GPU depreciation and useful life start as an estimate in an annual report. A company that owns GPU servers chooses how many years to spread their cost over, and in filings from 2023 to 2026 the nine companies reviewed here chose lives of up to six years [2, 4, 5, 7, 8, 9, 11, 12, 13]. Six of the seven changes those filings record made lives longer [1, 3, 4, 5, 6, 9, 10].
A book life states what management expects. How long a GPU keeps working and earning shows up in other records: tax rules, failure studies of large clusters, rents for older GPU classes, contract statements and a few resale disclosures. This report reads each one on its own terms and tags every GPU price figure with its evidence class, as evidence classes in GPU pricing defines them. A listing is a posted price, an indicative price is a level stated for reference, and a reported sale is a sale as a buyer, seller or broker reports it [35].
GPU depreciation and useful life in annual reports
The latest filing of each company states these lives.
| Filer | Asset line | Life | Source |
|---|---|---|---|
| Alphabet | Servers and network equipment | Generally six years | [2] |
| Amazon | Servers and networking equipment | Five to six years | [4] |
| Meta | Servers and network assets | Five to 5.5 years | [5] |
| Microsoft | Servers and network equipment | Two to six years | [7] |
| Oracle | Computer, network, machinery and equipment | 1 to 6 years | [8] |
| CoreWeave | Technology equipment | 6 years | [9] |
| Nebius | Servers and network equipment | Five years from 2026 | [11] |
| IREN | HPC hardware | 5 years | [12] |
| Nscale | Technology equipment | 5 to 6 years | [13] |
Alphabet, Amazon, Meta, Microsoft and Oracle report servers with network or other equipment as one class and state no life for GPUs alone [2, 4, 5, 7, 8]. Three filers tie a life to GPUs by name. Nscale's technology equipment "primarily relates to our GPU and related infrastructure assets" [13]. IREN depreciates the GPUs pledged under its GPU financing facility over five years [12]. Nebius wrote in April 2026 that "Our main assets, GPUs, are currently depreciated over four years" [10], and its second-quarter letter reports five [11]. CoreWeave gives six years for technology equipment, and its risk factors name GPUs among the infrastructure components whose useful life it estimates [9].
The changes ran mostly one way. Microsoft moved server and network equipment from four years to six from fiscal 2023 [6], Alphabet moved servers from four to six from fiscal 2023 too [1], CoreWeave moved computing equipment from five to six from January 2023 [9], and Amazon moved servers from five to six from January 2024 [3]. Meta extended most servers and network assets to 5.5 years from January 2025 [5], and Nebius moved from four years to five from January 2026 [10, 11].
The exception is Amazon. From 1 January 2025 it moved a subset of servers and networking equipment from six years back to five, citing an increased pace of technology development [3, 4]. It also retired some equipment early and recorded about $920 million of accelerated depreciation and related charges in the fourth quarter of 2024 [3].
What a change in useful life records
A change in useful life is a change in accounting estimate, applied from the change onward [16]. Earlier statements are not restated, and the company discloses the effect on the current year's income and per-share amounts [17]. So each filing gives one year's figure. Alphabet's change cut 2023 depreciation by $3.9 billion [1]. Meta's cut 2025 depreciation by $2.92 billion [5]. CoreWeave's cut 2023 expenses by $20 million [9].
The figure depends on the assets it was measured on. In February 2025 Amazon estimated that its shortening would lower 2025 operating income by about $0.7 billion, on assets held at the end of 2024 [3]. Its 2025 annual report put the effect at $1.4 billion of added depreciation and amortization [4].
The arithmetic is simple. A $1,000 asset depreciated straight-line costs $250 a year over four years and about $167 a year over six (derived). Both schedules expense the same $1,000 (derived). A longer life moves expense later. The change records no market value for the hardware, and PwC notes that a useful life can differ from the asset's economic or physical life [16].
GPU depreciation schedules for tax
Tax runs on its own schedule. IRS Publication 946 gives computers and peripheral equipment a 5-year recovery period and a 6-year class life, uses the 200% declining balance method for 5-year property, and reports that P.L. 119-21 reinstated a 100% special depreciation allowance for certain qualified property acquired and placed in service after 19 January 2025 [18]. The publication does not name GPU servers [18]. A full first-year deduction and a six-year book life can describe the same server, and neither measures how long it works. For how ownership costs compare with rental prices, see rent or buy GPUs: a break-even method.
