AI Divergence Amid Compute Shortage: Will AI Agents Become a Luxury for the Rich?

Compute Scarcity Is Creating a Class Divide in AI Access

Recently, a discussion about AI compute supply sparked widespread attention on Twitter. Its core argument points to a deep-seated contradiction in the current AI industry: Global compute is severely insufficient, and this will make complex AI Agent workflows extremely expensive—even as the cost of single-turn conversational chatbots continues to decline.

This perspective raises a sharp question: Will only the most well-funded companies and the most urgent use cases ultimately be able to afford AI Agents, while everyone else remains stuck at the chatbot level?

Single-Turn Conversations vs. Complex Agents: How Big Is the Compute Gap?

To understand this issue, we first need to see the enormous gulf in compute consumption between single-turn chatbots and complex AI Agents.

A typical ChatGPT conversation might require only a few hundred to a few thousand tokens of inference computation. But a complete AI Agent workflow—such as autonomously debugging code, writing market research reports, or performing multi-step data analysis—may require dozens or even hundreds of model calls, involving planning, execution, reflection, tool invocation, and many other steps. Its compute consumption can be 50 to 100 times or more that of a single-turn conversation.

The technical essence of AI Agent workflows lies in their autonomous decision-making and multi-step execution capabilities. Agent workflows are typically built on frameworks like ReAct (Reasoning + Acting) or Plan-and-Execute, where the model needs to perform a "think-act-observe" cycle at each step.