Everyone is watching what’s going up in the AI build-out. Almost nobody is talking about what’s coming out.

The construction story is well covered — gigawatts, land, grid connections, cooling, the race for capacity. Underneath it, something quieter is happening. As workloads shift into hyperscale AI facilities, a lot of traditional enterprise data centres are being consolidated, moved, or switched off entirely. And the kit inside them — servers, storage, network gear, racks by the thousand — has to go somewhere. Data Centre Decommissioning has quietly become one of the biggest operational problems in the industry and almost nobody is planning for it.

It gets wiped and redeployed. Refurbished and resold. Or scrapped.

So the construction boom everyone is writing about is quietly driving a decommissioning boom alongside it. And unlike the build-out, nobody issues a press release when a floor gets cleared.

The arithmetic of data centre decommissioning has changed

For most of the last decade, decommissioning was simple economics. Hardware got cheaper and faster every cycle. A three-year-old server was worth less than the labour required to work out what was still good in it. So end-of-life meant clear the floor, destroy the data, move on. Disposal was a cost line, and the goal was to make it small.

That assumption is now under pressure from an unexpected direction: memory.

The AI industry has poured extraordinary investment into overcoming infrastructure bottlenecks — GPUs, data centres, power, networking. Memory manufacturers have invested aggressively too. And yet memory keeps re-emerging as a constraint, with executives and analysts expecting demand to outstrip supply for years.

When memory is constrained, the arithmetic on what comes OUT of a site changes long before anything changes about what goes in.

A three-year-old server that would have been scrapped becomes worth harvesting. DIMMs that had no meaningful secondary market get one. The end-of-life decision stops being “recycle or dispose” and becomes “what can we actually recover here?”

We’re seeing it in the most direct way possible: we now get unsolicited approaches from brokers wanting to buy decommissioned GPU servers. That is not a forecast. That’s the phone ringing.

Disposal is a cost. Recovery is an operation.

This is the gap, and it’s a wide one.

Most enterprises still treat decommissioning as a disposal problem, and they’ve built their process around that assumption. Kit goes out of the door on a schedule. The emphasis is on clearing the floor by a date and destroying the data on the way. Nobody is asking what’s worth pulling, because for ten years the honest answer was “not much.”

Asset recovery is a different operation entirely. It needs someone who knows what’s in the rack before it comes out, not after. It needs functionality testing, documented. It needs a route to market. It needs different paperwork. And it needs to happen at the same time as the decommission, not as an afterthought — because once kit is on a pallet in a yard, unlabelled and unrecorded, its recoverable value has already dropped.

The organisations that make that switch will find their decommissioning projects paying for themselves. The ones that don’t will keep scrapping things they’ll want back.

The bit that’s data-bearing

There’s a security dimension that tends to get lost behind the water-and-power headlines.

A lot of the kit leaving these sites holds data. Drives leaving a decommissioned facility are a breach risk unless they’re certifiably wiped or destroyed — and “certifiably” is where this usually falls apart.

Certified by whose standard? The reference points are NIST 800-88 and, increasingly, IEEE 2883-2022. Within those, the distinction most people miss is between Clear and Purge. Clear is a logical overwrite. Purge is what’s needed for anything that held regulated data. They are not interchangeable, and a supplier who uses the words loosely is telling you something.

Then there’s the paperwork. A certificate of destruction has to name the method, the operator and the drive serial. “We wiped it” is not a certificate. If you can’t produce evidence per drive two years later when a regulator or an acquirer asks, you don’t have a defence — you have a hope.

This matters more, not less, as recovery economics improve. The more valuable a drive is, the more likely it is to be sold on rather than shredded — and a resold drive that wasn’t properly sanitised is a breach with a paper trail leading back to you.

The classification trap: used equipment or waste?

The other thing that catches people is a distinction that sounds like pedantry right up until it isn’t.

Your functional servers aren’t waste. Unless you can’t prove it — and then they are.

The distinction is UEEE (used electrical and electronic equipment) versus WEEE (waste). Functional equipment being relocated for reuse falls outside the Waste Shipment Regulation, but only if the conditions are met: functionality tested and documented before it moves, moving under a business-to-business contract to somewhere it will actually be used, with the documentation travelling with the load — not emailed afterwards, not filed at origin.

Miss any of those and the classification flips. It doesn’t matter that the servers boot fine. If you can’t evidence it at the border, the presumption isn’t in your favour, and a routine consolidation becomes an illegal waste shipment.

Where this bites hardest is exactly where the decommissioning boom is happening: consolidation projects. Someone decides to move a few hundred servers from one site to another, treats it as a logistics exercise, and nobody asks the classification question until the load is stopped.

Who feels it first

If hardware stays constrained, the organisations that feel it first probably aren’t the hyperscalers. They’re hedged, they buy forward, and they have the relationships to secure supply.

It’s the mid-market. The organisations that disposed of perfectly serviceable hardware last year, on the reasonable assumption that they could always buy more, and now find that replacing it costs more and takes longer than it used to.

That’s an unglamorous consequence of a very glamorous story, and it’s the one that will show up on someone’s budget.

What this means for infrastructure leaders

The practical implication is that decommissioning is no longer the end of a project. It is a project.

It needs planning at the same time as the migration, not bolted on at the end. It needs someone who knows what’s in the rack, what it’s worth, what has to be destroyed and what can be recovered — and who can evidence all of it afterwards. Those aren’t separable disciplines any more than they are on the way in.

And the questions worth asking are the uncomfortable ones. What’s actually in that room? Which of it is data-bearing? Who signs the certificate, and what does it say? If we’re moving it rather than scrapping it, can we prove it’s equipment and not waste?

The organisations that handle this well are the ones asking those questions before the decommission is scheduled. The ones that don’t tend to discover the answers under time pressure, with a floor to clear by Friday.

How DataMove can help

The AI story isn’t just about what gets built. It’s about what happens to everything it replaces — and that second half is a large part of what we do.

We handle data centre migration and relocation in over 58 countries, alongside asset audit and valuation, data erasure and disposal and onsite erasure and shredding including the cross-border documentation that decides whether your kit travels as equipment or as waste.

If you’d like the detail on what certified erasure actually involves, we’ve written about that here. If you’re planning a consolidation or exit, our migration strategy and site assessment service exists for exactly this — a senior engineer walks your floor and hands you a costed plan you own.

Get in touch to discuss your requirements.