Economics

GPU depreciation and useful life in 2026

The 2026 depreciation schedules for GPU servers at the hyperscalers and two listed GPU clouds, how they changed since 2020, the Burry critique and the evidence against it, and what sets a GPU's economic life.

All notesPublished Updated 9 min read
In this note
  1. GPU depreciation schedules by company
  2. How GPU server lives changed since 2020
  3. What Michael Burry argues about GPU depreciation
  4. The case for longer GPU lives
  5. Economic life vs accounting life
  6. What determines how long a GPU earns
  7. GPU depreciation numbers for a finance memo
  8. Owning or reserving GPU capacity

Most large operators now depreciate GPU servers over five to six years. Alphabet, Oracle and one listed GPU cloud use six, Microsoft uses two to six, Meta uses 5.5 for most servers, Amazon uses five to six after cutting a subset back to five in 2025, and a second listed GPU cloud moved from four years to five at the start of 2026.

Those are accounting estimates. How long a GPU keeps earning more than it costs to run is a separate question, and for the A100 and H100 the answer so far has been longer than the critics expected. Nobody has a full cycle of data for Blackwell yet.

GPU depreciation schedules by company

Every figure below comes from the company's own filing, with the asset class as the company names it. None of the hyperscalers report GPUs separately from other servers. The two listed GPU clouds are labelled A and B here, and each row links to the filing.

CompanyAsset class in the filingUseful life nowHow it movedSource
MicrosoftServers and network equipment2 to 6 yearsServers 3 to 4 years (fiscal 2021), then 4 to 6 years (fiscal 2023)10-K FY2026 (opens in a new tab)
AlphabetServers and network equipment6 yearsServers 3 to 4 years (2021), then 4 to 6 years (2023)10-K 2025 (opens in a new tab)
MetaServers and network assets5 to 5.5 yearsMost assets 4 to 4.5 years (Q2 2022), then 5 years (Q4 2022), then 5.5 years (January 2025)10-K 2025 (opens in a new tab)
AmazonServers and networking equipment5 to 6 yearsServers 4 to 5 years (2022), 5 to 6 years (2024); a subset cut back from 6 to 5 years (2025)10-K 2025 (opens in a new tab)
OracleServers and networking equipment6 years4 to 5 years (fiscal 2023), then 5 to 6 years (fiscal 2025)10-K FY2026 (opens in a new tab)
GPU cloud ATechnology equipment6 yearsData-center computing equipment 5 to 6 years, effective January 202310-K 2025 (opens in a new tab)
GPU cloud BServers and network equipment5 years4 to 5 years, effective January 202620-F 2025 (opens in a new tab)

GPU cloud B is the only one whose filing names GPUs directly. Its 2025 20-F says "our main assets, GPUs, are currently depreciated over four years" and then announces the move to five, and its Q2 2026 results (opens in a new tab) say the change has applied since the start of 2026.

How GPU server lives changed since 2020

Between 2020 and 2023 every company in the table lengthened its server lives, most of them more than once. The reasons given were much the same each time: software that runs the fleet more efficiently, and hardware outlasting the plan. Meta's 2022 10-K put it as "expected longer refresh cycles in our data centers". The money involved was large. Microsoft estimated that its move from four to six years would add US$3.7 billion to fiscal 2023 operating income.

Amazon was the first to go the other way. From January 1, 2025 it cut a subset of its servers and networking equipment from six years back to five, citing "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning". Its 2025 10-K puts the cost at US$1.4 billion of extra depreciation and US$1.0 billion less net income for the year.

Meta moved the opposite way that same month. Taking most servers and network assets to 5.5 years cut its 2025 depreciation by US$2.92 billion and added US$2.59 billion (US$1.00 a diluted share) to net income. The depreciation it did book on servers and network assets that year was US$13.36 billion. Oracle's move to six years took US$197 million off operating expenses in the first quarter of fiscal 2025 alone, and GPU cloud B expects its change to lower 2026 depreciation by about US$167.6 million.

A longer life raises reported earnings in the years when capital spending is highest. That timing is what the critics object to.

What Michael Burry argues about GPU depreciation

In November 2025 Michael Burry, the investor known for shorting the US housing market before 2008, said Meta, Oracle, Microsoft, Alphabet and Amazon were overstating the useful life of their AI chips and understating depreciation. As reported by CNBC (opens in a new tab), he puts the actual useful life of server equipment at around two to three years.

The reasoning goes like this. NVIDIA now ships a new data-center generation every year, where it used to take two, and each one does much more work per watt. The previous generation loses frontier training first and then the better-paid inference. Stretching book lives just as spending peaks pushes cost into later years and makes current earnings look better.

