Capacity in a data centre can be counted three ways, and the three rarely agree. What the utility will deliver, what plant is actually installed, and what has been sold to customers are separate numbers — and the gap between the second and the third is where oversubscription hides. This document sets all three out for the KOM Oman AI Factory at Knowledge Oasis Muscat and states the arithmetic that connects them.

The method is deliberately conservative. Installed capacity is derived bottom-up from node count and per-node power rather than asserted as a headline figure, and it is stated at both sustained and peak draw. Contracted capacity counts only executed, binding commitments — proposals and letters of interest are listed separately and are not treated as allocations.

Every figure in this statement is a design figure. Measured operating data does not exist for a facility that is not yet energised, and none is presented here as though it did.

Two utility numbers matter and they are not the same. The substation capability is what the campus feed can physically deliver. The confirmed allocation is what Prima is entitled to draw. Diligence should always ask for the second; a developer quoting only the first is quoting someone else's capacity.

Aggregate site draw is capped at the confirmed allocation of 37.5 MW and cannot exceed it. Sustained draw at full build-out is 35.9 MW, comfortably inside that cap, and the cap permits an IT load of 26.0 MW against a sustained IT load of 24.9 MW — so ordinary burst behaviour is absorbed without the cap binding.

The cap matters because dense GPU plant diverges sharply between sustained and instantaneous draw. Left unconstrained, the arithmetic product of every worst case at once — all 1,728 nodes at their 15 kW instantaneous peak simultaneously, with the cooling plant at its 1.44 peak PUE — would call for more than the allocation. That condition is synthetic and is never permitted to arise: power capping is a design requirement of this facility, not an operational preference. Five mechanisms hold the boundary.

1. GPU power capping. Per-GPU power limits are enforced at the platform layer. The HGX B300 cap of approximately 1,100 W per GPU is the basis of the 13 kW sustained per-node figure used throughout section 04.
2. Cluster-level draw ceiling. Aggregate site draw is monitored continuously against a hard ceiling set at the confirmed allocation. The scheduling layer throttles job admission and reduces per-GPU limits before the ceiling is reached, so the allocation is a boundary the site cannot cross rather than a target it might overshoot.
3. Peak coincidence. Simultaneous instantaneous peak across all 1,728 nodes is a synthetic condition. Real mixed workloads do not present it — jobs start, checkpoint and finish at different moments — so aggregate draw tracks the sustained figure, not the sum of per-node maxima.
4. Peak PUE, not average. Grid draw is computed at the peak PUE of 1.44 throughout. Because the plant will not sit at its peak in normal operation, every figure here overstates consumption and understates headroom — deliberately, so the reconciliation holds under the worst case rather than the expected one.
5. No reliance on the substation margin. The 2.5 MW between the confirmed allocation and substation capability is treated as the utility's headroom, not as capacity available to Prima.

Installed capacity is built up from node count and per-node power, not asserted. The per-node figures are the same ones used in Prima's commercial proposals, so the numbers a customer sees in a price quotation and the numbers in this statement reconcile to each other.

Contracted capacity counts executed, binding, take-or-pay commitments only. On the date of this statement there are none. Capacity at KOM is uncommitted and available in full.

Prima has commercial proposals outstanding with several counterparties. In aggregate the capacity under discussion exceeds the capacity installed at full build-out. That is ordinary pipeline behaviour — proposals convert at well under 100% — but a reviewer will find those proposals in diligence, so the allocation discipline is stated here rather than left to be discovered.

On the basis set out in this document, Prima states that at the KOM Oman AI Factory: grid capacity confirmed to Prima is 37.5 MW; sustained grid draw at full build-out of 13,824 GPUs is 35.9 MW, leaving 1.6 MW of headroom; aggregate draw is capped at the confirmed allocation and cannot exceed it; and capacity committed under executed take-or-pay contracts is nil. No capacity is allocated to more than one counterparty, and no figure in this statement relies on capacity beyond the confirmed allocation.

• Node basis. 8 × NVIDIA HGX B300 per node at approximately 1,100 W per GPU cap — 13 kW sustained, 15 kW peak per node.
• Ancillary basis. Approximately 180 kW per 1,024 endpoints, scaled linearly: fabric ≈ 130 kW, storage ≈ 45 kW, management ≈ 20 kW.
• PUE 1.44 is the peak, not an expected average. It is the maximum the plant is designed to reach under worst-case ambient and load conditions. Grid draw throughout this statement is computed at that peak, so every grid-draw figure is an upper bound — actual draw will sit below it. No measured annualised PUE exists and none is claimed; it follows twelve months of operation after RFS.
• Draw is capped, not merely forecast. Grid draw is bounded at the 37.5 MW confirmed allocation by platform-level power capping and a scheduler-enforced cluster draw ceiling. Figures above the cap are shown only as uncapped arithmetic, never as operating states.
• Tier III Standard concurrently maintainable power and cooling. Redundancy capacity is not counted as usable capacity anywhere in this statement.
• Phasing. Phase 1 of 9,792 GPUs ready for service December 2026; Phase 2 of a further 4,032 GPUs thereafter.
• Excluded. External connectivity, customer-side equipment beyond the demarcation, and any capacity at Prima's other sites. This statement covers KOM Muscat only.

1. Execution of any offtake contract, before signature.
2. Any change to the confirmed grid allocation or to the substation capability.
3. Any change to node configuration or peak PUE that moves grid draw by more than 0.5 MW.
4. Commissioning of either phase, when measured data replaces the corresponding design figure.