How long does a GPU last in a data center
The public failure research counts interruptions and replacements, not the end of working life.
- Two research clusters with over 150 million A100 GPU hours logged 6.50 and 2.34 failures per thousand node-days, and 1,024-GPU jobs failed every 7.9 hours on average [19].
- A 2024 report on a model family recorded 419 unexpected interruptions in a 54-day snapshot of pre-training its 405B model on up to 16,000 H100 GPUs, 58.7% of them from GPU issues [20].
- A study of 1,056 A100 and H100 GPUs over 2.5 years found H100 memory less resilient and H100 hardware more resilient than A100's, and estimated that 5% overprovisioning is needed to handle GPU failures [21].
- The only survival analysis found for this review, on the Titan supercomputer, covers 18,688 GPUs over the machine's nearly seven-year life [22]. About 11,000 GPU assemblies were replaced from late 2016 through much of 2017 because of a corroding resistor on the circuit board [22].
A credit agreement filed with Nscale's registration statement in September 2026 sets spare GPU server levels that change at fixed points before warranty expiry, and a trailing 12-month GPU failure rate that triggers cash reserves or added warranty cover above a threshold [14]. Every spares and failure-rate percentage is redacted [14].
Rents for older GPU classes
NVIDIA announced the Tesla V100 when it launched the Volta architecture on 10 May 2017 [24]. Lambda's pricing page lists it at $0.79 per GPU-hour on 8-GPU instances, against $4.59 for the HGX H100 and $3.19 for the HGX A100 80 GB, in prices that take effect 16 November 2026 [29]. These are posted prices (Listing). The V100 rate is about a sixth of the H100 rate (derived).
Support ends on separate schedules. Microsoft retired its V100-based NCv3 series on 30 September 2025, about eight years and four months later (derived) [24, 26]. In CUDA 13.0 the cuFFT and cuSPARSE libraries dropped support for Volta and older GPUs [25]. AWS cut posted On-Demand prices (Listing) from 1 June 2025 by up to 33% for P4d and P4de instances and up to 45% for P5 [28]. To set a year of rent against a system's price, see rent-to-value.
CoreWeave GPU useful life and re-contracting
CoreWeave depreciates technology equipment over six years, and its committed contracts averaged about five years at the end of 2025 [9]. On its call of 11 August 2026, management said "we recently signed an A100 contract that extends into 2029 at an attractive price", for a GPU class introduced in 2020 [30]. It said older-generation prices were at or above earlier levels [30]. These are management statements with no prices or counterparties. IREN's annual report states the opposite risk, that "pricing for services delivered on older generations of hardware may decline" [12].
KBRA's research of June 2026 frames the credit question as whether customers keep using the capacity at an attractive price after the original contract, beyond whether the GPUs still work [31]. Lambda's $926 million term loan, announced in August 2026, matures on 31 December 2030 and amortizes on a schedule aligned with the contracted cash flows and the useful life of the GPU infrastructure [32]. GPU financing structures explains how leases and loans carry that residual value. On GPU financing, Rillor prepares the evidence lenders weigh, including what the systems are worth today and what they are likely to be worth when the contract or lease ends, so borrower and lender can set the systems' market value beside their depreciation.
GPU resale value and residual value
The public resale record in this review has no verified price. HIVE Digital reported a gain on selling fully depreciated A-series GPUs, with no price, quantity or buyer, a reported sale [33]. American Compute, which offers to structure residual value insurance, says investors and lenders "can also book residuals, even north of 10% of equipment cost over five years", and that "Most lenders agree that over a three-year term a GPU server should fetch more than 5% of original cost" [34]. Those are the firm's own statements from proprietary data, at most indicative prices [34].
The Rillor Compute Index builds its hardware series from published listings, quotes and reported sales of GPU systems, and tags each observation with its evidence class so the classes stay separate.
Read the full report
The full report adds each filer's policy and every change with its disclosed effect, the MACRS rules, the failure studies in detail, contract and loan terms, and the evidence class of each price.