China Post Securities: Demand for Tokens Surges, Leading to a Shortage of Computing Power, and Prices Have Risen Along the Entire Computing Power Sector.

China Post Securities released a research report stating that as AI applications penetrate deeply, the consumption of Tokens continues to grow explosively, and the bottleneck of computing power supply is becoming increasingly prominent. There is a imbalance between supply and demand for computing power, and the future is promising for the continuous increase in prices from computing power leasing to cloud to large models. The suggestions focus on: 1) computing power leasing/IDC (expectations of price increase as computing power demand grows); 2) cloud services (the most direct way to "sell shovels" and enjoy the dividends of user growth); 3) large models (benefiting from the sharp increase in API calls); 4) IM entry; 5) computing power hardware (fundamental support for computing power). The main points of China Post Securities are as follows: Token demand expansion leads to computing power shortage The expansion of token demand has become a key driving factor for the scarcity of computing power. In China, according to the National Bureau of Statistics, as of March this year, China's daily Token (term element) call volume exceeded 140 trillion, more than a thousand times the early 100 billion in 2024, and more than 40% growth in three months compared to the end of 2025's 100 trillion. Overseas, due to unprecedented user demand in April, Anthropic has experienced 7 systemic crashes so far; according to the latest official disclosure, Anthropic's annual revenue has exceeded 30 billion US dollars, more than three times the approximately 9 billion at the end of 2025, showing strong user demand. Computing power leasing, cloud, and large model sectors have entered a price increase cycle The scarcity of computing power combined with rising upstream storage costs has led to computing power leasing entering a seller's market with rising prices. Ornn data shows that in the past few months, the spot leasing prices of the entire NVIDIA GPU series in cloud data centers have risen significantly. Ornn Compute Price Index shows that the hourly rent of NVIDIA's most advanced Blackwell series chip has reached 4.08 US dollars, an increase of 48% from 2.75 US dollars two months ago; SemiAnalysis data shows that the one-year rental contract price of H100 has increased from a low point of about 1.70 US dollars/hour/GPU in October 2025 to 2.35 US dollars/hour/GPU in March 2026, a nearly 40% increase. Regarding cloud and large models, in April, Alibaba Cloud in China adjusted prices three times within four days, involving large model service platform Bailing (MU unit increased by 2%-5%), DDoS high defense (elastic 95 increased from 100 yuan to 150 yuan/Mbps/month), and DataWorks API (after adjusting the free quota, excess usage is charged on a pay-as-you-go basis); overseas, Anthropic changed Claude Enterprise from a fixed subscription to billing based on computing power consumption, with a basic monthly fee of 20 US dollars, and costs for heavy users may double to triple. Prior to this, Tencent, Baidu, and other leading companies have also successively raised prices for large models. Risk alert: Security supervision risks, technological and commercialization uncertainties, legal and liability vacuum risks, supply chain damage risks, etc.