Metering, Power Quality, Monitoring & Electrical Operations

Earlier chapters built the electrical machine: the substation (Chapter 4.2), the transformers and harmonic mitigation (Chapter 4.4), the UPS and storage spine that absorbs transients (Chapter 4.5), DC distribution (Chapter 4.7), and the grid obligations that keep a gigawatt load online through a fault (Chapter 4.10). This chapter is about seeing and operating that machine. Metering shows where the power went. Power-quality instruments show whether the waveform is healthy. Telemetry-driven control loops make GPU load steps survivable. Commissioning and day-2 electrical operations keep the whole thing safe and inside spec. Without that visibility, you cannot confirm that any other Part 4 decision holds in production.

Every fork here has a downstream cost: how deep to meter, how fast to sample, open- or closed-loop smoothing, converged or federated DCIM/EPMS, and the commissioning level you hand over against. Underneath all of them is one break. AI loads broke the assumptions behind classic data-center metering and power-quality practice. A legacy hall drew a smooth, diversified, slowly varying load that a monthly revenue meter and an annual power-quality audit described well enough. A training cluster swings a power-electronic, sub-second-synchronized load by tens of megawatts in milliseconds, and instruments built for the old world are blind to exactly the events that matter.

Metering in an AI factory is a hierarchy, and the first decision is how far down it reaches. At the top sits the revenue meter at the point of interconnection, utility-owned or customer-owned per Chapter 4.3. The tariff is settled against it, and the utility's metering specification picks its standard and accuracy class. IEC 62053-22:2020 classes 0.2S and 0.5S, and ANSI C12.1-2026 classes 0.2 and 0.5, are typical examples. Utilities and system operators may also require operating telemetry, which is a separate requirement from the settlement meter. Below it, you choose how many layers of sub-metering to add: the medium-voltage feeders, the pod or block transformers, the LV switchboards and busway runs, the rack PDUs, and at the deepest tier, branch-circuit and outlet-level metering inside the PDU (Chapter 4.6).

Each layer costs money and bandwidth, and the payoff is attribution. Meter only at the POI and you know your bill and little else. You cannot tell a tenant what they owe, trace a PUE excursion to a pod, see which busway is approaching its ampacity, or reconcile the IT-reported power draw against what the facility delivered. To bill a tenant you need a meter where that tenant's share begins, whether that is a feeder, a branch circuit, a rack or an outlet. A rack power budget can be enforced from the rack's measured aggregate, and outlet metering adds device-level attribution. Attributing energy to a training job also needs a map from workload to hardware and a defined allocation method. PUE needs reconciled facility-input and IT energy at defined boundaries; Chapter 15.1 covers the PUE/TUE boundaries and computational-component accounting. Under-metering leaves a permanent gap in what the business can answer later. A retrofit may need planned isolation and de-energized switchgear work, so reserve safe metering access during design.

A second axis is what each meter measures. A cheap meter reports kWh and maybe RMS current. A power-quality-capable meter reports per-phase voltage and current, real, reactive and apparent power, power factor, individual harmonic magnitudes and THD, and timestamped sag, swell and transient events, which matter most for AI. Specifying PQ-capable meters at the right tiers, typically the MV feeders and the main LV boards, turns a metering hierarchy into a power-quality monitoring network without a second, parallel instrumentation buildout.

The first balance does not prove a meter fault. The residual is 1,000 − 950 − 30 = 20 kWh, or 2.0% of input. Worst-case meter uncertainty is 0.005 × (1,000 + 950 + 30) ≈ 10 kWh; add the ±4 kWh loss-model band and the bound is about 14 kWh. The residual sits 8 kWh from the 12 kWh loss estimate, inside that bound, so record the balance as consistent. It is not a calibration or efficiency certificate.

The decision flips when the gap exceeds the bound. If 15 kWh of IT load goes unmetered or misattributed, so the IT feeders read 935 kWh, the residual becomes 35 kWh. That is 23 kWh above the loss estimate, well outside the same 14 kWh bound. Hold billing allocation and check CT/VT polarity and ratios, omitted feeders, timestamp windows and the loss model before changing tariffs or blaming the load.

The same acceptance test procedure (ATP) checks the controls. In a controlled test environment, inject three conditions one at a time: let the BBU/BESS telemetry go stale, drop the EPMS network, and apply a protection test input. Stale telemetry should block new supervisory smoothing commands, the reachable outage reserve should stay available, and protection should take its independently approved action at its own approved settings. No software alarm substitutes for a trip test. The meter engineer signs off the energy reconciliation, the protection engineer signs the settings and the witnessed clearing, and the controls integrator signs the command and failure-state traces. Commissioning brings the three together in the ATP, and Chapter 14.7 takes over the operating response.

