伯恩斯坦：AI泡沫的“核心争议”，GPU真的能“用”6年吗？

Core Viewpoint - The debate surrounding the economic lifespan of GPUs is central to understanding the profitability of tech giants and the potential AI valuation bubble, with Bernstein supporting a 6-year depreciation period while critics like Michael Burry argue for a shorter lifespan of 2-3 years, suggesting accounting manipulation to inflate profits [3][14]. Group 1: GPU Depreciation and Economic Viability - Bernstein analysts argue that a 6-year depreciation period for GPUs is economically reasonable, as the cash costs of operating older GPUs are significantly lower than market rental prices, making it feasible to extend their usage [4][6]. - The report indicates that even 5-year-old NVIDIA A100 chips can still yield "comfortable profits," and only GPUs from the 7-year-old Volta architecture approach the cash cost breakeven point [4][6]. - The demand for computing power remains strong, supporting the value of older GPUs, as leading AI labs are willing to pay for any available computing resources, regardless of the model's age [9][10]. Group 2: Accounting Practices and Market Concerns - Michael Burry warns that tech giants are artificially inflating short-term profits by extending the useful life of assets, with predictions that this accounting practice could lead to an inflated profit of $176 billion for major tech companies from 2026 to 2028 [14]. - Burry specifically highlights that companies like Meta, Alphabet, Microsoft, Oracle, and Amazon are extending their depreciation periods to 6 years, despite the typical product cycle for AI chips being only 2-3 years [14]. - Amazon has recently shortened the expected lifespan of some servers and network equipment from 6 years to 5 years, reflecting differing views within the industry on hardware iteration speed [13].