Commissioning: proving the unit before load

The point of the stage

What commissioning actually proves

Installation proves that equipment is present and connected. Commissioning proves that it works — together, under load, and when something fails. The distinction is the whole point of the stage. Every subsystem can pass its own vendor checkout and the integrated machine can still fail on the interfaces between them: a transfer that works electrically but drops the cooling pumps long enough to trip a thermal limit, an alarm that reaches the controls system but never reaches a human, a fiber pair that lights but fails its loss budget at temperature.

The operational record justifies the rigor. In the Uptime Institute’s 2025 survey of more than 800 operators, about half reported an outage with meaningful impact in the previous three years^[11] — and the failure modes that commissioning exercises deliberately, chiefly power-path and cooling-continuity events, are the ones that recur. Commissioning is the last point at which those exercises are cheap: a failure discovered during integrated testing costs a schedule day, and the same failure discovered in service costs an outage.

Electrical commissioning starts static and ends dynamic. The static half verifies what cannot be measured once the system is live: grounding and bonding continuity, insulation resistance, conductor terminations and torque values, phase rotation, and protective-device settings against the coordination study. The dynamic half energizes in stages and then loads the unit — normally with load banks, so the test is repeatable and no production hardware is at risk — while measuring what the design promised. Reliability analysis of critical-facility distribution follows established practice such as the IEEE 3006 series.^[2]

Two measurement standards matter at this point because AI loads are electrically unusual. Power quality should be recorded with defined methods and aggregation intervals rather than a handheld spot reading — IEC 61000-4-30 specifies those methods^[3] — and harmonic distortion from large rectifier loads is assessed against the limits IEEE 519 sets at the point of common coupling.^[4] Step loading matters as much as steady state: GPU clusters swing power hard between idle and full synchronous training load, so the test script should ramp and dump load, not only hold it. How medium-voltage input becomes rack power is covered in the power-architecture explainer.

A liquid-cooled unit adds tests an air-cooled room never needed. The loop is flushed and filled to a documented chemistry, pressure-tested and held, then flow-balanced across every branch — federal-lab guidance treats fill, flush, and flow verification as core practice for direct-to-chip loops.^[6]Leak detection is proven by simulating a leak, not by confirming the sensor is installed. Supply temperature control is verified against the environmental envelope the IT equipment is specified for; ASHRAE’s thermal guidelines define the classes that verification is written against.^[5]

The load test then answers the only question that matters: at design ambient and full electrical load, does the unit hold its thermal setpoints with margin, and how fast does it drift when heat rejection is interrupted? That ride-through number — minutes to a thermal trip with rejection lost — belongs in the handover package, because operations will plan maintenance windows around it. The loop architecture being verified is described in direct-to-chip liquid cooling.

Network commissioning is mostly certification and documentation. Structured cabling, pathways, and cross-connects are built and tested against the data-center cabling standard^[7]; every fiber pair is certified for loss against the reach class of the optics it will carry; the demarcation to the carrier or campus fabric is proven end to end rather than assumed. The internal fabric choices this verifies are covered in networking and fiber.

Controls acceptance is where commissioning most often gets shortened and should not be. Every monitored point is proven end to end — sensor to controller to dashboard to alarm destination — by forcing the condition and confirming the notification arrives at a person. Out-of-band machine telemetry is verified through a vendor-neutral interface such as Redfish so the operator is not locked into one vendor’s agent^[8]; setpoints, sequences of operation, and trending intervals are recorded as-commissioned. Life safety is tested to its own codes: detection and notification devices are acceptance-tested and documented under NFPA 72^[9], emergency power off is proven to actually de-energize what it claims, and physical access controls, monitoring, and visitor records are verified against the physical and environmental protection controls in NIST SP 800-53.^[10] See monitoring and controls and safety and security for the systems themselves.

Level CX-05 is the difference between a commissioned unit and a checked-out one. With the unit under representative load and every vendor present, faults are introduced on purpose and the response is timed and recorded. A minimum script covers: loss of the primary power path; loss of a cooling pump and then loss of heat rejection entirely; loss of the controls network; loss of the upstream data link; and an emergency power off followed by a full restart to a serving state. Each exercise produces a measured number — transfer time, temperature slope, restart duration — not a checkbox.

These numbers become the operational baseline. Without them, the first real fault is also the first measurement, and nobody can tell degraded behavior from normal behavior.

Commissioning is not accepted by one signature. Five parties typically sign different scopes, and confusion about which signature closes which scope is a common cause of handover delay. Off-site construction adds a wrinkle worth planning for: the ICC/MBI off-site construction standard sets out in-plant inspection, third-party inspection, and on-site final inspection roles precisely because work completed in a factory is not inspected the way field-built work is.^[1]

In the product

How commissioning works for a factory-built unit

The modular model changes where commissioning happens, not whether it happens. Each PODOS Pod is designed as a standardized 1 MW building block and designed for 128 GPUs, so levels CX-02 and much of CX-03 are retired inside the factory against a repeatable test procedure, on the same configuration every time. What remains on site is the interface set: the service connection and protection settings, the grounding electrode system, heat rejection at the real ambient, the carrier handoff, and the integrated failure-mode script that only the assembled site can prove. That front-loading is one reason PODOS targets a 90-day window from order to commissioning for a standard unit — a target, not a guarantee, and the site-side items above still own their calendar.

The site-side prerequisites that make this package achievable on schedule are listed in the data center readiness checklist.

Commissioning is stage five of six in the deployment sequence. It begins where transport and placement ends and closes when the handover package transfers the unit into operations and maintenance. How the factory-versus-field split changes the inspection path is discussed in factory-built vs site-built, and unfamiliar terms are defined in the AI infrastructure glossary.