AI halls contain a high share of power-electronic loads. Active-PFC front ends can draw near-sinusoidal current, but aggregate distortion varies with topology, loading, controls and upstream impedance, and the aggregate injects harmonic current back toward the source. IEEE 519-2022 governs this. The 2022 revision made it a full IEEE Standard; the 2014 edition was a recommended practice. It sets limits at the point of common coupling (PCC), not at every individual load. The PCC sits on the public supply, at the closest upstream point where other customers are or could be connected. That may not be your point of interconnection. Internal tenant feeders are not PCCs under IEEE 519, so apply limits there by contract, at evaluation points the contract defines.

The limits depend on the PCC voltage band. Voltage distortion is capped at 8% THD-V at 1 kV and below, and at 5% THD-V with 3% per individual harmonic above 1 kV up to 69 kV. Current distortion (TDD) is bounded on a sliding scale keyed to the short-circuit-to-load ratio I_sc/I_L, from 5% TDD on a weak bus to 20% on a stiff one, for PCCs from 120 V to 69 kV. PCCs above 69 kV fall under the tighter limits in Tables 3 and 4 of the standard, so a transmission-voltage connection cannot use that 5–20% range. → Distortion mitigation, such as active front ends, active filters or a validated multipulse arrangement, is the subject of Chapter 4.4. A K-rated transformer or IEEE C57.110 derating gives thermal withstand and capacity for the measured spectrum. It does not reduce harmonic current or establish IEEE 519 compliance. The decision here is how you measure compliance.

Measurement is where the problem bites. Compliance is conventionally shown as a snapshot: a power-quality engineer parks an analyzer at the PCC for a week and produces a report from aggregated statistics. IEEE 519-2022 assesses both harmonic voltage and current, and the user's emissions are judged against the current limits, not voltage THD alone. It uses daily 99th-percentile 3-second values and weekly 95th- and 99th-percentile 10-minute values, and the weekly 99th-percentile test applies to current, not voltage. Instruments must meet IEC 61000-4-7 and the Class A methods of IEC 61000-4-30:2025 (corrected 2026-07). Whether a given survey shows compliance still depends on the site agreement and the instrument setup. Either way, it does not replace event capture for diagnosing equipment.

A synchronized training step is a millisecond-scale event. The voltage disturbance it causes on a stiff-but-not-infinite bus can begin within a cycle, and how long it lasts depends on the step, the source impedance and the storage and controls that respond. Harmonic statistics alone cannot characterize every sub-cycle event. Compliant instruments also offer shorter aggregation and event functions, depending on the standard and how they are configured. To see the events that actually throttle accelerators and trip undervoltage ride-through, you need continuous waveform capture with sub-cycle resolution and event triggering. A clean harmonic survey plus a throttling log is not enough to diagnose a sag. One correctly specified instrument can do both jobs, or you can use separate devices. The cost fork is continuous multi-point capture, with its storage and integration, against a periodic survey. Let the harmonic survey stand in for event diagnostics and you leave a blind spot.

The physics and mitigation of GPU power transients, meaning the synchronized load steps and the chip-to-BBU-to-BESS absorption spine, belong to Chapter 4.5. This chapter covers the observability and control half: the telemetry that feeds smoothing, and whether smoothing runs open- or closed-loop. That decision determines whether your facility is a passive victim of its own load steps or an actively managed grid citizen. Three tiers are worth separating: fixed supervisory policy, rack-local feedback, and facility coordination.

Open-loop smoothing fixes its behavior at commissioning: firmware ramp-rate limits, idle-time floors and power caps, set once and left alone. NVIDIA's GB300 NVL72 ships the building blocks, roughly 65 J/GPU of capacitive energy storage in the power shelf, a GPU-burn mechanism that bleeds energy on ramp-down, and ramp-rate controls. In NVIDIA's own test (July 2025, an instrumented GB200 rack training Megatron), the new shelf cut peak grid-side power by up to 30%. Configured open-loop, those features run on fixed parameters whatever the grid state. Rack-local feedback sits in between. On Vera Rubin NVL72 the power shelves hold about 400 J/GPU of capacitor storage, some six times the GB300 figure (NVIDIA, March 2026). Vera Rubin's GPUs track the rack capacitors' state of charge themselves, in a closed loop that needs no reconciliation with EPMS or DCIM. Facility closed-loop smoothing is an integration layer above NVIDIA's exposed controls: SMI and Redfish set GPU power parameters, while Mission Control provides cluster power management and building-management integration. Before that path takes on FRT or grid-services duty, the controls specification must pin down command cadence and latency, arbitration between controller and scheduler, RBAC, the loss-of-signal state, observability, and witnessed end-to-end tests.

