Our October 9 story on what happens to a data center's hardware at the end of its life covered the accounting side of a big question: how long does an AI chip last? It reported on Meta's and Amazon's changes to the assumed lives of their servers, the dispute between critics such as investor Michael Burry and defenders such as Yardeni Research, and used-chip prices from a dealer. This follow-up covers what that story left out: what renters are paying for older chips, how Nvidia's release schedule compares, and how lenders are treating chips that are booked as lasting six years. The evidence from 2026 leans toward the longer-life side so far, but nearly all of it traces back to a shortage.
What renters are paying
The clearest data point runs against the expectation that new chips make old ones cheap. SemiAnalysis reported that the one-year contract rate for an Nvidia H100, a chip that launched in 2022, rose from $1.70 per GPU-hour in October 2025 to $2.35 in March 2026, an increase of almost 40%. It said demand for H100s held firm even as newer Blackwell chips ramped up, and that some H100 contracts were renewing at the same rate as two to three years earlier, with some extended to 2028. Half the providers it contacted were sold out, and most had no Hopper-generation capacity coming off contract.
SemiAnalysis tied the squeeze to heavy demand from agentic AI workloads, open-weight models and AI lab fundraising, along with a surge in memory prices that led some operators to delay or drop deployments. Its authors drew a conclusion for the accounting debate: higher rental rates extend the economic life of existing GPUs, because the money invested keeps earning for longer before it needs replacing. That is the part to read carefully. The price rise happened because supply was short, so it shows an old chip is worth more when new ones are scarce. It does not show how much the chip will be worth when supply catches up.
What used chips sell for
Resale data gives a second, less tidy reading. Silicon Data, which tracks GPU listings, measures used and refurbished H100s against new ones, counting age from the H100's October 2022 launch. Its main table shows used H100s at a median of about 61% of the contemporaneous new price in 2024, when the chips were roughly two years old, and about 69% in 2025, at about three years. That sits a little below the 72% to 75% that dealer Caladan Semiconductor reported in the October 9 story. Both are listing prices from sellers, not audited sales, and Silicon Data says its numbers are not an accounting depreciation schedule. A different section of Silicon Data's page gives lower figures, 45% to 55% for used units at three years, so the page itself does not settle on one answer.
The baseline matters too. Silicon Data also compares resale prices with the highest price a new H100 ever listed at. By that yardstick, three-year-old chips sit at about 20% to 30%. A buyer who paid the peak price and a buyer who paid today's price have had very different experiences, and the share of a chip's value that survives depends on which one you are.
Nvidia's calendar
The supply side keeps moving. At its March 2025 developer conference Nvidia said it was “striving to announce new chip families on an every-year basis,” where before the AI boom it released new architectures every other year, according to CNBC. The company said Blackwell Ultra would ship in the second half of 2025, Vera Rubin systems in the second half of 2026, a Rubin Next generation in the second half of 2027, and a chip called Feynman in 2028. It also claimed cloud providers could earn up to 50 times the revenue from Blackwell Ultra compared with the Hopper generation, a company claim.
Whether older chips keep their rental value once newer ones arrive in volume is the test that has not yet been run. theCUBE Research argues they will, describing a cascade in which the newest hardware does foundation-model training for its first years, older chips move to real-time inference, and fully depreciated ones handle batch work. It expects hyperscalers to settle on a five-year schedule, down from six, but not two or three. That is an analyst framework with no measured utilization data behind it. The counterpoint the Motley Fool cited is a late-2024 Tom's Hardware report that data center GPUs running AI workloads at 60% to 70% utilization may last one to three years.
How lenders are treating the chips
The clearest example of money riding on the answer is CoreWeave, the specialist GPU cloud. It extended the assumed life of its technology equipment from five years to six at the start of 2023, a couple of years before its IPO, the Motley Fool reported. Its annual report shows what depends on it: technology and infrastructure depreciation and amortization rose from $843 million in 2024 to $2.3 billion in 2025, and total indebtedness stood at $21.6 billion at the end of 2025, with $4.4 billion of debt service paid during the year, including $3.4 billion of principal.
In March 2026 CoreWeave closed an $8.5 billion delayed-draw term loan that it described as the first investment-grade rated financing secured by high-performance computing infrastructure and an associated customer contract. Moody's rated it A3 and DBRS rated it A (low). The initial draw is up to about $7.5 billion, rising to $8.5 billion as the assets reach stabilization. Pricing is SOFR plus 2.25% on the floating portion and about 5.9% on the fixed portion, and the loan matures in March 2032, about six years after closing. Blackstone Credit & Insurance anchored the deal, with MUFG and Morgan Stanley as structuring agents, and the release says it was meaningfully oversubscribed. It describes the loan as non-recourse and secured by substantially all assets of a CoreWeave subsidiary, and it points to previously contracted cloud services without naming the customer.
The release does not say how the loan is sized against the chips. That matters because a lender repaid out of a long customer contract cares less about what the hardware is worth, so long as the chips keep running and the customer keeps paying. A lender counting on resale value cares a great deal. The sources reviewed here do not show which of the two the A3 rating leans on, and they do not include the lenders' or rating agencies' own analyses.
What to watch
Three things will say more than any forecast. Vera Rubin's shipments, due in the second half of 2026, will show whether the yearly cadence holds. Rental rates for H100s will show whether the 2026 increase reflects a lasting need for older chips or a shortage that eases once new supply arrives. And the next round of annual reports, due in early 2027, will show whether the big companies keep, stretch or shorten the lives they assign to their chips. For now the market is paying more for older chips than SemiAnalysis said many had expected, for reasons that may not last.
- The Data Center Digest — “The Afterlife of a Data Center: Resold GPUs, Mounting E-Waste, and Who Pays for the Cleanup”
- SemiAnalysis — “The Great GPU Shortage: Rental Capacity”
- Silicon Data — “H100 GPU market value trends”
- CNBC (via NBC New York) — “Nvidia announces Blackwell Ultra and Vera Rubin AI chips”
- theCUBE Research — “Breaking Analysis: Resetting GPU depreciation”
- The Motley Fool — “Michael Burry's latest warning could be bad news for CoreWeave”
- CoreWeave — “Form 10-K for fiscal year 2025”
- CoreWeave — “CoreWeave Closes Landmark $8.5 Billion Financing Facility Achieving First Investment Grade Rated GPU-backed Financing”