Sections 01–03 establish scope, vocabulary and what is delivered. Sections 04–06 state the measurable commitments; 07 and 08 the financial consequence of missing them. Sections 09–12 cover the operating disciplines that hold them, and 13–15 the limits, the exit right and the governance around both.

This document sets out the service levels Prima commits to for GPU compute capacity delivered at the KOM Oman AI Factory, Knowledge Oasis Muscat, Sultanate of Oman. It answers four questions a customer, a lender and an insurer each need answered in writing: what availability is promised and how it is measured; what environmental envelope is held at the rack; what compensation applies when the commitment is missed; and at what point persistent underperformance becomes a right to exit.

Commitments apply per Service Order Form and are measured per calendar month against the Committed Capacity for each data hall. Where a Service Order Form covers more than one hall or phase, each hall is measured and credited separately — a good month in one hall does not offset a bad month in another.

Capitalised terms have the meanings below. Where a term is also defined in the Master Services Agreement, the definition here applies for the purposes of measuring service levels.

The service is bare-metal GPU capacity in a purpose-built AI facility: racks powered and cooled to the committed Envelope, connected to a non-blocking internal fabric, with resident engineering support. The customer holds the workload; Prima holds everything the workload sits on.

• Contracted power draw. Operating a rack above the power draw stated in the SOF may take equipment outside the Envelope. Availability is not committed for racks operated above contracted draw.
• Named contacts. The customer maintains a current escalation list of at least two technical contacts reachable 24×7, and a named commercial contact for the monthly report.
• Access for restoration. Where physical access to customer-owned equipment is required to restore service, the customer grants or delegates access without undue delay.
• Change notification. The customer notifies Prima of firmware, image or workload changes materially altering power or thermal profile at least five business days before deployment.
• Ticket discipline. Availability events are raised through the NOC with timestamps. Informal channels do not start the Response Time clock.

Prima commits to a Cluster Availability of 99.982% per Measurement Period. A minute counts as available only if the Committed Capacity is powered, within the Envelope, and reachable at the leaf-switch customer port. Any whole minute failing any one of those three tests is an Unavailable Minute, whether or not the customer's workload was running at the time.

1. Sampling. Availability inputs are sampled at 30-second intervals across the power chain, CDU and BMS sensors, fabric port state and rack-level reachability probes. A minute is marked unavailable if any sample within it fails.
2. Ticketing. Every unavailability event is raised as a ticket carrying a start timestamp, restoration timestamp, affected racks, priority and root cause. The ticket record is the sole basis of the reported monthly figure.
3. Pro-rating. Where only part of the Committed Capacity is affected, Unavailable Minutes are multiplied by the affected share of committed GPUs, rounded up to the nearest whole rack.
4. Customer-raised events. The customer may raise an availability event through the NOC at any time. Where the customer record and the Prima record differ on start time, the earlier of the two timestamps applies.
5. Concurrency. Overlapping events are not double-counted; a minute is either available or unavailable, and the highest affected capacity share in that minute is used.
6. Retention. Availability telemetry and ticket records are retained for 24 months and made available to the customer on request within the retention period.

Each component of the service carries its own commitment, measured at its own boundary. Component figures are reported separately so that a customer can see where a bad month came from — but a component breach is credited through the Cluster Availability figure it produces, not through a second, parallel credit.

The facility is direct-liquid-cooled, so the binding environmental commitment sits on the liquid side — coolant supply temperature at the CDU primary inlet. An air-side commitment applies to ancillary and air-cooled equipment in the same hall. Both are stated against published ASHRAE classes so that compliance is verifiable against an external standard rather than an internal one.

The plant is capable of W32 operation at N+1; the service is committed at W27. That gap is deliberate and disclosed — it is the margin absorbed during a chiller failure or an ambient extreme without breaching the customer commitment.

Where measured Cluster Availability for a Measurement Period falls below the commitment, the customer is entitled to a Service Credit expressed as a percentage of the Monthly Service Charge for the affected capacity. Credits are the agreed financial remedy for missed availability and are applied against the following month's invoice.

