赤兔资本创始人刘迪凡在接受访谈时指出,费城半导体指数的大幅下跌标志着持续三年的AI硬件牛市结束 1。在2024年10月26日的访谈时点,该指数较10月高点下跌16%,而两周后发稿时跌幅扩大至29% 1。尽管市场出现回调,刘迪凡认为当前并非泡沫破灭,而是行业进入周期后段,主要特征为供给快速释放导致算力基建账目难以平衡 1。
支撑这一判断的核心逻辑在于巨大的供需缺口风险。数据显示,已投入的2万亿美元算力基建需要企业和消费者每年支付3.7万亿美元收入才能支撑回报要求,但云厂商的巨额资本开支面临终端需求不足的风险 1。此外,供应链的迅速扩张被视为周期见顶的信号,包括磷化铟供应商从3家扩至7-8家、长鑫上市以及三星和海力士宣布建厂 1。刘迪凡预计,80%-90%的AI硬件环节将在2027年从紧缺转向平衡 1。
针对市场是否发生根本性反转,刘迪凡提出了四个关键监测指标:GPU租赁价格下跌、云厂商债券利差扩大、积压订单停滞或取消、以及应收账款上升 1。以英伟达为例,其应收账款天数已从45.4天跳升至59.6天 1。在具体定价层面,亚马逊H100 GPU租金为7-12元/小时,而Neocloud仅为4-5元/小时;后者由于融资成本高且自由现金流为负,面临更大的经营挑战 1。
Liu Difan, founder of CT Associates, stated that the recent sharp decline in the Philadelphia Semiconductor Index marks the end of a three-year bull market for AI hardware. 1 As of the interview on October 26, 2024, the index had fallen 16% from its early October high, with the drop widening to 29% two weeks later at the time of publication. 1 Liu argued that the current situation is not a bubble bursting but rather a transition into the later stages of the industry cycle, driven by rapid supply expansion that makes it difficult to balance accounts for computing infrastructure. 1
A significant concern highlighted is the mismatch between massive capital expenditures and insufficient end-user demand. 1 The analysis indicated that the $2 trillion already invested in computing infrastructure requires annual revenues of $3.7 trillion from enterprises and consumers to meet return expectations. 1 To gauge a potential market reversal, four key signals were identified: falling GPU rental prices, widening bond spreads for cloud providers, stagnation or cancellation of backlog orders, and rising accounts receivable. 1 Specifically, NVIDIA’s accounts receivable days jumped from 45.4 to 59.6. 1 Additionally, supply-side indicators such as indium phosphide suppliers increasing from three to seven or eight, the listing of CXMT, and new plant announcements by Samsung and SK Hynix were viewed as signs that the cycle has peaked. 1
Market dynamics show distinct pricing pressures, with Amazon’s H100 GPU rental rates ranging from 7 to 12 yuan per hour compared to only 4 to 5 yuan per hour for Neoclouds. 1 However, Neocloud providers face higher financing costs and negative free cash flows. 1 Looking ahead, Liu predicted that 80% to 90% of AI hardware segments would shift from scarcity to equilibrium by 2027. 1
评论
还没有评论,欢迎留下第一条。