Reasons for the rise in AI computing power prices

Demand Side: Explosive AI Applications Trigger Token Consumption Surge
The most fundamental driver of this round of price hikes is the surge in computing power demand fueled by the explosion in AI applications. The industry has dubbed 2026 the "first year of AI applications." The deployment frenzy surrounding AI agents, particularly OpenClaw, has fundamentally altered the structural characteristics of computing power demand.

Unlike traditional conversational AI, OpenClaw is task-oriented, involving long-chain workflows encompassing multi-turn understanding, task decomposition, tool invocation, state management, time triggers, and sustained execution. This cycle of planning, execution, and observation leads to an exponential increase in token consumption. Data shows that the average daily token consumption per user for OpenClaw is 20 to 50 times that of traditional chatbot users.

Global AI invocation volumes are experiencing explosive growth. According to relevant data, in March 2026, the global weekly invocation volume for large AI models surpassed 10 trillion tokens, representing a more than tenfold increase compared to the same period in 2025. While the Chinese market has seen particularly rapid growth, markets in North America and Europe are also maintaining triple-digit year-over-year growth rates.

The accelerating deployment of multimodal technologies is further driving up computing power demands. Generating one minute of video consumes approximately 10 trillion tokens, and as commercial applications for video generation models become widespread, the global supply-demand imbalance for computing power intensifies.

Supply Side: Persistent Tightness in the Hardware Supply Chain
AI chips (GPUs, HBM, etc.) have extremely high production thresholds, with manufacturing capacity heavily concentrated among a few leading companies, preventing rapid short-term expansion.

In terms of capacity structure, NVIDIA dominates the AI training chip market, while three major manufacturers—Samsung, SK Hynix, and Micron—control over 90% of the global DRAM and HBM capacity. An intense "capacity battle" is underway between AI memory demand and consumer electronics demand. Meanwhile, manufacturers actively maintain a tight supply-demand balance, further exacerbating price transmission across the entire chain.

The scarcity of advanced process nodes is particularly acute. Capacity for 3nm and below process technologies at TSMC and Samsung Electronics has already been booked by major AI chip companies through 2027.

Model Evolution: From "Compute Leasing" to "Capability Delivery"
The deeper logic behind the current price surge lies in a fundamental transformation of the computing power supply model. The industry is shifting from a traditional "compute + storage" resource leasing model to an intelligent delivery model based on "Model-as-a-Service" (MaaS). As service providers offer not just GPUs but also value-added services like scheduling optimization, model deployment, and inference acceleration, they naturally open up profit margins.

This model evolution means that computing power service providers are no longer merely renters of hardware resources but deliverers of intelligent capabilities.

Capital Expenditure Cycle: The Tech Giants' Arms Race
Combined capital expenditures for the four major cloud service providers (Google, Meta, Amazon, Microsoft) are projected to reach $665 billion in 2026, a 58% year-over-year increase.

This round of capital expenditure is on a scale far exceeding any previous cycle. Giants like Microsoft, Google, Amazon, and Meta have all raised their AI-related capex guidance in recent earnings reports, primarily for procuring AI chips, building data centers, and deploying infrastructure like liquid cooling systems. This massive capex not only drives up hardware procurement prices but also strengthens the scale advantages of leading players.