Lounge
Connecting…
Connecting…
View all

GPU depreciation debate: why server life matters for Nvidia

How the GPU depreciation debate works: what useful-life assumptions do to Big Tech earnings, what Meta and Amazon disclosed, and how Nvidia responded.

StocksLatest9 min read30 viewsPublished Updated
Contents9 sections
  1. What is the GPU depreciation debate?
  2. Four years vs six years: the same server, different annual cost
  3. Server useful lives at Big Tech, by the filings
  4. Nvidia's rebuttal vs the bear case
  5. GPU collateral loans and used GPU values
  6. What Nvidia stock investors can track
  7. Common misreadings to avoid
  8. Frequently asked questions
  9. Sources and references

Key takeaways

  • The GPU depreciation debate is about how many years hyperscalers assume their AI servers stay useful. A longer life spreads the cost thinner and lifts reported earnings in the early years.
  • As of October 2026, Alphabet, Microsoft and Oracle use up to six years for servers, Meta uses 5.5 years and Amazon five to six, after cutting a subset of servers from six years to five in 2025.
  • Meta disclosed that its move to 5.5 years cut 2025 depreciation by about $2.92 billion; Amazon's shortening added about $1.4 billion of depreciation.
  • Nvidia told analysts its customers depreciate GPUs over four to six years and that 2020-era A100 chips still run at high utilization; critics such as Michael Burry argue the economic life is shorter.
  • Useful-life assumptions also underpin GPU-backed loans, so resale prices, rental rates and utilization are the outside tests of the accounting.

The GPU depreciation debate comes down to one estimate: how many years an AI server will stay useful. Write a GPU server off over four years and the annual expense is large; spread it over six and the expense shrinks, so reported profit rises. As of October 2026, most large US cloud providers depreciate servers over five to six years, while critics argue that Nvidia's near-annual product cadence makes the real economic life shorter. This guide explains the accounting, the dollar effects companies have disclosed, the arguments from Nvidia and its critics, and how the same assumption feeds into GPU-backed lending. It applies to companies reporting under US GAAP.

What is the GPU depreciation debate?

Depreciation allocates the cost of a long-lived asset across the years it is expected to serve. GPU servers and networking gear are expensive and last several years, so companies do not expense them in the year of purchase. They charge the cost to the income statement over an estimated useful life. That life is not set by law; management picks it based on operating experience, replacement plans and the pace of technology change, and auditors review it.

The estimate matters more now because the asset base is so large. According to a DECUS review of the filings, Microsoft's depreciation expense rose from $15.2 billion in fiscal 2024 to $22.0 billion in fiscal 2025 and $34.3 billion in fiscal 2026. At that scale, moving the useful life by six months can swing annual profit by billions of dollars. That is why investors now ask how much of AI-era earnings growth comes from operations and how much from accounting judgment.

Depreciation is not a cash outflow. The cash went out as capital expenditure when the servers were bought. Changing the useful life changes when the cost hits the income statement, not how much cash the business generates.

Four years vs six years: the same server, different annual cost

Four years vs six years: the same server, different annual cost

A simple hypothetical shows the mechanics. Assume a company buys $12 billion of GPU servers, expects no salvage value and uses straight-line depreciation. These are illustrative figures, not any company's actual numbers.

4 years

Annual depreciation
$3.0 billion
Cumulative cost by year 4
$12 billion
Total cost over life
$12 billion

5 years

Annual depreciation
$2.4 billion
Cumulative cost by year 4
$9.6 billion
Total cost over life
$12 billion

6 years

Annual depreciation
$2.0 billion
Cumulative cost by year 4
$8 billion
Total cost over life
$12 billion

The total cost is $12 billion in every case. What changes is timing. A six-year life books $1 billion less expense per year than a four-year life in the early years, so operating income looks higher. The cost does not disappear; it lingers into years five and six. If the servers actually become uncompetitive after four years and are retired, the remaining book value may have to be written off at once. That back-loaded risk is the heart of the bear case.

The effect compounds when capex grows every year, because each new cohort of servers carries the same timing gap. When spending slows, the opposite happens: book value accumulated during the boom keeps flowing through as expense and can weigh on margins.

Server useful lives at Big Tech, by the filings

Companies disclose the dollar impact when they change an estimate. Combining American Compute's schedule tracker, updated October 10, 2026, with the 10-K figures cited by DECUS gives the picture below. Note that Microsoft's latest annual report states a two-to-six-year range for servers and network equipment rather than a single number.

