Skip to content

Resources · AI Infrastructure Lifecycle

Disposition timing against the depreciation curve

Written for the finance function rather than the recycler. This page separates what has actually been disclosed or observed from what is being forecast — because a great deal of what circulates about AI hardware depreciation is the second sort presented as the first.

Why this page exists

The accounting question and the physical question have come apart

For most of the history of enterprise infrastructure, a depreciation schedule was a workable proxy for when equipment would leave a rack. For accelerated computing, that relationship has broken — and it has broken in both directions at once.

Operators are revising useful lives in opposite directions on comparable hardware, and separately, running equipment well past the point the schedule implies because it still earns.

That leaves the timing decision genuinely open, which is uncomfortable but also useful: it means the decision is yours to make on operating grounds rather than one the accounting makes for you.

Observed · disclosed in filings

What major operators have actually changed, and when

Operator Change Effective Disclosed impact
Amazon Servers and networking extended 4→5 years 1 Jan 2022 $3.6B reduction in depreciation expense
Amazon Servers extended 5→6 years 1 Jan 2024 $3.2B reduction in depreciation expense
Amazon A subset of servers and networking shortened 6→5 years, citing the increased pace of technology development in AI and machine learning 1 Jan 2025 $1.4B increase in depreciation expense
Meta Useful life extended to 5.5 years Disclosed 29 Jan 2025 Lower 2025 depreciation
Microsoft Extended 4→6 years 2022
Oracle Extended 4→5 years 2023

Source: company disclosures. The 2025 revision is the one worth reading twice — it is the first shortening in the set, and the stated reason names AI directly.

Signal and noise

Four things that are observed, and two that are being forecast at you

Observed

Operators disagree, in both directions, on identical hardware

Published useful-life assumptions currently span four to six years across major operators, and the direction of travel is not one-way — Amazon shortened its assumption in the same period Meta lengthened its. That is not a market converging on an answer. It is a market that does not have one.

Observed

Prior-generation accelerators are still contracted forward

Named operators have stated on the record that prior-generation accelerators remain fully utilised and are under contract years into the future. The only confirmed public retirement of a large accelerator fleet is a generation further back than the one the market talks about.

Observed

A power and cooling constraint is holding hardware in place

Legacy halls were built around roughly 20 kW per rack. Current-generation accelerated racks draw about 120 kW by the manufacturer's own figure. Facilities cannot host the newest hardware without a power and cooling rebuild, so the previous generation stays in service.

Observed

The published residual figures disagree by roughly a factor of two

Because they use different denominators. One index measures against today's new price, which has itself fallen sharply; another measures against original list price set during a shortage. Both can be internally consistent. Neither is comparable to the other.

Forecast

A large retirement wave arriving imminently

Widely asserted, including by parties selling disposition services. We could not find transaction volumes, filings or operator statements supporting it. Treat as a forecast.

Forecast

An 18–24 month accelerated refresh cycle

Circulated in vendor commentary and sponsored content. Every instance traced back to a party with a commercial interest, and none was accompanied by volumes actually being processed. Treat as a vendor assertion.

The decision

Four states, and what each one actually calls for

Still earning
Operating economics almost always outweigh a declining residual. The disposition question is premature, and the right action is to revisit it on a schedule rather than a hunch.
Out of service, decision unowned
The expensive case, and the most common one. Equipment that stopped earning months ago and is still in a staging area is losing value on a clock nobody is watching. This is where the largest avoidable losses in the category occur.
Out of service, decision owned
Move on a defined timetable. The relevant window is short — the observed period before material decline is measured in weeks, not quarters — and sequencing is a financial decision, not a logistical one.
Retiring against a lease or site deadline
Timing is fixed for you, so the variable is preparation. The projects that lose money here are the ones where discovery started too late for the inventory to be built before the trucks were booked.

The second row is where the money goes. It is also the only one of the four that nobody has put on an agenda, precisely because no single function owns it.

FAQ

What finance and infrastructure teams ask

Can we plan a disposition pipeline from our depreciation schedule?
Not reliably, and this is the most consequential misconception in the category. Book life and physical retirement have decoupled — the disclosed changes above show major operators revising useful-life assumptions in opposite directions within the same period, and separately, operators are contracting fully-depreciated-on-paper equipment years forward because it still earns. A depreciation schedule tells you how an asset is being accounted for. It does not tell you when it will come out of a rack.
What is a fair residual expectation?
The honest answer is that no one can give you a defensible single number today, and you should be wary of anyone who does. The published figures disagree by roughly a factor of two, and the disagreement is structural rather than a dispute about the data: the indices measure against different baselines. The useful discipline is to require any residual percentage to come with its denominator and its date, and to treat a rental-rate-derived residual as a different thing entirely from a liquidation value.
How much does waiting cost?
More than most organisations model, and the loss is invisible because it never appears as a line item. Two mechanisms compound: the equipment itself depreciates, and the market it would sell into is thin, dealer-mediated and slow to clear — so a delayed decision arrives at a worse price into a market that also takes longer to absorb it. The observed remarketing window before material decline is short. We would rather size that for your specific estate than quote you a generic percentage.
Is refurbished worth the extra handling?
The available data says yes, consistently. Tested and graded equipment commands a persistent premium over equipment sold as-is, and that premium is earned through inspection, grading and disclosed history rather than through anything about the hardware. It is one of the few figures in this market that points at an operating decision you control rather than at a market view you do not.
Should we sell outright or share the proceeds?
It is a question about certainty rather than about value. A structured buyback fixes a number early and transfers the placement risk. A revenue-share arrangement keeps the upside if placement goes well and leaves you exposed if it does not. In a thin market with wide spreads between asking and executed prices, the case for certainty is stronger than it would be in a liquid one — but that is a risk-appetite judgement, and we would rather set out both structures honestly than push you toward the one that suits us.
What is the single most expensive timing mistake?
Letting equipment sit while the decision is unowned. Every other error in this category is recoverable to some degree — a poor channel choice can be corrected, an inventory gap can be reconstructed at cost. Time cannot be recovered, and it compounds against you the entire time nobody is deciding.

Sizing the cost of waiting on a specific estate?

We would rather model your actual configuration and timeline than quote you a market-wide percentage that may not describe your equipment at all.