The cost runs in both directions. Open-loop is simpler and fixes its supervisory policy, but interacting firmware limits, source impedance and recovery can still create dynamic problems. Tuning for the worst case also leaves goodput on the table whenever actual conditions would permit more useful output. Closed-loop recovers that goodput but introduces a real governance problem. When grid stress and a training deadline collide, something decides whether to throttle the GPUs or stress the grid, and that authority must be explicitly assigned to the EPMS, the workload scheduler, or a negotiated handshake between them. Leaving it implicit is how a facility ends up silently throttling revenue workloads, or violating a grid-services commitment it forgot it had made.

Grid obligations can also cap consumption. ERCOT presented its large-load mitigation plans on 17 September 2026, and they were not yet active at that point. Once activated, they use ERCOT-provided group and participant consumption telemetry and allocate caps to each TSP/QSE. A site in an affected group needs a defined metering and control interface to stay under the limit its TSP/QSE assigns it (Chapter 4.10). → The grid-side obligation that constrains this loop is Chapter 4.10; the goodput-vs-availability framing is Chapter 12.2; the demand-response revenue that closed-loop control unlocks is Chapter 15.8.

Two platforms describe the electrical machine, and the decision is how they relate. The EPMS (Electrical Power Monitoring System) is the SCADA-grade supervisory system that watches the power chain in real time: breakers, relays (IEC 61850), meters, ATS/STS state and fault events. It is the system of record for electrical events. Calling it an EPMS does not make it deterministic, give it a safety integrity level or give it protective authority. Tripping lives in the protective relays, the trip circuits and the breaker logic, and the major EPMS vendors say plainly in their own documentation that their software and communications are unsuitable for time-critical protection. Schneider Electric's Power Monitoring Expert 2024 IT guide (October 2025) is typical: it warns against using the software for protection or time-critical control, and tells users to plan for communication delays and failures. Look for the equivalent warning in the documentation of any EPMS you are offered. DCIM (Data Center Infrastructure Management) is the broader operational layer: asset and capacity management, power, cooling and space, environmental sensors, increasingly a digital twin, and the home of cross-domain analytics and predictive maintenance. The two overlap on power telemetry, and blind spots form at the seam between them.

The choice is converged versus federated. A converged platform, one vendor's DCIM ingesting the EPMS natively, gives a single pane and one data model. The cost is lock-in and the risk that real-time electrical supervision becomes entangled with a non-deterministic IT platform. Federated systems keep the EPMS as an independent, hardened, real-time system and define an explicit data contract by which it publishes to DCIM, commonly over a message bus or a historian. Protection stays where it belongs, in the relays, trip circuits and breaker logic, and their required operation must not depend on the historian, the analytics or the supervisory network. Federated is the more defensible posture for anyone who holds that protection must never depend on the analytics layer being up, but it only works if the data contract is real: defined tags, defined latencies, defined failure modes. At an undocumented seam, a meter reads one number, the IT stack reports another, and nobody can say which is right.

The hardest part of the integration is facility-to-IT, not facility-to-facility. The EPMS speaks Modbus, IEC 61850 and BACnet; the IT fleet speaks Redfish, IPMI, SNMP, DCGM/Prometheus, and PMBus down at the power shelf. SoC management across hundreds of rack BBUs is a concrete case. The facility BESS state of charge lives in the EPMS, but each rack BBU's SoC is reported out-of-band via Redfish to the IT management plane. Facility-wide orchestration of BBUs and BESS needs both views reconciled in one place at one cadence, which is the converged-versus-federated decision made concrete. → The IT-side telemetry pipeline (DCGM, Prometheus) and the digital-twin layer are built out in Chapter 14.2; the agentic-ops direction is Chapter 14.13.

Commissioning is where everything above is proven, or where the gap between design and reality shows up, ideally before handover and not during a production training run. The industry runs a five-level (L1-L5) taxonomy: L1 factory acceptance testing of individual components, L2 site acceptance on delivery, L3 pre-functional checks, L4 functional performance testing of each system, and L5 integrated systems testing (IST). IST is the coordinated failure-injection exercise in which utility, generator, UPS, BESS, cooling and IT load run together through loss-of-power and load-step scenarios. The handover decision is which level you accept against. For an AI factory it must be L5, because the failure modes that matter are interactions, not component faults. A loss-of-power test proves a clean transfer to backup. It does not prove that the load stays on the grid through a normally cleared voltage sag. ERCOT's NOGRR282 and SPP's large-load rules now require that of qualifying new loads, so the L5 plan needs both tests. Chapter 13.5 and Chapter 13.6 cover thermal and integrated acceptance.

