What Is GPU Depreciation? How AI Servers Wear Out on Paper
GPU depreciation is how a company spreads the cost of AI servers over the years it expects them to work. The cash left on day one. The useful life decides how fast that cost hits profit.
EXPLAINED · Mechanism · Last reviewed: 05 Oct 2026 · Next review: after the next hyperscaler or CoreWeave 10-K / 10-Q useful-life note
GPU depreciation is how a company spreads the cost of its GPU servers over the years it expects them to be useful. The cash leaves on the day the chips are bought. Depreciation is the slice of that cost that shows up as an expense each year, and the useful life a company picks decides how big each slice is.
In short. GPU depreciation sits on the Hyperscalers and Neoclouds layers. The buyer pays for GPUs up front. Each year after, a slice of that cost hits profit. The big buyers use five or six years. The real test is whether billed hours earn the cost back before newer chips cut the hourly price.
A simple example. A company buys $1 billion of GPU servers. On a six-year life, it books about $167 million of expense a year. On a five-year life, it books $200 million. Same chips, same cash out the door, different profit. For the buyers who set these numbers, see what a hyperscaler is. For the specialist GPU clouds that carry the most GPU weight per dollar of revenue, see what a neocloud is.
How does GPU depreciation work?
Most big buyers use straight-line depreciation. Take the cost, divide by the useful life, and book the same slice each year. The slice usually lands in cost of revenue, because the servers are what the company sells time on.
Three things are worth keeping apart. The cash leaves when the servers are bought. The expense arrives slowly over the useful life. And the GPU earns money only when someone pays for its hours. Depreciation is the middle one. It is an estimate, and companies change it.
When a company changes the estimate, it applies the new life going forward. The servers already on the books get spread over the new, longer or shorter, remaining life. Nothing is restated. That is why each change comes with a dollar effect for the year it starts.
What useful life do the big buyers use?
None of the filings has a line called GPU. They report servers and network equipment, which is where AI accelerators sit. Here is what the latest annual reports say.
| Company | Server useful life (latest 10-K) | Recent change | Effect the company disclosed |
|---|---|---|---|
| Amazon | Six years; a subset of servers and networking at five | 5→6 years from 1 Jan 2024; subset 6→5 years from 1 Jan 2025, citing AI/ML pace | 2025: +$1.4B depreciation, −$1.0B net income |
| Meta | Five to 5.5 years | Most servers and network assets to 5.5 years from 1 Jan 2025 | 2025: −$2.92B depreciation, +$2.59B net income |
| Alphabet | Generally six years | None stated in FY2025 10-K note | — |
| Microsoft | Two to six years | None stated in FY2026 10-K note | — |
| CoreWeave | Six years (technology equipment) | 5→6 years from 1 Jan 2023 | 2023: −$20M total expenses |
Amazon is the odd one out. Its FY2025 10-K, filed 6 February 2026, says it moved its servers from five to six years in 2024. Then, from 1 January 2025, it moved a subset of servers and networking equipment back from six to five years. The filing gives the reason in plain words: the increased pace of technology development, particularly in artificial intelligence and machine learning. That cost it $1.4 billion of extra depreciation and $1.0 billion of net income in 2025.
Meta went the other way on the same day. Its FY2025 10-K says most servers and network assets moved to 5.5 years from 1 January 2025. That cut 2025 depreciation by $2.92 billion and added $2.59 billion to net income, or $1.00 a share.
Why does the useful life move profit so much?
Because the fleets are huge. One year of difference on a big fleet is billions of dollars of expense moved between years. The table below runs the arithmetic on $1 billion, and on CoreWeave’s $33.8 billion of technology equipment at 30 June 2026.
| Useful life | Yearly expense on $1B of GPU servers | Yearly expense on $33.8B (CoreWeave technology equipment, 30 Jun 2026) |
|---|---|---|
| 3 years | $333M | ~$11.3B |
| 4 years | $250M | ~$8.5B |
| 5 years | $200M | ~$6.8B |
| 6 years | $167M | ~$5.6B |
Read the right-hand column as an illustration only. Not all of that equipment was in service for a full year, and CoreWeave does not book it this way line by line. The point is the gap. Moving from six years to four would add nearly $3 billion a year of expense on that base. Cash would not change at all.
Who gets paid, and where is the stuck step?
The chip and server makers are paid first, on the day of purchase. Their cash does not depend on the useful life the buyer picks.
The owner of the GPUs is paid last. A hyperscaler or a neocloud earns money only as customers pay for billed hours. That is how GPU as a Service works, and why long contracts such as take-or-pay GPU contracts matter. They lock in hours before the chip ages.
The stuck step is the race between the hours and the clock. The GPU has to earn back its cost before a newer chip cuts the price an hour can fetch. A six-year life on paper does not promise six years of good pricing. If the price per hour falls faster than the paper life assumes, profit was overstated in the early years.
Lenders feel this too. When a GPU is collateral, its real earning life sets how safe the loan is. Recent examples on the site: Lambda’s investment-grade GPU loan, Amazon’s Nvidia chip leaseback, and the SPV chip leaseback analysis. For who funds the fleets in the first place, see who finances GPU clusters.
The clocks to watch now
The expense is already climbing as the 2025 and 2026 buying shows up. Meta’s Q2 2026 10-Q puts servers and network depreciation at $4.62 billion for the quarter, up from $3.12 billion a year earlier. That is about 48% more, by our arithmetic.
