Over a thousand times growth in two years! The concept of tokenization is gaining momentum, and computing power stocks are surging.

Ask AI · How the surge in token usage is driving transformation in the AI industry?

On March 25, the concept of tokens caught on. Tezheng Information and 263 both hit their daily trading limits. Companies including Huanjuan New Network, Tianqi Technology, Changyingtong, BOWAY Storage, Qiyu Technology, AoFei Data, and others surged sharply one after another.

On the information front, the National Data Bureau stated that in early 2024, China’s average daily Token (token) usage was 100 billion; by the end of 2025, it had surged to 100 trillion; and in March of this year, it has already exceeded 140 trillion—up more than 1,000-fold over two years.

Cai Zhiyu, an engineer from the Artificial Intelligence Division of the National Information Center under the National Development and Reform Commission, said in an interview that the rapid growth in average daily token usage fully shows that artificial intelligence is accelerating its transition from laboratories to thousands of industries and millions of households, becoming a tangible tool for productivity—just like water, electricity, and networks—turning into a fundamental resource for the operation of an intelligent society. The amount of token consumption is an important indicator of AI industry development and a “weather vane.”

The explosion in token demand directly drives up computing power costs, and major domestic cloud providers have recently launched dense price adjustments. On March 18, Alibaba Cloud and Baidu Intelligent Cloud officially announced price increases for products related to AI computing power in succession. Among them, Alibaba Cloud’s AI computing power, storage, and other products saw the highest increases of up to 34%; and computing card products such as the PingTouGe ZhenWu 810E led the gains. Baidu Intelligent Cloud raised AI computing-related services by 5%-30%, parallel file storage by about 30%, and the new pricing took effect starting April 18.

In fact, tokens are not only the basic unit for large-model processing of text; their commercial logic is being redefined. At the 2026 GTC conference, NVIDIA CEO Huang Renxun proposed “Token Economics,” stating that tokens are the new base currency. In the future, data centers will no longer be warehouses for storing files, but “AI factories” that produce tokens. Huang Renxun even predicted that in the future, a portion of engineers’ annual salaries will include a token budget that is as important as cash.

Fuyi Fu, a part-time researcher at SuShang Bank, said in an interview with the media that tokens have become the core measurement unit of AI, marking the industry’s shift from model competitions to commercialized deployment. It is the smallest unit for processing information in models, directly corresponding to computing power consumption and service costs, enabling clear pricing, billing, and accounting.

From industry trends, the AI industry is entering a system-level reconfiguration. A research report from Soochow Securities pointed out that AI is continuing the system-level reconfiguration trend driven by the “explosion in inference demand.” The computing power ecosystem is shifting toward an inference-oriented direction, accelerating along three evolution paths: chip in-house development, ultra-large-scale clusters, and long-term computing power binding.

This trend has been validated by overseas tech giants’ deployments. On one hand, NVIDIA proposed “Token Economics” at GTC 2026. To reduce token production costs, it released an end-to-end hardware system including the VeraRubin platform, CPO switches, and space data centers, and strengthened the CUDA and NemoClaw platform ecosystem. This demonstrates that NVIDIA is upgrading from providing system-level infrastructure for computing power production and scheduling to covering it comprehensively.

On the other hand, Tesla is advancing the Terafab wafer-fab project, attempting to achieve vertical integration of AI chips from design to manufacturing. This reflects that amid supply bottlenecks and demand uncertainty, leading companies are strengthening their ability to build, develop, and independently control computing power by establishing their own production capacity. At the same time, Nebius and Meta signed a long-term computing power agreement totaling $27 billion, and will provide dedicated computing power capacity based on the VeraRubin platform. The contract locks in long-term computing power supply.

Soochow Securities believes that, overall, driven by inference demand and agent application adoption, AI infrastructure is accelerating its parallel evolution toward systematization, heavy asset deployment, and supply lock-in.

(Statement: The article content is for reference only and does not constitute investment advice. Investors act on it at their own risk.)