Skip to content

Resources · AI Infrastructure Lifecycle

What the evidence says about retiring AI infrastructure

This category has generated a great deal of confident commentary and not much checkable evidence. This is our attempt at the opposite: what can be observed, what is merely forecast, and where the two are being quietly conflated.

Why we wrote this

The gap between what is happening and what is being marketed

We researched this category before building a service around it, and the most useful finding was uncomfortable: much of what is published about retiring accelerated infrastructure describes a market that has not arrived yet, using standards language that does not say what it is claimed to say.

That is not a reason to avoid the category. It is a reason to be precise inside it.

What follows is the market as we found it. Where the evidence is thin, this page says so, because a buyer making a disposition decision on a large estate is better served by a stated uncertainty than by a confident number with nothing behind it.

Market reality

Six findings that change how a retirement should be planned

The retirement wave is anticipated, not observed

Most published commentary assumes a flood of retired accelerators already arriving. The evidence does not support that yet. Hyperscale operators have stated publicly that prior-generation accelerators remain fully utilised and are being contracted years into the future, and the only confirmed public retirements are a generation further back. What is observed is strong growth in conventional data centre disposition — memory, drives and processors — rather than accelerator volume.

A power and cooling wall is holding hardware in service

The reason is physical rather than commercial. Legacy halls were designed around roughly 20 kW per rack. Current-generation accelerated racks draw several times that. Existing facilities cannot host the newest hardware without a power and cooling rebuild, so the previous generation stays where it is and keeps earning. That constraint pushes real retirement volume outward, not inward.

Book life no longer predicts physical retirement

Published useful-life assumptions across major operators range from four to six years, and the direction of travel is contested — some have shortened their assumption in the same period others lengthened theirs. Depreciation schedules and physical retirement have decoupled. Planning a disposition pipeline from an accounting policy will produce the wrong answer.

Every published residual figure needs a denominator

Credible sources disagree by roughly a factor of two on what a three-year-old accelerator retains — because one measures against today's new price, which has itself fallen sharply, and another measures against original list price set during a shortage. Both can be internally consistent and still mislead. The only safe response to a residual percentage is to ask what it is a percentage of, and as at what date.

Refurbished carries a persistent premium over used

Across the available data, tested and graded equipment holds a materially higher price than equipment sold as-is. That premium is earned through inspection, grading and disclosed history rather than through the hardware itself, which makes it one of the few numbers in this market that points at an operating capability rather than a market view.

Matched sets and packaging decide value

Accelerators mounted to a shared baseboard sell as a matched set because the buyer wants the interconnect fabric that binds them. Splitting the set destroys value. Card-format accelerators behave completely differently and are frequently more liquid despite lower performance, because far more buyers can actually use one. A valuation that treats a retired system as a stack of loose parts will be wrong in both directions at once.

Where the value is

The recovery order is close to the opposite of the attention order

Memory
Currently the strongest-evidenced recovery stream in data centre disposition, and it has nothing to do with accelerators. Pricing has moved further and faster than accelerator pricing, and it is straightforward to harvest, grade and place.
Interconnect fabric
High-speed switching and transceivers depreciate on a much cleaner curve than accelerators — current generation holds five figures, one generation back still holds five figures, two generations back collapses. Critically, switches are fungible: no matched-set problem, and a conventional facility can test them.
Processors and storage
Conventional, well-understood, and with an established secondary market and a standards-backed sanitization path.
Accelerators
The headline item and the least liquid one. Posted asking prices sit well above executed prices, matched-set constraints narrow the buyer pool, and the equipment loses value measurably while it sits in a staging area.
On-package memory
Not a recovery stream. It is mounted on the same package as the processor die and cannot be harvested as a saleable component. Any business case built on recovering it, or on precious-metal content, is built on nothing.

The unanswered question

No standard covers sanitizing an accelerator

It is the first question a security team asks about this equipment, and there is no standards body answer to give them. The sanitization standards address storage media. The manufacturer utilities stop at the drives and the management controller.

Several providers cite standards for this that do not contain the word. We think the more useful thing to publish is exactly where the guidance runs out, and what a defensible position looks like on the other side of it.

Read the full analysis

FAQ

What infrastructure and finance teams ask

Should we retire accelerated hardware now or hold it?
That is a finance question before it is a disposition question, and the honest answer is that it depends on whether the equipment is still earning. Where a cluster is fully utilised, the operating economics usually outweigh a declining residual. Where it has come out of service, the calculus inverts sharply — this class of equipment loses value measurably while it sits, and the observed window before material decline is short. The expensive outcome is the middle case: equipment that stopped earning months ago and is still in a staging area because nobody owns the decision.
Is the accelerator the valuable part of the rack?
Often not. Memory, interconnect fabric, processors and storage frequently carry more recoverable value in aggregate, and they are considerably more liquid. The accelerator is the part everyone talks about because it is the part everyone recognises. Inventorying and valuing at component level rather than by chassis is what surfaces the difference.
How liquid is the secondary market for accelerators?
Thinner than the marketing suggests. Executed transaction volumes for current-generation accelerators remain low, posted asking prices sit substantially above prices actually achieved, and the market is dealer-mediated rather than open. A liquidation model that assumes a large fleet can be placed quickly at posted prices is not supportable by the available evidence.
What makes decommissioning an AI estate physically different?
Weight, density and cooling. Current-generation accelerated racks weigh several times a conventional rack — enough that the floor loading of the room becomes a planning input rather than an afterthought — and draw several times the power. Some are entirely liquid-cooled with no air-cooled variant, which changes both removal and any subsequent testing. This is why the removal is planned as an engineering exercise rather than scheduled as a collection.
Why does no one publish a grading rubric?
It is a fair question, and the absence is conspicuous. Accelerated hardware carries a readable service history — error counts, memory remapping consumption, operating state — that is a close analogue to the reallocated-sector count everyone already accepts on a hard drive. Disclosing it would let a buyer price risk properly. Most of the market prefers the phrase tested and graded without publishing what that means.
What should we ask a provider that others will not answer?
Four things. What exactly does your certificate cover, component by component. Which parts of the removal do your own people perform and which are subcontracted, named in writing. What is your grading rubric and what do you disclose about service history. And what standard are you citing for accelerator sanitization — because if the answer names one, it is worth checking whether that document contains the word at all.

Planning a retirement rather than reading about one?

Our GPU and AI infrastructure lifecycle service covers decommissioning engineering, national field execution, custody, sanitization and recovery — with the scope boundary written down.