Electrical Power Acceptance (L3/L4)

By the time a cluster reaches electrical acceptance, the power chain has been designed, studied, manufactured, installed and point-to-point checked. Every component has a datasheet saying it works. Electrical acceptance, Levels 3 and 4 in the commissioning taxonomy of Chapter 13.1, converts that paper into instrumented results. L3 is standalone (pre-functional): each subsystem is energized and run on its own. The switchgear racks and trips, the generator starts and assumes load, and the UPS transfers and rides through, each as its sequence of operations says. L4 is functional performance: the electrical subsystems run together under load, proving that a utility loss, a generator start, a transfer and a UPS ride-through chain into one another without a gap the IT load can feel. Cross-discipline integration waits for the L5 IST (Chapter 13.6). The question here is what "under load" must mean for an AI facility, because the load this building was designed for sits outside anything a conventional data-center commissioning script was written to test.

At every gate there is a choice between a cheaper, faster acceptance that proves the steady-state case and a slower, more expensive one that proves the dynamic case. The cheap path sends its bill months later. At the first real workload, a synchronized load step that the load banks never reproduced trips an under-voltage relay, sags a UPS or de-syncs a generator paralleling scheme, and the operator learns that the L4 sign-off certified a building that cannot hold its design load. The sections below walk the chain in energization order (utility, switchgear, generators, UPS/BESS), then take the three decisions specific to AI: load-bank realism, redundancy-topology validation under load, and protection/arc-flash verification against the study. They end on the criterion conventional commissioning has no vocabulary for: dynamic-load-swing tolerance.

Electrical acceptance follows the direction power flows, and each stage is a prerequisite for the next. You cannot commission the UPS until the switchgear that feeds it is accepted, and you cannot commission the switchgear until the utility or generators can energize it. A deficiency anywhere upstream stalls everything below it.

Utility energization (backfeed / first energization). The interconnect and on-site substation (Chapter 4.2) are energized for the first time, often by backfeeding the customer transformer from the utility side, with the utility's protection and the customer's relays both live. This is the highest-energy moment in the build: the first time full fault current is available behind the customer's breakers. Protection dominates acceptance here. Relay settings are loaded and verified, CT/PT ratios and polarity confirmed, and the trip scheme proven before the bus is left energized. Get the polarity of a differential CT wrong and the first through-fault, not the commissioning team, finds it.

Switchgear commissioning. Medium- and low-voltage switchgear and the busway/PDU distribution (Chapter 4.6) are tested to ANSI/NETA ATS, the acceptance-testing specification the industry treats as the as-installed bar. ATS covers the equipment. Its companion, ANSI/NETA ECS-2024, covers system commissioning, with dedicated UPS and transfer-switch system sections, and is used with the current ATS. ATS-2025 also added a battery-energy-storage section (7.28), the test basis for the BESS layer below. Insulation resistance and the equipment/OEM-approved dielectric test establish the tested insulation condition after shipping and installation; DC hi-pot is not a universal choice for every cable or electronic assembly. Contact-resistance (ductor) testing on every bolted joint catches the loose lug that becomes a hot spot. Breaker timing and trip verification prove the mechanism operates within spec. And primary injection proves the protective relays trip on real current, not just on a settings file. With thousands of low-voltage breakers, even a small per-device failure rate produces a long deficiency list. The punch-list management of Chapter 13.2 is the only way to track which of 3,000 breakers has been proven and which has not.

Generator commissioning. Each generating set is started, run and load-tested, first on its own load bank (L3), then in parallel and onto the building (L4). Acceptance proves cold-start time to accept load, voltage and frequency recovery on a block load step, fuel-system and exhaust performance, and protection (reverse-power, loss-of-field, over/under-frequency). AI sites that lean on on-site generation as prime or island power need paralleling and load-sharing acceptance deep enough for its own treatment in Chapter 13.4. Here the gate is simpler: the gensets can pick up the design block load within the voltage and frequency window the downstream UPS can tolerate.

UPS and BESS commissioning. The ride-through layer (Chapter 4.5) is the last electrical subsystem and the one the IT load is most intimate with. The acceptance set is the battery/capacitor discharge test to proven autonomy; the transfer tests (normal-to-bypass, bypass-to-normal, and the all-important loss-of-utility transfer that hands off to the generator); harmonic and power-factor behavior under non-linear load; and, for the increasingly common BESS layer, the fast-response transient absorption that conventional double-conversion UPS was never asked to provide. Here the AI-specific criteria start to bite, because the UPS/BESS must absorb the GPU load swing that the gensets and the grid cannot follow fast enough.