Alphabet

Current disclosed life
6 years
Change
Jan 2023, 4 to 6 years
Disclosed effect
About $3.9B less 2023 depreciation

Microsoft

Current disclosed life
2 to 6 years (servers, network)
Change
Fiscal 2023, 4 to 6 years
Disclosed effect
About $3.7B higher fiscal 2023 operating income

Oracle

Current disclosed life
6 years
Change
Fiscal 2025, 5 to 6 years
Disclosed effect
About $733M lower fiscal 2025 operating expenses

Meta

Current disclosed life
5.5 years (most servers)
Change
Jan 2025, 4-5 to 5.5 years
Disclosed effect
About $2.92B less 2025 depreciation; net income up about $2.59B

Amazon

Current disclosed life
5 to 6 years
Change
Jan 2024 5 to 6; Jan 2025 subset 6 to 5
Disclosed effect
About $1.4B more 2025 depreciation; net income down about $1.0B

CoreWeave

Current disclosed life
6 years (technology equipment)
Change
Jan 2023, 5 to 6 years
Disclosed effect
About $20M lower 2023 expenses

In its fourth-quarter 2024 results, Meta said it completed an assessment in January 2025 that raised the estimated life of certain servers and network assets to 5.5 years, which it expected to reduce 2025 depreciation by roughly $2.9 billion based on assets in service at year-end 2024. The same month, Amazon went the other way for a subset of servers and networking equipment, citing the increased pace of technology development in artificial intelligence and machine learning. Two buyers of similar chips reached opposite conclusions at the same time.

American Compute counts 17 changes since 2020, and 16 of them lengthened lives; Amazon's 2025 cut is the exception. The same five-to-six-year extension was worth about $20 million to CoreWeave and about $2.92 billion to Meta. The hardware did not change; the size of the asset base did.

Nvidia's rebuttal vs the bear case

The debate became a Nvidia stock story when Michael Burry, the investor known from The Big Short, argued that hyperscalers were stretching GPU lives and that the AI buildout resembled the late-1990s telecom boom. As CNBC reported, Nvidia circulated a seven-page memo to sell-side analysts in November 2025 rebutting his claims.

Nvidia said its customers depreciate GPUs over four to six years based on real-world longevity and utilization. It added that older GPUs such as the A100, released in 2020, still run at high utilization and retain meaningful economic value well beyond the two to three years critics claim. The logic is a value cascade: chips displaced from frontier training move to inference and other lighter workloads and keep earning. The idea that a GPU is money and a durable asset comes from this argument.

The filings cannot settle the question. DECUS points to three outside tests: whether resale prices for previous-generation accelerators hold into years four to six, whether rental rates for older chips still generate revenue in those years, and whether utilization stays high. If those indicators weaken, the six-year assumption comes under pressure; if they hold, Nvidia's case strengthens.

GPU collateral loans and used GPU values

The assumption reaches beyond the income statement. GPU cloud providers, often called neoclouds, borrow heavily against GPU servers and customer contracts. On August 10, 2026, CoreWeave closed a $2.6 billion delayed draw term loan (DDTL 5.5). The facility has an approximately five-year maturity, while the underlying customer contracts average about three years. CoreWeave framed that gap as lenders underwriting renewal risk on the long-term value of Nvidia GPUs. The loan priced at SOFR plus 5.50% and was rated Ba2 by Moody's and BB+ by Fitch.

Those extra two years are where the useful-life question becomes a credit question: the GPUs must be re-leased at adequate rates for the structure to work. American Compute stresses that book life is not resale value; in the conservative band of its own residual report it puts a used H100 at 11% to 19% of original list price in 2027. That is one firm's estimate, not a market quote, but it illustrates how far a six-year accounting life and secondary-market value can diverge.

Neoclouds also show the depreciation load most clearly. In CoreWeave's second quarter of 2026, as summarized by DECUS from the 10-Q, revenue was $2,575 million, depreciation on property and equipment was about $1.4 billion, net interest expense was $640 million and the net loss was $626 million. Accumulated depreciation was only 11.2% of gross cost, so the charge on equipment already owned has room to rise. Big Tech can absorb depreciation with other profits; a company whose main business is renting GPUs cannot.