The defining commissioning fork for AI is the load profile you validate against. Traditional commissioning uses resistive load banks: a steady, linear, well-behaved load that proves the power chain can carry rated current and reject heat. That is necessary but not sufficient. A default resistive test does not reproduce the two things that define an AI load. First, the load is non-linear. A resistive bank injects no harmonics, so it tests nothing about IEEE 519 behavior or the active front ends. Second, the load is stepped. Switched resistive banks can execute coordinated steps across networked units, but only to the rise time, step resolution and repetition rate you wrote into the specification, and never with the constant-power response of a GPU front end. A steady-state resistive test therefore proves nothing about transient ride-through, BBU/BESS response or the closed-loop smoothing logic. Specify rise time, resolution, repetition, power factor and waveform first, then select equipment against them. A cluster that passes a resistive test at rated power can still trip on the first 50-megawatt synchronized training step, because the step was never in the test plan.

Step-load validation fills that gap. Programmable non-linear AC loads, electronic or regenerative, can inject a specified harmonic spectrum, and switched resistive and reactive banks can execute specified load steps. Together they test whether capacitance, BBUs, BESS and the smoothing loop catch the transient as designed, and whether ride-through holds through the induced voltage dip. Reactive banks alone prove displacement power factor only and inject no harmonics. Before you accept a test load bank, require the offered model, firmware, operating modes and connection arrangement to reproduce the acceptance profile, because vendors gate waveform modes on model and firmware. For regenerative loads, identify where the energy returns and measure net power through the feeder or source under evaluation, since the returned energy changes the net test load upstream. Rack capacitance and GPU-side smoothing are tested only if those components are in the test path. An external AC load bank bypasses them and cannot prove them. The cost of skipping step-load validation is the worst kind: the facility passes commissioning, the SLA clock starts, the first real workload lands, and the transient behavior nobody tested takes the cluster down. That is now a production incident with a customer attached, not a punch-list item.

The protection studies behind a safe, selectively coordinated electrical system are design-phase deliverables, canonically in Chapter 4.2: short-circuit and fault-duty analysis, time-current-curve coordination (the active IEEE 3004 series; IEEE 242 has been inactive since 2021) and arc-flash incident-energy calculation (IEEE 1584 / NFPA 70E). This chapter owns the acceptance handoff: transfer approved studies, settings and test records to Chapter 14.12, which owns change control and the procedures for day-2 electrical work. Selective coordination is proven only for the equipment, settings and source states in the approved study. Before changing a breaker setting, adding a transformer or shifting a source configuration, check the proposed state against that study. If it lies outside what the study analyzed, update the coordination, fault-duty and arc-flash analyses and the operating records before using it. A change can preserve or improve coordination. The risk is operating outside the analyzed range. The discipline is a maintained coordination study and an as-operated single-line that the EPMS reflects in real time, so a fault clears at the nearest device and takes down one pod, not the whole hall.

Arc-flash operations follow the same pattern. The IEEE 1584 study yields predicted incident energy at the working distance and the arc-flash boundary for equipment within its scope. NFPA 70E then requires PPE selection by either the incident-energy method or the PPE-category method, as applicable, not both. Labels must track the as-operated configuration, and energized work must follow the NFPA 70E hierarchy of controls. For an AI factory at scale, that increasingly means designing for de-energized maintenance, with redundant paths that let you isolate and lock out a section without dropping load, rather than relying on PPE for live work.

The new operational hazard is DC touch-safety. The shift to ±400 V and 800 VDC distribution (Chapter 4.7) brings a class of risk that AC-trained electrical operations teams have not internalized. DC fault current has no periodic zeros for a breaker to exploit, so DC arc-flash and sustained-arc behavior differ fundamentally from AC. The system may be solid-midpoint, HRG or isolated, a first ground fault can itself be hazardous, and an opposite-pole second fault can create a high-energy short (canonical DC grounding and isolation monitoring is in Chapter 4.11). Solid-state DC breakers behave nothing like the electromechanical devices technicians know. The metering and monitoring network must therefore surface DC isolation status and first-fault alarms as first-class operational signals, and the safety program must retrain for DC. A real risk is not a design error but an experienced AC technician applying AC-correct instincts to a DC bus.