显卡抵押成风？“大空头”唱衰英伟达芯片债：经济逻辑已崩！

Core Insights - Emerging AI cloud companies are leveraging loans backed by GPUs from a $5 trillion chip manufacturer to expand their operations and challenge tech giants [1] - These companies are renting data centers filled with AI hardware and leasing computing power to tech firms for AI model training and operation [1] - Jim Chanos has raised concerns about the debt levels of these "neoclouds," which exceed $20 billion, as they lack clear paths to profitability [1] Group 1: Financing and Debt - The financing trend initiated by CoreWeave involves asset-backed loans, allowing loss-making companies to secure funding using their assets as collateral [1] - CoreWeave and its competitor Fluidstack each have approximately $10 billion in debt backed by GPU reserves, while Lambda and Crusoe have $500 million and $425 million in GPU-backed debt, respectively [2] - CoreWeave reported a loss of about $65 million in 2024, while Fluidstack's loss was under $1 million [2] Group 2: Risks and Valuation - A significant risk associated with GPU-backed debt is the potential depreciation of Nvidia's GPUs, which could occur faster than accounted for by companies [2] - Most cloud providers, including CoreWeave, expect their AI chips to generate revenue within approximately six years, but Amazon has reduced the depreciation period for its AI chips from six years to five [2] - Chanos indicated that the rapid introduction of new AI products by Nvidia every 18 months could accelerate chip depreciation, complicating debt repayment [3] Group 3: Profitability Concerns - Chanos emphasized that many new cloud providers, including CoreWeave and Nebius, are currently unprofitable even with extended depreciation periods of 8 to 10 years [3] - High capital costs are exceeding the returns from AI data center transactions, raising concerns about the economic viability of these business models [3] - If the economic lifespan of Nvidia's chips is only three years, the overall economic logic of many transactions could collapse [3]