From July 7 to 12, 2026, the most exciting scene in the market was the performance competition among AI models.

OpenAI unveiled GPT-5.6 and ChatGPT Work, while in China, open-weight models including GLM-5.2 are rapidly narrowing the performance gap with US frontier models. With Grok and Meta also releasing models emphasizing price-performance, the cost of using frontier-grade AI is now likely to continue to decrease.

However, the exact opposite is happening in the semiconductor and data center markets.

Memory is scarce, advanced packaging orders are overflowing, and there isn't enough power and water to build data centers. Demand is spreading to metals like copper and aluminum, cooling systems, and optical communication equipment.

While models are becoming cheaper, the physical world required to actually run these models is becoming more expensive.

This is the key point I see this week.

The decline in AI model prices may not be a factor that reduces AI infrastructure investment, but rather one that enables more agents and longer tasks, thereby exploding computing consumption.

1. The Next AI Competition Is Not Model Performance, But Workload

How smart GPT-5.6 is, and which model, Claude or Grok, leads in benchmarks, are certainly important.

However, from an investment perspective, there's a more significant change than model rankings: the length and complexity of tasks people entrust to AI are rapidly increasing.

OpenAI revealed that its internal agentic token usage has increased by approximately 22 times over the past six months, and the proportion of computing used for internal coding inference has grown 100-fold. Codex has been integrated into ChatGPT Work, and Codex's weekly users have already surpassed 5 million.

In the past, AI generated one answer for one question.

Today's agents search for information, read documents, write code, correct errors, review results, and then rework them. Even if a user gives a single command, dozens to hundreds of model calls can occur internally.

A new input factor has entered the production function of knowledge work.

Previously, it was roughly as follows:

Human Time × Skill Level

In the age of agents, one more factor is added:

Human Judgment × Supervised Agent Labor × Computing

Even if the number of users no longer significantly increases, the tokens and computing consumed per person can continue to grow. AI infrastructure demand will not solely depend on 'user growth' but will gain a second engine: increased consumption per user.

However, just because many tokens are generated does not mean their value increases at the same rate.

More time may be required to review and revise AI-generated output. As agents take on execution, a company's bottleneck shifts from execution itself to defining what to instruct and judging whether the results are correct.

AI is not eliminating domain expertise; rather, it is likely to increase the value