The same CNBC piece quotes two executives in ways that help his case. Jensen Huang said at NVIDIA's GTC conference in March 2025 that "when Blackwell starts shipping in volume, you couldn't give Hoppers away." Satya Nadella, talking about how fast NVIDIA now moves between generations, said "I didn't want to go get stuck with four or five years of depreciation on one generation."

Burry has kept at it on his Substack. On October 1, 2026 he responded (opens in a new tab) to an NVIDIA investor slide titled "Nvidia AI Infrastructure Retains Value Beyond Accelerated Depreciation Schedules", which plots A100, H100 and B200 retained value against a five-year accelerated depreciation curve.

The case for longer GPU lives

The other side points to what older GPUs have done since 2023.

The A100 launched in May 2020 and is still under contract. GPU cloud A's September 2026 investor presentation (opens in a new tab) shows an A100 customer renewal running into 2029, signed as a three-year contract at pricing in line with typical one-year terms, next to an H200 renewal priced at a premium to the original contract. GPU cloud A's CEO told CNBC in November 2025 that a batch of H100s from 2022 came free when a contract expired and was immediately rebooked at 95% of its original price, and that the company's A100s were fully booked. SemiAnalysis (opens in a new tab) wrote in April 2026 that H100s were "getting renewed at the exact same rate they were signed at 2-3 years ago".

Rents went up in 2026. SemiAnalysis has one-year H100 contracts rising from a low of US$1.70 per GPU-hour in October 2025 to US$2.35 by March 2026. Sprout, citing Silicon Data's index, has H100 rates going from about US$2.00 to US$3.00 between January and August 2026.

Fleets also outlast their book lives. Sprout, which resells retired data-center hardware, reports (opens in a new tab) that Microsoft ran its K80, P100 and P40 fleets for 6.9 to 8.8 years before retiring them, and its V100 fleet for 8.4 years against a six-year book life. It also cites AWS's Matt Garman saying in February 2026 that AWS had never retired an A100.

Weigh this with some care. Nearly all of it comes from parties that gain from high GPU values: NVIDIA, a neocloud, a reseller, and data and index vendors. Burry is working from outside the fleets, without their operating data. And Amazon, which has more data on its own fleet than almost anyone, is the company that shortened its lives.

Economic life vs accounting life

People mean one of four things by a GPU's "life", and the four move independently.

LifeWhat sets itTypical figure
Accounting (book) lifeManagement's estimate under GAAP, reviewed by auditors5 to 6 years at the large operators
Tax lifeTax law; in the US, MACRS5 years for computers, with 100% bonus depreciation for qualified property acquired and placed in service after January 19, 2025 (IRS Pub 946 (opens in a new tab))
Economic lifeHow long rental or token income exceeds power, space and operating costVaries by generation and market; over six years so far for the A100
Physical and support lifeFailure rates and software supportCUDA 13.0 removed support for Maxwell, Pascal and Volta, GPUs launched between 2014 and 2017 (NVIDIA (opens in a new tab))

Economic life is the one that matters for a buy-or-rent decision. It usually ends in one of two ways: the compute hour the GPU can sell is worth less than the power, cooling and operations it needs, or the megawatts it occupies would earn more with a newer GPU in them. In a power-constrained market the second is more common, and it retires GPUs that still work and still have customers.

Most GPUs pass through a cascade before then. They spend a year or two on frontier training, move to production inference, and end up on batch inference, evaluation, fine-tuning and smaller models. Ornn Data found (opens in a new tab) that on gpt-oss-120b, a sparse model with 5.1 billion active parameters, an A100 produces output tokens more cheaply than an H100 at spot prices and at three- and five-year term prices. That's part of why A100 demand has held up.

What determines how long a GPU earns

New GPU generations

Each NVIDIA generation raises throughput per megawatt by a large multiple. NVIDIA's October 2026 post, citing SemiAnalysis data, claims Vera Rubin NVL72 delivers over 30 times the throughput per megawatt of GB300 NVL72 on DeepSeek V4 Pro. A vendor multiple on one chosen model overstates the typical gain, but the larger the jump per megawatt, the sooner the older part loses the best-paid work.

Power and data-center space

When power and finished data-center space are the bottleneck, GPUs already installed and powered keep earning because new capacity arrives late. Sprout calls power and shell, rather than chips, the binding constraint on new capacity in 2026. If power gets easier to find, the logic reverses and an operator with a fixed allocation swaps old GPUs out to push more tokens through the same megawatts.

Rental prices

Higher rents extend economic life directly, since the GPU covers its running cost for longer. The term structure shows what the market expects. Ornn reports that a five-year A100 rental is priced at 80% of the one-month rate, against 44% to 60% for the Hopper and Blackwell families, so the market expects rents on the newer parts to fall further over five years than A100 rents.

