Amazon and Meta just told the SEC they believe two opposite things about the same technology. Line up five hyperscalers’ disclosed useful lives and a pattern emerges, one that’s still moving.
For nearly two decades, useful-life assumptions generally moved together because the market largely agreed on how enterprise infrastructure aged. That consensus has disappeared. The important story isn’t whether servers last five years or six. It’s that the industry’s largest operators are no longer making the same assumptions about similar technology.
“The accounting is guessing. The secondary market isn’t.”
— ZACK SEXTON, CEO, REMARKETS
That’s not a rounding error, it’s the whole story of what AI is doing to the economics of retired infrastructure, and almost nobody outside a small circle of analysts is talking about it. For twenty years this number was boring: servers depreciated over three to five years, hit zero, got recycled, and nobody argued about it because nobody needed to. AI broke that. Useful life was never a physical fact about hardware, it’s a management estimate, and for the first time in two decades, that estimate is diverging sharply, company by company.
Book Value vs. What the Market Actually Pays
When a server hits zero on the books, the assumption is it’s worth close to nothing. The secondary market disagrees, and disagrees by a lot: five- and six-year-old A100- and H100-class GPUs are still trading at 40–70%+ of original value.
As memory makers redirect fab capacity toward high-margin HBM for AI accelerators, standard DDR4/DDR5 supply has tightened, pulling ordinary legacy DRAM into the same valuation gap, even in servers that never ran an AI workload. This isn’t an AI-hardware story anymore. It’s becoming an all-hardware story.
What This Means
Markets aren’t assigning value to products in isolation. They’re assigning value to the role those products continue to play within different organizations, and right now, that role is worth more than the accounting reflects.
The Bottom Line
Whichever direction a hyperscaler’s assumption moves, the hardware still has to go somewhere. Any very large volumes of AI-era compute is going to physically exit hyperscaler fleets over the next 24–36 months, at a pace and scale this industry has never had to manage. ReMarkets decommissions infrastructure for the world’s largest hyperscalers and gets real, primary-source pricing on what’s actually inside it, GPUs, CPUs, memory, and beyond.
Different assumptions create different outcomes.
ReMarkets turns insight into advantage.
Sources: Amazon, Meta, Microsoft, Alphabet, Oracle SEC filings (2019–2026 Forms 10-K/10-Q), earnings-call transcripts, and analyst commentary (Bernstein, Barclays, Introl Research, Michael Burry).