A load bank puts load on the power chain before the GPUs arrive, or before you dare put them at risk. Choosing the technology is choosing which question you are answering, and the answers overlap: impedance, switching capability and heat-rejection path are separate specifications.

Resistive load banks dissipate real power (kW) as heat through resistive elements at unity power factor. They are the workhorse of data-center commissioning: cheap, robust, and enough to prove that the transformers, switchgear, busway and cooling can carry the design kW and reject the design heat. They draw no VARs, so they say nothing about how the chain behaves under the leading/lagging and harmonic-rich conditions a real GPU power supply imposes.

Reactive load banks add inductive (and sometimes capacitive) load, so the team can test the chain at a realistic, non-unity power factor, at the specified lagging or leading PF. Generator sets are rated in kW at a stated power factor (typically 0.8 lagging) and in kVA. For an illustrative 0.8 PF load, 100% rated kW needs 1/0.8 = 1.25 times that kW in kVA, so compare that apparent load to the actual generator rating. Many modern UPS are rated at unity PF (kVA = kW), so check each nameplate. Reactive banks prove the generator's automatic voltage regulator (AVR) and kVA rating, the transformer's heating at rated PF, and the UPS's behavior at the leading or lagging PF its loads actually present. Whether a resistive or reactive bank also proves the transient depends on its switching granularity, control bandwidth and slew rate. Conventional banks do perform load-step tests, so specify the step size, rate and repetition you need rather than letting the technology class settle it.

AI-load-emulating (dynamic) load banks are the newest and least standardized class. They reproduce the workload's electrical fingerprint with programmable, fast load steps that mimic the synchronized ramp-up, steady-state oscillation and abrupt ramp-down of a GPU cluster entering and leaving collective operations. They exercise the di/dt the power chain was designed to survive only where their measured step size, slew and repetition cover it. Switched resistive and reactive banks also prove transients inside their switching range: the under-voltage and frequency excursions on a step load, the UPS/BESS transient-absorption response, the generator's load-acceptance and load-rejection recovery, and the interaction between rack-level capacitance and facility-level storage. Dynamic banks are more expensive and harder to source, and they do not perfectly reproduce the target cluster. Use them to prove the transient range you specified, then cover the normal-operation behavior they cannot reproduce with a staged production-representative run in Chapter 13.9. Keep hazardous protective-function tests in injection/HIL or an isolated controlled scope.

The practical pattern at a 2026 AI site is layered, not single-choice. Resistive banks prove capacity and that the cooling can reject the heat. (Most facility scopes specify resistive banks that reject to air, which is itself a realism limit. Liquid-cooled banks exist and are the tool for the loop-side proof, treated in Chapter 13.5.) Reactive banks then prove two things: the generator at its kVA rating, and the UPS at whatever power factor its real loads present. Load banks capable of the required steps then prove the transient range. Any remaining duty goes explicitly to a valid test at IST (Chapter 13.6) and ultimately to the first proxy training run (Chapter 13.9), which loads the chain with the real, software-synchronized cluster after its protective functions pass. The decision a strategist signs off is how much transient risk to retire at L4 (expensive, but caught before GPUs are at risk) versus carry into IST and first workload (cheaper to commission, but found with $100M+ of accelerators energized).

A redundancy topology (N+1, 2N, distributed-redundant, or the fault-domain designs cataloged in Chapter 12.1) is a claim that the load survives the loss of any one component, or any one block of components. The drawing makes the claim. L4 acceptance makes it true by proving isolation and transfer behavior with secondary injection, simulation, test sources and staged load banks, before any specifically approved live-load demonstration, and by watching whether the load notices. The scoping question is how aggressively you test: which transfers and source losses you genuinely trip, and which you accept on the single-line diagram's authority.

