The forward P/E ratio of large-cap US tech stocks is only 1.7 percentage points higher than the S&P 500, which is severely oversold, comparable to a financial crisis. I now retract my "AI bubble burst theory" because I've observed that the Agent Harness paradigm is driving exponential growth in API token consumption by large-scale model vendors, while capital expenditure growth for producing state-of-the-art (SOTA) large-scale models remains linear. Now, if just one or two Palantir-like companies based on the Agent Harness paradigm and targeting end consumers emerge (like Microsoft's Copilot or Claude Cowork), the capital cycle for large-scale models will successfully shift from solely relying on narrative-driven growth to a dual-engine model driven by both narrative and cash flow. 🧐