At CoreWeave, depreciation and amortization on property and equipment was $1.4 billion in Q2 2026, up from $553 million a year earlier. Revenue was $2.575 billion. So that one expense was about 54 cents of every revenue dollar in the quarter, by our arithmetic. Technology equipment on the balance sheet rose from $20.9 billion at the end of 2025 to $33.8 billion at 30 June 2026.
Three things to check in each new filing. Did the useful life change? Is the hourly price holding up? And are the GPUs busy enough that the hours pay? The first is an accounting choice. The second and third are the cash.
Common misconceptions
- People assume depreciation is cash going out each year. Actually the cash left on the day the servers were bought. Depreciation is a non-cash expense that spreads that cost.
- People assume a longer useful life means the company found more value. Actually it lowers the yearly expense and lifts reported profit. It does not create cash. Meta’s 2025 change added $2.59 billion of net income with no change in cash spent.
- People assume every hyperscaler agrees on how long AI servers last. Actually Amazon shortened a subset to five years from 1 January 2025 while Meta lengthened most assets to 5.5 years the same day.
- People assume a fully depreciated GPU is worthless. Actually it can keep running and billing. The risk runs the other way too: a chip can lose pricing power before its paper life ends.
FAQ
What is GPU depreciation?
GPU depreciation is the yearly expense a company books for the GPU servers it already paid for. It spreads the purchase cost over the useful life the company picks, usually five or six years at the largest buyers. The cash went out on the day of purchase.
How many years do companies depreciate GPUs over?
None of the big filings has a GPU-only line. They use server lives: Alphabet generally six years, Meta five to 5.5 years, Amazon six years with a subset cut to five, Microsoft two to six years, and CoreWeave six years for technology equipment (latest 10-Ks, filed Jan–Jul 2026).
Why did Amazon shorten server life while Meta lengthened it?
Amazon’s FY2025 10-K says a subset of servers and networking moved from six to five years from 1 January 2025 because of the faster pace of AI and machine-learning technology. Meta moved most servers and network assets to 5.5 years on the same date after its own review. Same hardware class, opposite calls.
Does a longer useful life mean more cash?
No. The cash already left. A longer life lowers the yearly expense, so reported profit goes up. Meta said its change cut 2025 depreciation by $2.92 billion and raised net income by $2.59 billion. Cash flow did not change because of the estimate.
Is depreciation the same as the GPU wearing out?
Not exactly. Depreciation is an accounting estimate. A GPU can still run after it is fully depreciated, or lose its price power earlier when a newer chip arrives. The useful question is whether its billed hours earn back the cost before that happens.
Why does GPU depreciation matter for neoclouds?
Neoclouds carry GPUs as their main asset. CoreWeave booked $1.4 billion of depreciation and amortization on property and equipment in Q2 2026 against $2.575 billion of revenue (10-Q). If hours are priced too low or sit idle, that expense still arrives.
Is this a list of stocks to buy or sell?
No. Second Order Explained pages explain how the money moves. They do not give buy or sell calls.
Related terms
- Hyperscaler — the biggest GPU buyers and who sets the useful life.
- Neocloud — specialist GPU clouds where depreciation is the main cost.
- GPU as a Service — how billed GPU hours pay the cost back.
- Take-or-pay GPU contracts — contracts that lock hours before the chip ages.
- Remaining performance obligation — contracted revenue not yet earned.
Stack doors
Layers: Hyperscalers and Neoclouds.
Get the Wire — short daily notes when useful-life, pricing, or fleet clocks move.
How we check this
Useful lives and dollar effects come from each company’s own 10-K or 10-Q, dated in the list below. We quote the filing’s scope word for word where it matters (Amazon’s “subset”, Meta’s “most”). None of these filings has a GPU-only line, so we say server life, not GPU life. The $1B and $33.8B tables are our own straight-line arithmetic with no salvage value, shown in full. They are illustrations, not any company’s reported expense.
Last reviewed: 05 Oct 2026 · Next review: after the next hyperscaler or CoreWeave 10-K / 10-Q useful-life note
Spot an error? Tell us and we will correct it and note the change here.
Sources
- Amazon Form 10-K FY2025 — servers 5→6 years (1 Jan 2024); subset of servers and networking 6→5 years (1 Jan 2025), citing AI/ML pace; +$1.4B depreciation, −$1.0B net income in 2025 (filed 6 Feb 2026)
- Meta Form 10-K FY2025 — most servers and network assets to 5.5 years (1 Jan 2025); −$2.92B depreciation, +$2.59B net income in 2025 (filed 29 Jan 2026)
- Meta Form 10-Q Q2 2026 — servers and network assets depreciation $4.62B vs $3.12B a year earlier (filed 30 Jul 2026)
- Alphabet Form 10-K FY2025 — servers and network equipment generally six years (filed 5 Feb 2026)
- Microsoft Form 10-K FY2026 — servers and network equipment two to six years (filed 29 Jul 2026)
- CoreWeave Form 10-K FY2025 — technology equipment six years; 5→6 change effective 1 Jan 2023 (filed 2 Mar 2026)
- CoreWeave Form 10-Q Q2 2026 — D&A on property and equipment $1.4B vs $553M; revenue $2,575M vs $1,212M; technology equipment $33,823M (filed 12 Aug 2026)