The acceptance set runs in ascending order of how much it tells you and how much it risks. (1) Controlled transfers: operate every static transfer switch, automatic transfer switch and breaker tie through its full sequence under load, confirming break/make timing and that no downstream bus drops. (2) Single-source loss: simulate or open an approved test source at a defined boundary, a UPS module, a CDU power feed or a PDU, and confirm the redundant path picks up within the IT load's ride-through window. (3) Approved single-fault and specifically justified combined-fault cases: the IST-grade test (Chapter 13.6) where you stack faults to confirm the topology survives the design-basis worst case, up to the whole-system utility-loss proof selected by the approved risk-based test plan. For an AI cluster, workload recovery changes the consequence of a failed test state but does not erase the state from the acceptance basis. Checkpoint-and-resume after a node failure does not prove that a distribution-path or whole-hall interruption meets the uptime target. L4 must prove the topology's own states: transfer interruption, post-event loading, path and control independence, common modes, and recovery. Workload and fleet resilience are proven separately, then exercised end to end.

The protection coordination study and the arc-flash incident-energy study (produced in Chapter 4.11) are analytical artifacts. They specify relay settings, breaker trip curves and PPE/labeling on the assumption that the as-built plant matches the model, and electrical acceptance is where that assumption is tested. Verification depth is the real choice. You can confirm coordination by reading settings back from the relays, which is fast and cheap and proves only that someone typed the right numbers. Or you can prove it by primary injection and recorded event data, which is slow and expensive and proves the device actually trips at the studied current in the studied time.

The acceptance set is concrete. Settings verification: every protective relay's pickup, time-dial and curve is read back and reconciled against the coordination study and against the device labeling and addressing, because a correct setting loaded into the wrong device silently breaks selective coordination. Primary/secondary injection: current is injected to confirm the relay operates within the study's time-current band, so that a downstream fault is cleared by the nearest upstream device and not by one two levels up that would needlessly drop a whole bus. Coordination proof under fault: controlled functional demonstrations, with injected or simulated fault current at the study's coordination points and relay event records showing which device operated first, close the gap between a correct settings file and selective behaviour. Creating a real high-energy fault in the installed plant is not the required final proof. The proof package is the as-built study inputs, the settings and wiring verification, the device and chain injection tests, and those demonstrations, each stating which uncertainty it closes and which it leaves open. Arc-flash verification: the as-built clearing times from those injection tests feed back into the incident-energy calculation, because arc-flash energy scales with clearing time. A relay that trips slower than the study assumed raises the incident energy at that bus, which can invalidate the PPE category on the label a technician is relying on. Acceptance is therefore more than "the relay trips"; it is "the relay trips fast enough that the arc-flash labels on this gear are still true."

Conventional data-center commissioning has no script for this criterion, which also makes it the one most likely to be skipped: the load banks most scopes specify cannot produce it, and the relays that would trip on it sit upstream of the IT the script focuses on. A modern rack actively fights its own transients. The GB300 NVL72 power shelf carries roughly 65 J/GPU of electrolytic capacitance, about half the PSU volume. It combines that with power-capping on ramp-up and a deliberate "GPU burn" on ramp-down to taper the load gracefully, cutting peak grid demand by up to 30% in NVIDIA's instrumented GB200/Megatron demonstration of those GB300 power-shelf features. The Vera Rubin generation raises rack storage to ~400 J/GPU (≈6x) and adds closed-loop state-of-charge control; NVIDIA puts the resulting drop in peak current demand at up to 25% (March 2026). That 25% is NVIDIA's design statement with no stated baseline, not a measurement like the 30% demonstration, so the two figures do not rank the generations. Behind the rack, facility BESS adds the next layer of fast-response absorption (Chapter 4.5). Rack capacitance, then BBU/UPS, then facility BESS form a layered mitigation stack, and electrical acceptance is where you prove the stack actually engages and that the residual swing it does not absorb stays inside the power chain's tolerance.

The acceptance criteria are quantitative pass/fail gates in the L4 script (script anatomy is in Chapter 13.2). The voltage excursion at the bus on a defined load step stays within band. The frequency excursion (on generator or island power) stays within the relay's no-trip window. The UPS/BESS sources the swing without dropping to bypass or sagging the output. The rack-level smoothing holds the contracted upstream peak and slew under the specified test waveform and initial storage state. No protective relay anywhere in the chain trips on the largest synchronized step the cluster can produce.

