By Viktor Yoshimura – 26th of August 2026
The Circularity Brief — independent intelligence on IT asset disposition and the circular economy, Asia-Pacific & Japan.
For most of its history, the ITAD business has been measured in tonnes. Trucks, dock space, shredder throughput, weight processed — the whole operation was scaled on mass. AI is about to make that the wrong measure, and the operators who don’t see it coming are optimising for an axis that’s disappearing.
The old world: cheap, heavy, and volume-driven
A conventional server fleet is, bluntly, a lot of low-value metal. To reach a megawatt of computing you needed a lot of boxes — legacy enterprise racks might run 200 servers to the megawatt — and most of that mass was chassis, steel and commodity components worth little on resale. When it retired, you got big, heavy piles, and the economics were driven by how fast you could move and shred tonnage. The binding constraint was the loading dock and the shredder.
In that world, measuring the business by weight made sense, because weight was the work.
The new world: light, dense, and value-concentrated
AI and high-density infrastructure invert this. GPU-dense, liquid-cooled racks pack enormous compute into very few, very expensive units. To reach that same megawatt you need a fraction of the boxes — and the value concentrates hard into a handful of components: GPUs, high-bandwidth memory, high-end CPUs and NVMe storage.
So two things happen at once as racks densify: the weight per megawatt collapses, and the value per tonne climbs. A rough, deliberately-flagged skeleton of tonnes-per-megawatt by rack profile makes the shape clear (these are directional estimates, not a sourced figure — treat them as illustration):
- Legacy enterprise: ~140 tonnes/MW — volume-driven, the dock and shredder bind.
- Modern colocation: ~65–100 tonnes/MW — standard ITAD economics.
- AI training (air-cooled): ~30 tonnes/MW — testing and grading bind, not tonnage.
- AI training (liquid-cooled): ~11–14 tonnes/MW — roughly ten times the value density, roughly a tenth of the mass.
The direction is the point, not the decimals: as racks densify, tonnes-per-megawatt fall by something like an order of magnitude, while value-per-tonne rises by more.
The implication: the bottleneck moves
If the value is no longer in the mass, then the scarce, value-capturing capability is no longer the truck bay or the shredder. It’s the test bench — the ability to functionally test and grade memory and SSDs at volume, to verify GPUs, and to route each high-value component to its highest-value outlet.
A business scaled on weight is scaled on the wrong axis for what’s coming. The pile gets smaller and the value gets bigger, so throughput at the test bench — not tonnage through the shredder — becomes the constraint that determines whether you recover ten percent of a rack’s value or eighty.
Liquid cooling adds a further wrinkle. Retiring GB200/GB300-class liquid-cooled racks isn’t a shredding problem at all — it’s a de-manufacturing problem: coolant handling, cold-plate and CDU drains, dielectric fluid streams. That wave lands toward the end of the decade, and almost no one in APAC is equipped for it yet.
The read, and the challenge
My read: the operators, OEMs and hyperscalers who win the next cycle will be the ones who re-tooled around value density — test, grade, harvest — rather than weight processing. If you’re sizing a decommissioning opportunity off tonnage, or investing in a bigger shredder, you’re building for the last cycle. The next one rewards the test bench.
That’s where I land, and I’d genuinely like it tested. If you run a processing operation and you think weight still wins — where’s the flaw in this?
The tonnes-per-MW figures above are a derived, directional model built from rack-density assumptions, not a published statistic — no public tonnes-per-MW figure for decommissioned data-centre capacity exists. Treat them as illustration of the direction, not measurement.