and AI computing power solutions for various industries. Its general-purpose GPU products cover the Tiange and Zhikai series and are compatible with mainstream AI ecosystems both domestically and internationally, as well as various deep learning frameworks. In January of this year, the company announced its fourth-generation chip architecture roadmap, expecting to surpass NVIDIA's Rubin architecture by 2027, and simultaneously launched its "Tongyang" series of edge computing power products.