The grid has already shown what transfer settings decide at scale. About 1,500 MW in Northern Virginia in 2024, and about 3.8 GW near Ashburn on 22 July 2026 after one normally cleared fault, left the grid as sites transferred to backup as designed. At the utility interface, prove two separate things with settings readback and secondary injection or HIL: that UPS transfer and PSU undervoltage settings keep the computing load connected through a normally cleared sag inside the site's grid ride-through envelope (set by ERCOT NOGRR282, SPP's rules or another interconnection rule, or by the industry OCP grid-resilience specification), and that they transfer cleanly for disturbances outside it (deeper or longer sags, or loss of source). The integrated test belongs to IST (Chapter 13.6); Chapter 4.10 covers the rules and their status.

A load bank, even a dynamic one, can approximate these steps but cannot reproduce the exact, software-driven, thousands-of-GPUs-in-lockstep spectrum of a real workload. So this gate spans complementary test results. Dynamic load banks at L4 retire the bounded gross-transient case. IST combines only the approved interactions inside an isolated or otherwise protected test boundary. Product-representative emulation or a staged workload validates the remaining electrical and thermal behavior, with live high-energy testing exceptional and governed by explicit authority, hold points and abort criteria (Chapter 13.6; Chapter 13.9). Where you draw that line is the call: draw it too early and the cluster's first real all-reduce is also its first real transient test.

E-01 trace. The electrical lead freezes the one-line boundary, protection/settings version, load-bank waveform and storage state. The test authority verifies isolation, backfeed prevention, abort and restore paths. Simulate loss of the upstream test source at t = 0 while commanding the load step. The UPS delivers (1.00 MW × 2.0 s) + (0.80 MW × 10.0 s) = 10 MJ, below the assumed 30 MJ usable output reserve. Apply uncertainty to the captured extrema: 0.925 − 0.005 = 0.920 pu ≥0.90 pu; 59.6 − 0.1 = 59.5 Hz and 60.3 + 0.1 = 60.4 Hz stay inside the frequency band; 12.0 + 0.1 = 12.1 s ≤15.0 s. Simulated bypass and unexpected-trip counts are zero.

Result and action. The simulated profile passes the gates above. Installed release waits until attributable raw waveforms, relay/breaker events, storage telemetry, calibration and witness records exist. Stop on a boundary violation, unexpected trip or excursion; restore the approved source state, remove the step, recharge the UPS and repeat the baseline before further tests. In a real record the electrical lead, the CxA and the owner's release authority sign it. NETA's acceptance standard sets the equipment tests but no voltage threshold for this case, so the limits come from the OEM and the sequence of operations, and the evidence follows Chapter 13.2.

Flip. With ±0.005 pu uncertainty, the indicated minimum must be at least 0.905 pu. A simulated 0.904 pu gives 0.899 pu and fails even if transfer completes at 12.0 s. Fix the transfer response and repeat the dependent tests. Carrying that miss into a GPU run buys an earlier start at the cost of an unproven ride-through boundary. Chapter 13.4 adds the separate island reference and reconnection record.

For an AI facility, electrical acceptance is a gate the density ramp re-opens with every GPU generation, not a one-time event. A power chain accepted for GB300 NVL72 at the peak or busway basis of the OEM configuration (Lenovo's 155 kW peak; NVIDIA's reference architecture says up to 142 kW) faces a different acceptance problem for VR200 NVL72, shipping since June 2026, at NVIDIA's 330 kW cabinet TDP facility design basis and Pegatron's 228 kW Max P. Rubin Ultra/Kyber adds an announced ~600 kW roadmap planning point on 800 VDC (NVIDIA targets H2 2027; SemiAnalysis reported 2028 in July 2026). GB200 is the installed base the substrate was originally accepted against, not the generation setting today's gate. Every density change needs a new normal-load and transient-load acceptance profile. Recalculate short-circuit duty, protection coordination, relay-injection scope and arc-flash energy only when the source/network model, equipment, settings or operating configuration changes.

The irreversible electrical substrate (interconnect capacity, voltage class, switchgear fault rating, busway ampacity) has to be accepted not just for today's load but for the ramp it must absorb (the density-ramp trap of Chapter 12.1). The strategist's acceptance decision is therefore forward-looking: do you commission the protection and the transient envelope against the generation you are installing, or against the headroom the substrate must eventually carry? Accepting only to today's load is cheaper and re-opens the gate at every refresh. Accepting to the substrate's ceiling costs more now and retires that capacity test early, but a new rack generation still needs its own transient and interface acceptance.