Nvidia's A100 chips hold steady at $5,000 in residual value, as sustained rental demand and long-term contracts from firms like CoreWeave challenge bearish depreciation forecasts.
Nvidia's A100 chips are maintaining a residual value of nearly $5,000, which challenges Michael Burry's thesis that Big Tech is overstating profits by under-depreciating AI hardware. The ability of these older chips to transition into inference and fine-tuning workloads suggests a longer economic life than critics anticipate, supporting Nvidia's long-term infrastructure value.
They remain economically useful by shifting from cutting-edge training to inference, fine-tuning, and lower-cost workloads.
CoreWeave's commitment to A100 GPUs through 2029 serves as a practical counter-argument to the theory that AI hardware becomes obsolete quickly. By securing long-term contracts for older architecture, CoreWeave validates the sustained economic utility of Nvidia's legacy hardware.
CoreWeave has signed customer contracts for A100 GPUs that extend through 2029.
Meta is a focal point in the depreciation debate, having explicitly extended the useful life of its server and network assets to 5.5 years. This accounting decision resulted in a $2.9 billion reduction in depreciation expense, which Michael Burry cites as a potential method for inflating reported profits.
By extending the useful life of certain servers and network assets to 5.5 years, Meta cut its 2025 depreciation expense by approximately $2.9 billion.
Nvidia's A100 GPUs are maintaining a residual value near $5,000, defying Michael Burry's warnings that Big Tech is overstating profits by extending hardware depreciation schedules. Data from Silicon Data and long-term contracts from CoreWeave suggest that older chips remain economically viable through inference and fine-tuning workloads, complicating the debate over hyperscaler accounting practices.
If companies assume their hardware remains useful for longer than it actually does, they record lower annual depreciation expenses, which can artificially inflate reported profits.
Even as newer generations arrive, older chips like the A100 are repurposed for inference, fine-tuning, and lower-cost workloads, allowing them to continue generating rental income.