Electrical & Energy Infrastructure

Every conversion stage between the grid and the chip taxes efficiency, capital, and floor space, so topology design means choosing the fewest stages and highest voltages code and ecosystem allow.

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

A transmission-connected campus must settle who owns the substation, who operates the switches, and whether it registers as a NERC entity — choices locked into a thirty-year interconnection agreement.

The modern accelerator PSU is an active rectifier drawing near-sinusoidal current, so the AI hall's real power-quality threat is the synchronized idle-to-full load step of thousands of clean rectifiers moving as one.

A rack swinging from idle to 150 kW and back in milliseconds turns the UPS into a transient absorber, judged by how fast and flat it clamps the spike along the chip→BBU→BESS spine.

In the last thirty meters of copper between floor PDU and chip, resistive (I²R) loss quietly sets how much power you can actually land on a 600 kW rack.

Past ~200 kW the AC chain hits a wall, and the choice is which DC bus — ±400 V to ride the EV supply chain, or 800 V straight to the rail.

A behind-the-meter plant must take a gigawatt load stepping from idle to peak in milliseconds without tripping — decided by how you parallel prime movers, where storage sits, and whether inverters form the grid.

The fuel-supply chain — pipeline tie-in, conditioning, compression, and a firm-or-interruptible contract — is a second lead-time gate behind the turbines; order the prime movers without locking the gas and they arrive to sit idle.

NERC's 2026 ride-through push makes a gigawatt AI load that trips itself during a routine grid fault a planning contingency at interconnection — a hard design expectation, though not yet a penalty-backed standard.

On a gigawatt of millisecond-stepping GPU load fed by an ungrounded 800 VDC bus, the earthing regime is a safety hazard, a goodput killer, and a poured-in-concrete decision at once.

At sub-cycle resolution, metering and power-quality telemetry are the closed loop that makes a gigawatt of synchronized GPUs a controllable load — without it you cannot bill correctly or stay on the grid.