On October 12, according to Wall Street Insights, Professor Stijn Van Nieuwerburgh of Columbia Business School pointed out in a recent paper from the Brookings Institution that the total investment in AI data center construction in the U.S. from 2025 to 2032 is expected to reach $10.3 trillion, averaging about 3.63% of GDP. This is 1.6 times the peak period of railroad construction from 1870 to 1890 and more than three times that of the internet and fiber optic boom. The construction cost for each 100-megawatt AI training facility is approximately $4.1 billion, with about two-thirds allocated for IT equipment. In terms of financing, Meta's Hyperion data center project, valued at around $30 billion, carries $27 billion in debt, resulting in a project-level debt-to-asset ratio of 90%. However, due to the debt being held within joint ventures and special purpose vehicles, it has not been included in Meta's balance sheet. Currently, approximately $1.4 trillion in debt is allocated for such structures, with about $700 billion to $800 billion held by banks. While he did not assert that AI has formed a bubble, he believes that off-balance-sheet financing makes it difficult for the market to see the complete economic responsibility.