价值 300 亿美元的博客：金融基础设施中的 AI 能力感知系统性风险

On February 23, 2026, Anthropic published a blog post describing Claude Code's capabilities for COBOL modernization. IBM shares fell 13.2% in a single session, erasing approximately $30 billion in market capitalization — the largest single-session percentage decline since October 2000. This paper analyzes the event as a paradigmatic instance of "AI capability perception risk": systemic risk arising from the gap between demonstrated and market-priced AI capability, particularly for firms whose revenue depends on the perceived difficulty of legacy system modernization. We present an anatomy of the market event, including IBM's "stack multiplier" revenue exposure ($5.1B infrastructure revenue driving $15–20B+ in adjacent services), survey AI agent capabilities for legacy code transformation, and examine governance gaps in AI-driven code transformation for regulated financial systems. We conclude that AI capability perception risk is structurally underpriced, that COBOL workforce attrition will intensify it, and that existing regulatory frameworks do not address the intersection of AI capability, legacy infrastructure dependency, and market information asymmetry.