How power reaches the chip

Every watt an AI accelerator uses has to get there somehow, and it travels through a surprising number of hands on the way. What leaves a power plant as high-voltage alternating current arrives at a die as a fraction of a volt of steady direct current, delivered through more than a dozen distinct pieces of equipment. Understanding that chain matters for two reasons: it explains why building an AI campus takes years even when the servers themselves ship in months, and it explains why some of the least glamorous parts in the catalog — transformers, switchgear, busbars — are also some of the tightest bottlenecks.

From the grid to the fence line

Power arrives at a campus at transmission voltage, typically hundreds of kilovolts, because moving current at high voltage and low amperage loses less energy over distance. The first job on site is stepping that down. large power transformers (345–765 kv) & substations do this at the substation, alongside the switchyard equipment that connects the campus into the grid. These are custom-built, oil-filled machines wound to order, and they are the single biggest schedule risk for a gigawatt-class campus — lead times measured in years, not months, because only a handful of factories worldwide can build units at this voltage class.

Once power is inside the substation boundary, hv switchgear, gis, breakers, protection relays handles the switching and fault protection at high voltage: breakers that can interrupt a short circuit before it damages the transformers, and relays that decide when to trip. A campus ordering several large transformers at once is also ordering multiples of this switchgear, which pulls on the same constrained supply chain.

Stepping down through the building

From the substation, voltage drops in stages rather than one jump, because each stage of equipment is only rated for a certain voltage window. Medium-voltage switchgear, sometimes packaged into pre-built E-houses or skids, distributes power across the campus. MV/LV transformers then step that down again to the low voltage a building actually uses, and LV switchgear takes it from there into switchboards and panelboards. Every new data hall needs its own set of these smaller transformers and switchboards, so a multi-hall campus multiplies unit counts even as each individual unit gets easier to build than the substation-class gear upstream.

Somewhere in this run, the power usually passes through a ups (static, modular, rotary/drups) & ups power semis, which sits between the utility feed and the load so that a grid outage doesn’t interrupt the racks before backup generation can start. energy storage: ups batteries & bess back the UPS with batteries sized for that ride-through window. AI training workloads add a wrinkle here: GPU clusters swing their power draw sharply and quickly as jobs start, pause, checkpoint, or resume, and UPS systems are increasingly asked to smooth those swings for grid stability, not just bridge outages.

Into the building and down to the rack

Inside the building, power moves along busway / busbar trunking — rigid, tap-off-equipped busbar run overhead or along a hall’s perimeter — because running individual cables to every row doesn’t scale at these densities and busway is easier to reconfigure as layouts change between generations. From there it reaches floor-mounted PDUs and remote power panels, which route it toward the UPS, generator, or utility source as needed, then down to the rack pdus that meters and distributes power to an individual rack.

Inside the rack, the last stretch is short but carries enormous current. A fully populated GPU rack needs to move more current internally than a bundle of discrete cables handles cleanly, which is why rack-scale systems increasingly use a solid in-rack busbar & power whips — Nvidia’s own spec for its rack-scale liquid-cooled systems runs ±48 V at 1,400 A — rather than conventional whips. That busbar feeds power shelves & psus, the rack-mounted units that convert facility AC or DC input into the regulated rail distributed to each tray. Because GPU trays draw far more continuous power than general-purpose compute, these shelves carry substantially more capacity per rack than a conventional server power supply, and wide-bandgap semiconductors (gallium nitride and silicon carbide) are increasingly used inside them in place of plain silicon.

The last few centimeters

At the board, board-level power: vrms, power stages, pol, power semis take that regulated rail and step it down further, to the low voltage and very high current a GPU die actually consumes — often multiple separate voltage rails per chip, each tightly regulated because a die that experiences a voltage droop under a sudden load spike can throw errors or shut down. GPU power draw is both higher and more transient than CPU power draw, which is why VRM design has pushed toward faster response and higher current density than board-level power delivery historically needed. From the VRM, power crosses the package to the die itself, where it turns into computation — and, by simple physics, heat: whatever electrical power goes in comes back out as heat that has to be removed (see “How heat leaves the building”).

Where 800 V DC changes the picture

Every step described above involves converting or re-transforming power at least once, and each conversion step is a place where a small fraction of the power turns into waste heat instead of useful current. As rack power draw climbs, converting AC to DC and back multiple times between the grid and the chip becomes a meaningful efficiency and space cost rather than a rounding error. 800 v dc architecture & solid-state transformers is the industry’s response: an emerging architecture that distributes DC power deeper into the building, in place of the AC busbar and power-shelf approach current racks use, paired with solid-state transformers that could collapse several conversion stages into one solid-state device. It isn’t yet how today’s racks are built, but it’s the direction equipment vendors across switchgear, power shelves, and power semiconductors are already designing toward for the generations after the current one.

Where to go next

• large power transformers (345–765 kv) & substations — the transformers that set the schedule for a whole campus.
• power shelves & psus — where facility power becomes rack power.
• board-level power: vrms, power stages, pol, power semis — the last regulation step before the die.