By Viktor Yoshimura – 4th November 2026
High-density AI hardware is turning decommissioning from a logistics job into a specialist engineering one.
Liquid cooling is being sold as a facilities story: more heat out, less power in, denser racks. For ITAD it changes the job itself. Decommissioning stops being logistics and becomes de-manufacturing, and in my experience most of the ITAD capacity in Asia-Pacific is still set up for logistics.
The racks are already here. SoftBank switched on a GB200 NVL72 platform in Japan in December 2025: 1,224 Blackwell GPUs on direct-to-chip liquid cooling, designed to scale past 4,000. KDDI’s AI data centre in Osaka Sakai, announced with HPE in June 2025, began operating on the same rack-scale system in January 2026. NVL72 racks are liquid-cooled by design, and AI builds across the region are installing the same class of rack.
What changes at the rack
An air-cooled rack leaves a data hall the way it arrived: powered down, unplugged and rolled out on its casters. A liquid-cooled rack rolls out the same way, but only after a set of steps that air-cooled kit never needed. Table 1 sets out the difference.
| Step | Air-cooled | Liquid-cooled |
|---|---|---|
| Before it moves | Power down, unplug, unrack | Isolate the coolant, drain or recover it as the design requires, disconnect at the specified interfaces, confirm it’s safe to handle and transport |
| Disconnection | Power and network | Power, network and the facility cooling interface (manifolds, CDU connections), in the order the design requires |
| Waste streams | E-waste | E-waste plus drained coolant, a separate regulated stream |
| Test and grade | Power up on a bench, burn in, grade | Can’t be powered up without a cooling loop: needs a liquid-cooled test rig, or testing in the client’s hall before removal; follow the OEM procedure for cold plates, hoses and manifolds |
| Paperwork | Asset list, certificate of destruction | Asset list, certificate of destruction, a record of how the coolant was handled |
Table 1. Basis: author’s operating experience; published GB200 NVL72 specifications. Procedures vary by cooling architecture and OEM instructions.
None of this is exotic. HPE, for one, lists a fill and drain hand pump kit as a service part for its GB200 NVL72, so draining is a planned step, not an improvisation. It’s plumbing, test engineering and waste handling. But each line is a capability an ITAD either has or doesn’t, and the client’s facilities team won’t wait on site while you find out which. Get it wrong and a leak during transport or at the receiving dock means damaged equipment, cleanup costs and a serious loss of client confidence.
It’s also the practical side of the argument I made about GPU-backed lending in The collateral is in a rack: what a lender recovers depends on getting the hardware out cleanly. For liquid-cooled kit, this is what “cleanly” means.
Why it’s harder in Asia-Pacific
Three things make this a regional problem rather than a general one.
The coolant is a separate waste question. In Japan, a liquid removed during decommissioning may fall under industrial waste rules depending on its composition, its condition and how it is classified, and that decides the collection, transport and disposal route. A provider’s existing permissions shouldn’t be assumed to cover it. A regional programme needs the same assessment market by market: identify the fluid, agree who is responsible for it, and document the authorised route for anything classed as waste.
The floors were built for air. Moving the rack is a planning job, not an equipment one. Racks well over a tonne get pushed by hand every day, but at roughly 1.5 tonnes fully loaded it’s a two- or three-person push on a checked route (ramps, thresholds, lifts, dock plates), and in markets with strict workplace-safety rules that crew and route go into the method statement. Testing is the harder part. Every test and grading bench I know in this region assumes an air-cooled server you can plug in and burn in. A liquid-cooled tray can’t be powered up without a cooling loop, so testing for resale needs a liquid-cooled test rig on the ITAD floor, or has to happen in the client’s hall before the kit is removed, which brings its own access and time problems. Without a test, there’s no resale grade, and the hardware drifts toward scrap.
And the hub model is closing. As I set out in the export doors piece, cross-border processing in the region is getting harder, not easier. A rack with coolant residue is the last thing you want to explain at a border, so this capability has to exist in-market, market by market. I haven’t yet identified an Asia-Pacific ITAD provider publicly offering liquid-cooled rack decommissioning as a defined service. That doesn’t mean the capability doesn’t exist; it means the service descriptions I’ve reviewed don’t show who is ready to deliver it.
Decommissioning stops being logistics and becomes de-manufacturing.
The runway is shorter than it looks
The obvious answer is that this is a 2030 problem. If a deployment stays in service for four to five years, much of the hardware installed in 2025 and 2026 might not reach full retirement until around 2030. That’s an assumption, not a forecast. Individual components leave earlier through failures, RMAs and mid-life upgrades. Those jobs won’t match the volume of a full retirement, but they arrive in ones and twos from sites that expect them handled properly the first time, and they expose gaps in a provider’s procedures long before the main wave.
Building the capability takes longer than buying a test rig. Waste routes, written procedures, trained crews and, hardest of all, getting through a client’s facilities audit before you’re allowed into a live hall. I’d put that at a year or more. An ITAD that starts in 2028 will be learning on the client’s floor.
There may be some value in the rack too. Some liquid-cooling components contain recoverable metals, including copper; whether that offsets a meaningful share of the cost depends on the design, the amount recovered and the cost of separating it. Either way, the ITADs best placed to keep data centre work in this region will be the ones that understand these requirements before the first large retirement programme arrives. Some will build the capability; others will bring in qualified engineering or waste partners. The work has to be planned before the equipment reaches the loading dock.
That’s my read. Where am I wrong?
The four to five year cycle and the time to build capability are the author’s estimates. Observations about regional ITAD facilities and service offerings are the author’s experience and review, not a market survey.
Sources
SoftBank Corp., launch of AI computing platform built on NVIDIA GB200 NVL72, press release, 25 December 2025.
HPE and KDDI, GB200 NVL72 deployment at the Osaka Sakai AI data centre, announced June 2025 (Converge Digest, 25 June 2025); KDDI newsroom, start of operations January 2026.
HPE, NVIDIA GB200 NVL72 by HPE QuickSpecs: about 3,245 lb (1.47 t) fully loaded; HPE Cray Fill and Drain Hand Pump Kit listed as a service/deployment part.
NVIDIA DGX GB rack scale systems user guide (2026): general handling guidance only; no rack-specific moving equipment specified.