Risk can travel in one direction. If a new GPU generation pushes used prices down quickly, collateral values fall, lending terms tighten and neoclouds have less capacity to buy new chips. That would weaken one of the customer groups supporting Nvidia demand.

What Nvidia stock investors can track

Useful-life assumptions do not sit on Nvidia's own income statement; Nvidia sells the chips and its customers carry the depreciation. The link is indirect, through customers' returns on AI spending, replacement cycles and access to financing. A practical routine each earnings season:

  1. 1

    Read the useful-life footnote

    In the property and equipment policy and any change-in-estimate paragraph, note the server and network life, the effective date and the disclosed impact on depreciation and earnings per share.

  2. 2

    Compare depreciation growth with capex

    When capex grows much faster than depreciation, more expense is queued up for future years. Record what management guides for next year's depreciation.

  3. 3

    Strip out the estimate effect

    Subtract the disclosed impact before comparing year-over-year earnings so you can separate operating growth from accounting-driven growth.

  4. 4

    Watch the outside indicators

    Track rental rates and resale prices for older GPUs, neocloud loan pricing and ratings, and, in Nvidia's results, data center revenue, customer concentration and product cadence.

The theddari earnings calendar lists report dates for Nvidia and its largest customers in one place; confirm dates against each company's investor relations page.

Depreciation arguments often spread through social media before anyone reads the footnotes. The theddari stock Telegram trends page and the AI Telegram trends page show which claims are circulating; check any viral number against the original filing.

Common misreadings to avoid

  • A longer life is not automatically fraud. The change and its effect are disclosed in audited statements; the question is whether the estimate is reasonable.
  • Lower depreciation does not mean more cash. Only the timing of expense recognition moves.
  • Physically running is not the same as economically useful. Power efficiency and rental pricing matter.
  • The disclosed effect is not the whole risk. It covers the year of change, not the size of a future write-down if the estimate proves wrong.

In short, the GPU depreciation debate is less about accounting fraud than about who is estimating the economic life of AI hardware more accurately. As of October 2026 most hyperscalers use five to six years, and only Amazon has shortened part of its fleet. Used GPU prices, rental rates, utilization and the terms of GPU-backed loans will test the six-year assumption. Reading the useful-life footnote alongside headline earnings, and computing growth with the estimate effect removed, lets you judge the competing claims for yourself.

Frequently asked questions

Does a longer depreciation schedule create more cash?
No. The cash left the company when it bought the servers, and depreciation only allocates that cost across future income statements. A longer schedule raises reported operating income in the early years but leaves operating cash flow and capital expenditures unchanged. Tax depreciation follows separate rules in each jurisdiction, so the book estimate and the tax effect should be analyzed separately.
Are hyperscalers manipulating earnings by extending server lives?
The filings alone cannot establish intent. A useful life is a management estimate under US GAAP, and changes are disclosed with their dollar impact in audited statements. The real question is whether the estimate reflects how fast AI accelerators become uneconomic. Meta lengthening its schedule in the same month Amazon shortened part of its fleet shows that reasonable judgments can diverge.
Would shorter GPU lives be good or bad for Nvidia stock?
It cuts both ways. Faster replacement cycles could support demand for new chips, but higher depreciation lowers the reported return on AI spending and could make customers revisit capex budgets. Neoclouds that borrow against GPUs could also face tighter credit if collateral values fall. There is no settled answer, which is why investors track both Nvidia's results and its customers' filings.
Does a useful-life change restate past results?
Generally not. Under US GAAP a change in useful life is a change in accounting estimate and is applied prospectively, from the date of the change onward. Companies disclose the effect on depreciation and earnings per share for the year of the change. When comparing year-over-year profit growth, it helps to strip that effect out.

Sources and references

  1. 1Meta Reports Fourth Quarter and Full Year 2024 Results (Meta Investor Relations)(opens in a new tab)investor.atmeta.com
  2. 2GPU Depreciation Schedules (American Compute, October 2026)(opens in a new tab)www.amcompute.com
  3. 3Why Do Hyperscalers Depreciate GPUs Over Six Years? (DECUS, October 2026)(opens in a new tab)www.decus.co
  4. 4Nvidia pushes back on charges that AI investment is a bubble (CNBC)(opens in a new tab)www.cnbc.com
  5. 5CoreWeave Closes $2.6 Billion Loan Facility (CoreWeave Investor Relations)(opens in a new tab)investors.coreweave.com