Lenders are starting to treat GPUs as collateral, so how long the chips last and what they’re worth afterward now matter to financing. Sprout co-founder and CEO Shelly Li joined Squawk Box to describe what Sprout sees as that hardware leaves the data center. Sprout has processed about 46,000 GPU systems, roughly 13,000 of them AI hardware from Volta through Blackwell. Li says hyperscalers are keeping AI servers for about six years before letting them go, and the largest wave reaching Sprout today is 2020-era A100 systems. Hopper systems, three and a half years after the first installs, are holding 60 to 70% of their value.
Li sees power as the long-term limit on AI capacity, a problem she expects to last 10 to 30 years. She also raises a concern for resale value: if a $150 part on a $20,000 GPU is no longer made, the GPU is worth little more than its scrap metal. Sprout refurbishes retired racks for buyers who want to run inference and other lower-level workloads without paying for the newest chips.
How Long Does a Data Center GPU Actually Last?: Sprout’s report on GPU useful life, built from fleet retirement records, rental rates and Sprout’s own sales data.
The First Major AI Refresh Cycle Has Arrived: why A100 refreshes don’t fit traditional server lifecycle models.
Inside Sprout’s New Garland, Texas Flagship: the facility where Sprout refurbishes and redeploys AI hardware.