1. Single tier per period. Only the tier corresponding to the measured Cluster Availability applies. Tiers are not cumulative and are not stacked.
2. Per affected capacity. The credit is calculated on the Monthly Service Charge for the racks or hall that missed the commitment — not on the whole contract.
3. Issued by Prima. Credits are identified in the monthly service report and applied automatically to the next invoice. The customer is not required to claim them.
4. Aggregate cap. Total credits in any Measurement Period, from every head in section 07, will not exceed 50% of the Monthly Service Charge for the affected capacity.
5. No set-off against power. Credits are calculated on the service charge only, and are not netted against pass-through power, taxes or non-recurring charges.
6. Credit, not refund. Credits are applied against amounts otherwise payable. On termination with a credit balance outstanding, the balance is paid in cash within 30 days of final invoice.
7. Survives assignment. Accrued credits follow the SOF on any permitted assignment or novation.

Incidents are classified on service impact, not on cause. Response Time runs from the earlier of NOC detection or customer ticket. Every response target below is met by the resident 24×7 team — there is no dependence on business-hours coverage or on-call callout for P1 and P2.

• Upgrade on customer request. The customer may request an upgrade in priority at any time; the request is honoured for one level unless Prima can demonstrate on the telemetry that impact does not warrant it.
• Downgrade requires agreement. Prima will not downgrade a P1 or P2 without the customer's written agreement or restoration of service.
• Clock does not reset. Reclassification does not restart the Response Time clock; the original detection or ticket timestamp continues to apply.
• Ageing. A P3 unresolved beyond its target restoration is escalated automatically to P2. A P2 unresolved beyond 48 hours is escalated automatically to P1.

Escalation is driven by the clock, not by the customer. A P1 reaches executive level on a fixed schedule whether or not the customer asks for it, and the customer is told at each step.

A P1 unresolved at two hours is on the COO's desk with a written status; at four hours it produces a formal incident notice to the customer's named contacts within the hour; at eight hours it is a CEO-level matter with a written recovery plan.

The facility is designed to Tier III Standard — every element of the power and cooling path can be taken out of service for maintenance without interrupting the load. Planned Maintenance is therefore expected to be non-disruptive, and non-disruptive work does not consume the customer's availability allowance.

1. Preventive maintenance programme. A published annual calendar covering the electrical chain, cooling plant, generators, fire systems and fabric, executed by the resident Facilities team with OEM vendors under supervision.
2. Change approval. Any change touching the critical path requires written approval from the Facilities Manager and the Operations Manager, with a documented rollback plan and rollback trigger recorded before work begins.
3. Customer-affecting change. Notice is issued to the named customer contacts with scope, window, expected impact, rollback trigger and the engineer accountable on the night.
4. Post-change verification. Environmental and availability telemetry is verified against the pre-change baseline before the window is closed; the result is recorded in the monthly report.
5. Failed change. Where a change is rolled back and the window overruns, the overrun is treated as Emergency Maintenance and counted as unavailability.

Physical and information security are operated in-house. This section states the service levels attaching to them — access provisioning times, incident notification obligations and the assurance programme the customer can rely on in its own audits.

The list above is exhaustive. Any cause of unavailability not on it counts against the commitment — including, for the avoidance of doubt: OEM and subcontractor failure; utility loss at the site boundary; fuel supply failure; failure of Prima-owned or Prima-operated equipment; staffing shortfall; and failure of a managed-service provider engaged by Prima.

Service Credits compensate a bad month. They are not a substitute for a service that works. Where availability is persistently below commitment, the customer may terminate the affected SOF without termination charge and without loss of any accrued credit, on thirty (30) days' written notice, if any of the following occurs.

• Remediation plan. On the second consecutive Measurement Period below commitment, Prima issues a written remediation plan within ten business days, with named owners and dated milestones.
• Executive review. The plan is reviewed with the customer at COO level monthly until availability has been at or above commitment for two consecutive Measurement Periods.
• Right preserved. Accepting or participating in a remediation plan does not waive or defer the termination right once a trigger above is met.
• Orderly exit. On termination for a trigger, Prima supports migration of the customer's workload and data for up to ninety (90) days at the then-current rate, with no exit, de-installation or early-termination fee.
• Per SOF. Triggers attach to the affected SOF, not to the Master Services Agreement. A customer terminating one hall retains all rights and pricing under any other SOF in force.

Prima reports against every commitment in this document each month, without a request, by the tenth business day following the Measurement Period. The report is the single document in which measured performance, credits and open actions are reconciled.