Retiring AI infrastructure is an inventory problem wearing an engineering problem's clothes

Confirmed: decommissioning engineering and planning, project scoping, technical project management, national technical field crews, on-site technical execution, equipment removal and de-racking within a managed project, serialized custody, NIST 800-88 Revision 2 sanitization of storage media with per-serial certificates, technical inspection and grading to the level our current capability supports, parts harvesting, reuse and redeployment assessment, remarketing, structured buyback and certified recycling. Not claimed: verified sanitization of accelerator memory, sustained functional testing of accelerator compute nodes, manufacturer diagnostic programmes, teardown of liquid-cooled systems, fluid handling, and specialist heavy rigging beyond managed project removal. Where one of those is required we name who performs it before the project starts.

No, and neither can anyone else, which is a more useful answer than the one usually given. On-package accelerator memory is volatile — it is DRAM, and it does not retain contents across a power-down, transport and storage cycle. But there is no published standard defining a sanitization procedure for it, no manufacturer procedure to follow, and therefore no verification step that would let us issue a meaningful certificate. Any provider telling you they perform standards-compliant accelerator memory sanitization is describing something the standards do not contain. What we certify is the storage media, per serial, to NIST 800-88 Revision 2 — and we tell you exactly which components that covers.

Yes. WesternTechSystems Technical Services operates national technical field crews, and on-site technical execution is performed under our own project management against a written scope statement agreed before a date is booked. Where a project requires a specialist function outside our scope, that is named in the same document rather than discovered on the day.

It is graded on what can be established rather than on what would be convenient to assert — configuration, physical condition, documented removal from service, and the health and error history readable from the equipment itself. Where our current capability does not support sustained functional testing of a given class of hardware, we say so and price the disposition accordingly rather than describing untested equipment as certified. We would rather give you a defensible number than an optimistic one.

Because much of it becomes unobtainable afterwards. Error counters, remapping history and operating state can be read from equipment while it is still running in your rack, and once a system has been removed and shipped, recovering that history requires bringing it back up in a test environment. Capturing it during on-site execution costs a documented procedure and produces a materially better grading record — which is why it is part of our sequence rather than an optional extra.

Frequently not, and assuming it is leads to poor decisions. Memory, high-speed fabric switching and transceivers, processors and storage all carry real and often more liquid value. A valuation that prices a retired system as a stack of loose accelerators will usually be wrong in both directions at once — overstating what the accelerators will fetch and ignoring what the rest of the estate is worth. We inventory and value at component level for exactly that reason.

Through the Certificate of Value Recovery in the closeout package. It records, per serial, the disposition outcome, the grade, gross recovery, and whether that figure was observed on the individual asset or allocated from a multi-asset sale — and where it was allocated, the methodology is named on the certificate rather than held internally. Deductions carry a contract reference and are supported by processing recorded against that specific serial. Where components were harvested, their proceeds attribute back to the parent asset they came from.

Both structures are available, and which one suits you depends on how much certainty you want and when. A structured buyback gives you a fixed number early. A revenue-share arrangement gives you the upside if placement goes well. For estates being retired against a deadline the timing matters as much as the structure — this class of equipment loses value measurably while it sits, so a decision made in week one is worth more than the same decision made in month three.

It is consigned under manifest to qualified, certified downstream recyclers, who perform the physical recycling. Our role is qualifying those vendors under R2v3 Appendix A, making the routing decision, and producing the documentation — and the downstream manifests form part of the closeout package rather than arriving separately months later. We are a certified processor, not a shredding or materials-recovery operation, and we would rather say so than let the distinction blur.

It is the same programme. Where an AI estate is part of a wider facility retirement, the accelerator hardware is scoped into the decommissioning project rather than handled as a separate engagement, and the closeout reconciles as one record. See data center decommissioning.

The engineering, field execution, custody, sanitization, recovery and downstream routing described on this page are current capability. The functions we have not yet built are named on this page rather than omitted from it. We are building this category deliberately and early, and we would rather be the provider whose published scope matches its actual scope than the one that discovers the gap during your procurement process.