Redundancy buys time; it does not repair anything. When one half of an N+1 pair fails, the Facility is running without protection and stays that way until the part is replaced. The interval between the failure and the replacement is the real exposure, and it is set almost entirely by whether the part is on the shelf.

This document therefore does two things. It defines the method by which a part is judged critical — so that the list is defensible rather than a collection of opinions — and it sets out the register itself, with the holding level for each line and the reasoning behind it.

The KOM Oman AI Factory at Knowledge Oasis Muscat: the electrical chain from utility intake to rack, the cooling plant and secondary loops, the network fabric, the compute platform, fire detection and suppression, the security estate, and the BMS and EPMS control layer. Design basis is 13,824 GPUs in 1,728 eight-GPU nodes, direct liquid cooled, at a 37.5 MW site allocation, Tier III Standard topology, ready for service December 2026.

The register covers Prima-owned plant and Prima-owned IT. Where a customer owns equipment in the hall under a colocation arrangement, spares for that equipment are the customer's responsibility and sit outside this register — the boundary in each case is the one drawn in PRM-DEM-2026-001. Prima will hold customer-owned spares on request as a chargeable service, under the same storage and handling regime at section 09.

It is not an inventory record, and nothing here should be read as evidence that a part is in stock. No equipment has yet been delivered. Quantities marked indicative are derived from the design basis and manufacturers' published failure data; they become firm at the equipment order, when model numbers, actual port counts and rack layouts are fixed. The first physical stock take is scheduled against gate L3 in the commissioning programme.

A part is critical if it passes any one of three tests. The tests are applied to the part, not to the system it sits in — a cheap relay inside a switchboard can be more critical than the switchboard, because the relay is what actually fails.

Two categories are kept off the register on purpose, because carrying them would be capital held against a risk that is already managed elsewhere.

• Annually, against actual consumption. A part consumed twice in a year is under-stocked regardless of what the model says; a part untouched in three years is reviewed for de-listing.
• On any change to the plant — equipment replacement, firmware change that alters a part number, or a topology change that alters redundancy.
• After any incident in which the absence of a part extended the outage. This is recorded as a finding against the register, not only against the incident.
• On a manufacturer's end-of-life notice, which triggers the obsolescence path at 10.3.

Every line on the register carries one of four classes. The class states where the part lives and how quickly it can be in an engineer's hands — not how important it is.

Bands are used throughout the register in place of specific dates, because a published lead time from a manufacturer is a forecast and not a commitment. The band is what the stocking decision is made against.

The exposure window is the period the design can run without the failed part before the availability commitment is genuinely at risk. It is not a comfort figure; it is derived from what redundancy remains after the failure.

Three different methods produce the quantities in the register, depending on how the part fails. Mixing them would be the common error — a wear part and a random-failure part need different arithmetic.

Optics, drives, fan modules, power supplies. These fail at a roughly constant annual rate across a large installed population, so the holding is a percentage of the installed base, floored at a minimum that covers a cluster failure.

Filters, seals, coolant, gaskets, batteries, fuel filters. These fail predictably against hours run or service interval, so the holding is one full service set plus one: enough to complete the next scheduled service and to deal with an unplanned replacement before the next order lands.

Control cards, PLC modules, protection relays, static switches, pumps. These are one-off items whose failure removes redundancy immediately. The arithmetic is not statistical.

The electrical chain carries the shortest exposure windows in the Facility, because a single failure here can take redundancy to zero in one step. Holdings are correspondingly conservative.

Three electrical items are large enough, and long enough on lead time, that holding them would be capital tied up against a risk the topology already manages. Each is named here rather than quietly omitted, with the exposure that results from not stocking it stated on the line.

Direct liquid cooling changes the spares picture materially against an air-cooled facility. The secondary loop sits inside the hall, at the rack, so a proportion of the register is wet parts in the same room as the compute — and thermal ride-through at these densities is measured in seconds, not minutes.

Three thermal items are large enough, and long enough on lead time, that holding them would be capital against a risk the topology already manages. Each is named rather than omitted, with the resulting exposure stated.

This is where the percentage method at 4.1 does the work. Quantities scale with the installed base of 1,728 nodes and 13,824 GPUs and are held per type — a correct total made up of the wrong variants is a zero holding.

Quantities scale with the installed base of 1,728 eight-GPU nodes. Nodes themselves are Band C on new supply, which is why a cold spare exists at all — it is the only route to restoring capacity inside a month while the failed unit goes to RMA.

Nothing in this section carries load, and every line still passes Test 1. These are the systems that make a failure survivable and an access decision auditable — their loss does not stop the Facility, it stops the Facility being defensible.

The control layer does not carry load and does not cool anything. What it does is make the plant visible and automatic — and every line here fails in the same way, by leaving the plant running while nobody can see or sequence it.

A part that is on site but degraded, uncalibrated or unfindable at three in the morning is not a spare. This section covers the conditions the store must hold and the disciplines that keep the stock genuinely usable.

1. Every issue is recorded against the work order or incident that consumed it, under PRM-SOP-OPS-009. An unrecorded issue is how a register silently becomes fiction.
2. Reorder is triggered on issue, not on the next stock take. The replenishment order is raised the same business day.
3. A part removed as suspect is quarantined, not returned to stock. Suspect parts are tagged, segregated and either tested or sent to RMA — never put back on the shelf.
4. Serial numbers are recorded both ways. The part fitted and the part removed are both logged against the asset, so warranty and failure history follow the equipment rather than the paperwork.
5. Emergency issue outside hours is permitted to any Shift Lead without prior approval; the record is completed within the shift. Access to a spare is never the thing that delays a repair.

The register is a model until it meets reality. Consumption is reviewed quarterly against the assumptions in section 04, and the model is corrected rather than defended.

• Consumed more than modelled — the holding is raised and the failure rate assumption revised. Repeat consumption of the same part triggers a root-cause review under PRM-IRP-2026-001, because a part failing twice is usually a symptom rather than a fault.
• Untouched for three years — reviewed for de-listing, unless it is a single-unit critical item, where a zero consumption rate is the expected and desirable outcome.
• Consumed and not replaced within lead time — recorded as a supply failure against the supplier, and considered at the annual sourcing review.

Reported to the COO on the cadence shown, and to customers and lenders on request. Fill rate is the measure that matters most: it states how often the shelf answered the question it exists to answer.

Stated plainly so that a reviewer is not left to infer it.

• Quantities are indicative against the design basis, and become firm at the equipment order when part numbers, port counts by optic type and rack layout are fixed. The register is reissued at version 1.0 at that point.
• No stock exists yet. First procurement is scheduled against the L2 and L3 commissioning gates in PRM-CDP-2026-001, so that spares are on site before the plant they protect is carrying anything.
• Class S2 lines depend on contracts not yet placed. Each in-region holding obligation must be written into the relevant supply or maintenance agreement; until it is, those lines carry the published lead time and not the contracted response.
• Preventive maintenance completion rates — the other half of the diligence line this register answers — require operating history and will first exist twelve months after service commences. They cannot be produced now by any means.