📌 Older GPUs May Extend the Payback Period for Compute Assets
The scale of AI capital expenditures is bringing a question beyond “how many new chips to buy” to the forefront: whether equipment already deployed can continue generating revenue.
Gina Martin Adams noted that U.S. AI capital expenditures already amounted to 1.3% of GDP in 2025 and are expected to rise to 2.5% this year, with investment intensity exceeding the historical peaks of telecommunications and fiber-optic investment. Such investment first drives spending on procurement, power supply, and data center construction, but the ultimate return depends not only on the performance of new equipment, but also on whether existing assets can maintain high utilization rates.
A CoreWeave contract provides a case worth watching. Its A100 contract pricing is fully locked in through 2029, while this GPU was launched in 2020. If the contract can be executed as planned, it indicates that older-generation GPUs can continue generating cash flow throughout this decade in data centers with adequate power conditions, rather than immediately losing economic value as new-generation products are introduced.
This has significant implications for assessments of cloud compute leasing, data centers, and related chip supply chains. The slower the equipment refresh cycle and the longer the remaining useful life, the lower the payback pressure on early capital expenditures; conversely, if utilization declines or customer demand shifts toward higher-performance equipment, locked-in older equipment may also face revenue pressure.
However, one A100 contract cannot represent all compute assets. The actual economic life of older-generation GPUs also depends on subsequent contracts, data center power conditions, and equipment utilization rates.
▌ Sources