Memory

A GPU keeps earning while current models fit in its memory. For inference that often matters more than FLOPS: Barkr writes that an H200 with 141 GB "may command twice the resale value" of an 80 GB H100.

GPU depreciation numbers for a finance memo

Depreciation changes when the cost reaches the income statement. It doesn't change the cash. The table shows straight-line depreciation for each US$10,000 of server cost per GPU, spread over every hour of the year, so you can scale it to your own quote. The figures are arithmetic, not price quotes.

Book lifeDepreciation per GPU-hour, per US$10,000 of cost per GPUAnnual rateBook value left after 3 years
3 yearsUS$0.3833%0%
4 yearsUS$0.2925%25%
5 yearsUS$0.2320%40%
6 yearsUS$0.1916.7%50%

Going from six years to four raises annual depreciation by 50%. At Meta's scale, an extra half year of life was worth US$2.92 billion in one year.

Book value at year three only means something next to what the hardware would sell for then. Barkr has GPUs losing about 30% to 40% of list price in the first 12 months and then holding for 24 to 48 months. Sprout reports six-year-old 8-GPU A100 systems clearing at US$75,000 to US$85,000, or 38% to 43% of the roughly US$199,000 a DGX A100 cost at launch.

If you own the cluster, pick a life your auditors will accept and test the plan at three and four years too. If you rent or reserve, the provider carries the hardware's depreciation and resale risk and you carry the commitment.

Owning or reserving GPU capacity

If useful life is the main thing putting you off owning, a reservation leaves the depreciation and resale risk with the provider. Lightyard reserves single-tenant HGX B300 clusters and GB300 NVL72 racks in the region you choose, run to an SLA, with failed parts replaced as part of the service. The contract includes an exit option, on terms set out in your proposal. Part of the commitment can also roll into a newer GPU generation when it ships, so a faster generation cycle doesn't hold the whole 60 months to one generation. See contract terms and pricing.

Questions this note answers

Sources

  1. Microsoft Form 10-K, fiscal 2026 (opens in a new tab)
  2. Microsoft Form 10-K, fiscal 2022 (four to six years) (opens in a new tab)
  3. Microsoft Form 10-K, fiscal 2021 (three to four years) (opens in a new tab)
  4. Alphabet Form 10-K, 2025 (opens in a new tab)
  5. Alphabet Form 10-Q, Q1 2023 (four to six years) (opens in a new tab)
  6. Alphabet Form 10-Q, Q1 2021 (three to four years) (opens in a new tab)
  7. Meta Form 10-K, 2025 (opens in a new tab)
  8. Meta Form 10-K, 2022 (opens in a new tab)
  9. Amazon Form 10-K, 2025 (opens in a new tab)
  10. Amazon Form 10-K, 2021 (four to five years) (opens in a new tab)
  11. Oracle Form 10-K, fiscal 2026 (opens in a new tab)
  12. Oracle Form 10-Q, Q1 fiscal 2025 (five to six years) (opens in a new tab)
  13. Oracle Form 10-Q, Q1 fiscal 2023 (four to five years) (opens in a new tab)
  14. GPU cloud A, Form 10-K, 2025 (opens in a new tab)
  15. GPU cloud A, investor presentation, September 2026 (Form 8-K) (opens in a new tab)
  16. GPU cloud B, Form 20-F, 2025 (opens in a new tab)
  17. GPU cloud B, Q2 2026 results (Form 6-K) (opens in a new tab)
  18. CNBC: How long before a GPU depreciates? (Nov 2025) (opens in a new tab)
  19. Michael Burry: Don't Believe Your Lyin' Eyes, GPU Depreciation & Useful Lives (Oct 2026) (opens in a new tab)
  20. NVIDIA: Productive, Durable, Fungible (Oct 2026) (opens in a new tab)
  21. Sprout: The Productive Life of a Data Center GPU (Sep 2026) (opens in a new tab)
  22. SemiAnalysis: The Great GPU Shortage (Apr 2026) (opens in a new tab)
  23. Ornn Data: The Economics of Open-Weight Inference (Sep 2026) (opens in a new tab)
  24. Barkr: GPU Resale vs Rental (Apr 2026) (opens in a new tab)
  25. NVIDIA CUDA 13.0 release notes (opens in a new tab)
  26. IRS Publication 946, How To Depreciate Property (opens in a new tab)
Get started

Start with one rack.

Tell us the size and timing you have in mind. We usually reply within a few hours and come back with a written proposal.

sales@lightyard.ai