Neocloud stocks fell hard the week before this interview. Worries about borrowing costs and AI spending were growing, and a Gartner analysis suggested more than 80% of the sector may not operate independently by 2030. Shelly Li told Squawk Box that investors had priced good and bad neoclouds alike. She compares the companies to businesses that buy planes for hire. Some line up a customer for a GPU’s entire first life before they buy it. Others park the plane on the runway and wait for demand, which means higher cost and higher risk.
Asked how to tell strong operators from weak ones, Li lists the three questions she would ask on an earnings call:
-
Time to revenue: How many days pass between drawing on the credit line for a GPU and booking revenue from it?
-
Uptime: How deep is the spares pool, given the penalties customers like Meta and Microsoft charge when GPUs sit idle or go down?
-
Fleet math: Does the company know when an older GPU earning $2 an hour should give its power to one earning $8?
She also pushes back on the idea, which she attributes to the Gartner report, that GPUs are a rapidly depreciating asset. Across the tens of thousands of servers Sprout has processed, she cites residual values of roughly 40% for A100s and over 80% for H100s.
Related reading
-
Gartner: 80% of neocloud providers expected to cease independent operations by 2030: the analysis discussed in the segment.
-
How Long Does a Data Center GPU Actually Last?: Sprout’s data on A100 resale values and GPU useful life.
-
Your Next AI Infrastructure Investment May Already Be Funded: using recovered value from retired hardware to fund new AI capacity.
