AI 公司正以未证实的假设支撑其资本开支：GPU 的折旧年限普遍被估计为 5–6 年，但顶级芯片通常不足一年即被下一代产品取代，且 NVIDIA 承诺每年更新一代，使设备可能在 4 年或更短时间内大幅贬值。行业快速扩张导致巨额债务购买 GPU，而折旧期限每缩短数个月即可令季度利润变动达数十亿美元。对折旧乐观化的质疑正在升高，尤其是 Michael Burry 估算 2026–2028 年间超大规模数据中心运营商可能低估折旧高达 1,760 亿美元。
在需求端，老旧 GPU 仍支撑推理业务，例如 OpenAI 仍在使用 2020 年的 Ampere 芯片；NVIDIA CFO 表示 Hopper（2022）仍在销售，CoreWeave 与 Nebius 的老芯片合同续签率亦高。然而这种需求部分源自新芯片缺货，不确定能否长期持续。部分公司甚至以“10 年有效期”假设发债融资。随着市场强调数据中心“可替换性”，若设施无法灵活重新部署老芯片，成本将迅速上升。若库存老芯片只能提供低价服务，则会占据昂贵机柜与电力资源，降低整体回报。
为降低风险，基础架构公司尝试签订 5–6 年锁定特定芯片代际的长期合同，并在部分案例中选择租赁而非购买 GPU，例如 OpenAI 正与 NVIDIA 商谈租赁协议。但风险只是重新分配：债务、设备抵押与需求波动仍可能在系统内累积。投资者与监管者担忧若折旧假设普遍失准，将放大财务脆弱性。然而截至目前，CoreWeave 等公司的五年前 Ampere 资源仍“全部满载”，反映低成本算力的持续需求。
AI companies are relying on an unproven assumption: that GPUs retain value for 5–6 years, even though top-tier models often become outdated in under a year and Nvidia releases a new generation annually. This raises financial risk because firms are borrowing heavily to buy GPUs, and even a depreciation shift of a few months can move quarterly earnings by billions. Concerns are mounting, highlighted by Michael Burry’s estimate that hyperscalers may understate depreciation by $176 billion from 2026–2028.
On the demand side, older GPUs still support inference; OpenAI continues using 2020 Ampere chips. Nvidia’s CFO says 2022 Hopper chips still sell, and CoreWeave and Nebius report strong renewals for older hardware. Yet this demand is partly driven by shortages of newer chips, making long-term viability uncertain. Some firms have issued debt assuming chips will hold value for 10 years. As data-center “fungibility” becomes critical, facilities unable to repurpose older GPUs face rising costs. If outdated chips serve only low-margin workloads, they occupy expensive rack space and electricity, reducing returns.
To mitigate risk, infrastructure providers seek 5–6-year contracts tied to specific chip generations and in some cases rent GPUs instead of buying them, as in OpenAI’s reported negotiations with Nvidia. But shifting ownership does not eliminate system-wide exposure: debt structures and collateral values still hinge on chip longevity. Investors and regulators warn that incorrect depreciation assumptions could amplify financial fragility. For now, firms like CoreWeave report that five-year-old Ampere inventory remains “fully booked,” indicating sustained demand for lower-cost compute.