As computing power increasingly becomes a foundational infrastructure, what segment of the industry chain holds the highest potential for returns in the next stage? A recent report from Guolian Minsheng Securities suggests the industry is transitioning from merely "selling computing resources" to a higher-value model of "pricing based on intelligent utility," with the Token operations sector poised to become the next high-growth investment theme.
Demand for AI computing power continues to expand beyond expectations, driving up GPU leasing prices, while major telecom operators are accelerating their deployment within the Token operations ecosystem. The essence of Token operations lies in transforming non-standard intelligent services into standardized, tradable utility commodities. This shift upgrades the industry from "selling computing resources" and "selling model APIs" to "pricing based on intelligent utility," providing a unified framework for value measurement and profit-sharing across the industrial chain.
The value transmission logic unfolds as follows: AI Agents drive an explosion in Token demand → sustained increases in computing power prices → commoditization of computing power → shifting the value focus from "selling GPUs" to "selling Tokens" → ultimately extending to "Token operations." The core competitiveness of Token operators lies in resource integration and ecosystem synergy, with their long-term value potential expected to surpass that of traditional computing power leasing models. Telecom operators and platform providers are likely to emerge as new value hubs in the AI era.
Persistent Token "Inflation": GPU Leasing Prices Rise Across the Board The tight supply-demand situation for computing power is being repeatedly confirmed by hard data. During Nvidia's Q1 FY2027 earnings call, CFO Colette Kress stated that leasing prices for H100 chips have risen by 20% since the start of the year, while cloud pricing for A100 chips has increased by nearly 15%. Notably, customers can still generate profitable revenue from these GPUs even beyond their standard depreciation period. Cloud service provider Nebius announced an increase in the on-demand leasing price for H100 GPUs from $2.95 to $3.85 per hour.
Data from the demand side is equally striking. Citing a People's Daily report, the analysis notes that China's average daily Token call volume surged from 100 billion at the beginning of 2024 to 100 trillion by the end of 2025, reaching 140 trillion by March 2026—a more than thousandfold increase in just over two years. The explosive growth of AI Agent applications is identified as the core driver of this non-linear expansion.
Price hikes from leading domestic cloud providers further corroborate robust demand: - Alibaba Cloud: Implemented multiple rounds of price increases since March, raising computing card services by 5% to 34%, CPFS Intelligent Computing storage by 30%, DDoS protection (monthly 95th percentile model) by 20% to 50%, and large model MU unit services by approximately 2% to 7%. - Tencent Cloud: Increased prices for all AI computing-related products and services by 5% effective May 9, 2026. - Baidu AI Cloud: Raised prices for AI computing-related services by about 5% to 30% starting April 18, 2026, with parallel file storage and other products seeing a 30% increase. Some models have even experienced call throttling due to demand.
Token Operations: The Triple Logic Behind the "Next Stop" for Computing Power Leasing Token operations represent not a simple resale of computing power, but a fundamental upgrade in business model, supported by three underlying dimensions: 1. Standardization on the Supply Side: Tokenized operations can unify heterogeneous computing power from different architectures, vendors, and nodes into a standardized Token supply. Users no longer need to concern themselves with underlying hardware specifications or cluster deployment details; they can simply purchase the required number of Tokens to access large model services. This significantly lowers the barrier to using computing power while improving resource utilization through dynamic scheduling. 2. Upgrade in Commercial Value: The business model evolves from a "one-time sale of hardware resources" to the "continuous operation of intelligent services." By intelligently matching computing power with models to reduce users' per-Token consumption, it enables diversified revenue models like tiered pricing and subscription packages, transforming one-time computing revenue into sticky, recurring operational income. 3. Shift in Industry Trend: The industry is moving from the infrastructure construction phase to the phase of efficient computing power scheduling and operation. As the core medium for AI value circulation, Tokens are gradually becoming a widely recognized unified standard within the industry. For computing power leasing companies, extending services into Token operations can both unlock the monetization potential of existing computing resources and connect with rapidly growing AI inference and training demand.
The long-term competitive moat in Token operations lies in the closed-loop capability of "industry data accumulation → large model training and optimization → high-quality Token output → user feedback iteration." Companies that control important datasets and can consistently produce Tokens with low hallucination rates and high professional value are expected to build defensible advantages and enhance long-term profit margins.
Policy and Demand: Dual Drivers Propel Token Operations into a Rapid Development Window On the policy front, the state has incorporated computing power networks into its key "Six Networks" plan, alongside water networks, new-type power grids, communication networks, urban underground pipelines, and logistics networks. According to National Development and Reform Commission estimates, investment in these "Six Networks" and related fields will exceed 7 trillion yuan this year. The Ministry of Industry and Information Technology and the National Data Bureau have simultaneously launched the "Model-Data Resonance" initiative, explicitly supporting Token technology to empower 20 key industrial sectors, including steel and industrial machine tools.
The National Data Bureau has also solicited public comments on the "Implementation Plan for Promoting the Construction of High-Quality Industry Datasets," proposing to explore new dataset transaction models like token trading and to build a quantifiable, pricable dataset value system based on tokens.
Major Telecom Operators Fully Enter the Fray, Token Ecosystem Takes Initial Shape The three major telecom operators have completed their intelligent computing infrastructure layouts, becoming core participants in the Token industry with differentiated competitive strategies: - China Telecom: Has adopted a clear "All in Token" strategy, launched Token packages, and completed the first 10-billion-level Token factory centralized procurement among the three operators (with a scale of approximately 16.451 billion yuan, tax excluded). Its total intelligent computing scale reaches 91 EFLOPS, with inter-hub latency under 12 milliseconds. Its Token packages cover major domestic large models like its self-developed Xingchen, DeepSeek, and GLM. For consumers, the entry point is as low as 1 yuan for 400,000 Tokens, with a minimum monthly package of 9.9 yuan for 10 million Tokens. For SMEs, the starting cost for AI applications has dropped from the traditional 500,000 to 1 million yuan range to the thousand-yuan level. China Telecom has deployed over 250 internal AI applications, achieving cost savings exceeding 2 billion yuan. - China Mobile: Focuses on computing infrastructure, having released a Token operations ecosystem and an application ecosystem alliance, and pledged to open trillion-level Token experience packages. Its core platform, MoMA, pioneered an intensive Token operation model, intelligently matching the optimal model based on three modes: "cost priority," "effect priority," or "balance priority," with end-to-end latency under one minute. China Mobile has explicitly set a goal to double its revenue from computing services and intelligent services by the end of the "16th Five-Year Plan" period. - China Unicom: Promotes an "Agent + Token + AI Cloud" model, with an intelligent computing scale of 45 EFLOPS and the construction of seven 100-megawatt-level AIDC parks.
In 2026, all three major operators plan to increase the proportion of investment allocated to computing power within their overall declining capital expenditure budgets, signaling a comprehensive shift of resources toward intelligent computing businesses.
Comments