The shortage of computing power has reached a critical point, and we will face the largest scale computing power shortage in recent years. According to OpenAI's official disclosure at https://(openai. com)/zh Hans CN/index/accelerating the next phase ai/, as of the end of March, the API's processing volume per minute has exceeded 15 billion tokens, while in October 2025, this data was processing 60 tokens per minute, an increase of 2.5 times in less than half a year. The rapid popularization and outbreak of AI agents should be the main driving logic behind the shortage of computing power. The boom of agents is depleting the "computing power", leading to a serious shortage of computing power. It seems that the gold rush exhausted the "pickaxe and shovel" (computing resources) first. We have already seen the impact of computing power shortage: 1) Limited supply: Multiple AI companies are rationing their products, causing dissatisfaction among high-frequency users. 2) Reduced reliability: frequent outages. For example, Anthropic's Claude tool has experienced multiple interruptions and has tightened usage restrictions (some "unlimited" package users use it up in 45 minutes). 3) Product adjustment: OpenAI has partially removed the Sora video generation tool in order to free up computing resources for coding, enterprise level products, and new models (codenamed Spud). 4) Price increase: GPU rental prices have risen by 48% in two months. According to the computing power price index released by Ornn, a data service provider in New York, the hourly cost of renting Nvidia's latest generation Blackwell chips is currently $4.08, up 48% from $2.75 two months ago. The question is, why not deploy more devices directly? Because the delivery cycle is too long, the data center construction cycle is too long, and all the available power capacity before 2026 has been pre booked. We may be experiencing the severe impact of computing power shortage for the first time, which means that the prosperity of AI is facing a 'physical limit'. Yesterday, the underlying logic of the bottleneck transmission in the AI computing power industry chain was also discussed in this long article: it is no longer due to insufficient chips, but rather the overall infrastructure (GPU/TPU, chips such as HBM, optical modules, data centers, and power) cannot keep up. The key bottleneck links in the computing industry, computing infrastructure, and new could all benefit further from the shortage wave.