Electrical & Energy Infrastructure

Every conversion stage between the grid and the chip costs efficiency, capital and floor space. Choose voltages and stages whose losses, isolation, protection and serviceability work together in the failure state as well as at the full-load efficiency point.

The switchgear topology, insulating gas and protection scheme you freeze in the on-site substation decide whether one fault drops a pod, a hall or the whole gigawatt factory.

A transmission-connected campus has to settle who owns the substation and who operates the switches under its tariff and interconnection agreement, then work out separately whether it registers as a NERC entity. Each choice shapes long-lived assets, and reversing it later costs real money.

Harmonics matter less when the accelerator PSUs are active rectifiers whose current stays near-sinusoidal at the load point you specify. But thousands stepping from idle to full load together can still sag the bus. Test harmonic heating and that step separately before buying transformer capacity.

UPS and storage now do two jobs: carry the load when the utility disappears, and absorb the synchronized GPU load step every few seconds. Size and prove each duty separately at the chip, rack and facility layers, so smoothing never spends the outage reserve.

Between the floor PDU and the rack sit the last thirty meters of copper. For AC-fed racks up to Vera Rubin's 330 kW design basis, I²R loss, connector temperature and surviving-feed duty decide how much power that copper can land.

As rack current pushes copper and connectors to their limits, the choice is 415/480 VAC or an 800 V-class DC interface. ±400 V counts each pole from a midpoint; 800 V counts the full span. Fault protection and service design decide whether either works.

A synchronized GPU load step arrives in milliseconds and tests every source state on a gigawatt campus. Storage placement, prime-mover load sharing and the voltage and frequency reference decide whether IT and cooling survive grid-parallel, islanded and transition operation.

Fuel supply is a second lead-time gate behind the turbines: the pipeline tie-in, conditioning, compression and a firm-or-interruptible contract all run on other parties' schedules. Order prime movers without securing the gas and they arrive to sit idle.

A gigawatt that drops out during one normally cleared fault becomes a planning contingency. In October 2026 NERC’s computational-load standards passed final ballot without a ride-through curve, so regional rules and your connection agreement decide the ride-through, P/Q, frequency and recovery behavior the utility models.

On a gigawatt-class GPU campus, the AC and 800 V-class DC earthing can decide whether a fault is cleared locally or trips running compute, costing goodput. Grounding is poured early: set the reference first, then prove touch limits, first-fault response and safe access for service.

A gigawatt of synchronized GPU load needs sub-cycle power-quality capture to explain throttling and trips, reconciled meters to bill correctly, and protection that holds its own trip authority. Handover must show that load, storage and recovery meet the design basis.