Lesson 02/12Beginner15 min read·5 diagrams

Power Infrastructure

From the utility substation to the chip socket, every watt that powers an AI training run passes through six levels of equipment. This lesson walks each one, then explains why GW-scale AI is now competing with cities for grid capacity.

1 · The path from grid to chip

Electricity entering an AI data center is transformed and conditioned six times before reaching a GPU. Each stage exists for a reason — fault isolation, voltage matching, instantaneous backup, or final distribution.

Six stages between the utility line and the GPU socket. Generators are a parallel safety path, not in series.

Utility grid — high-voltage transmission (115 kV / 230 kV typical) from the local utility.
Substation — steps voltage down to medium-voltage (typically 13.8 kV or 34.5 kV) and provides the legal handoff between utility and customer.
Switchgear — high-current breakers and protective relays. Isolates faults before they propagate.
Transformer — final step-down to 480 V (US) or 400 V (EU) for distribution inside the building.
UPS (Uninterruptible Power Supply) — battery or flywheel buffer that bridges the <10-second gap before generators come online.
PDU (Power Distribution Unit) — rack-level branch circuits, often 415 V three-phase, feeding individual server power supplies.

2 · UPS + generators: why both

UPS and generators solve different problems. UPS handles millisecond-to-second outages — the lights blink, generators haven't started yet. Generators handle seconds-to-days — the grid is properly down.

UPS response

<10 ms

Battery or flywheel

Generator start

~10 sec

Cold start to rated load

Generator runtime

48–72 h

On-site fuel typical

UPS efficiency

94-98%

Modern lithium-ion online

3 · Why power density matters

Power density (kW per rack) is the single most important number in modern data center design. Every other system — cooling, networking, raised floors, even the room dimensions — is sized to it.

Per-rack power has 30×'d in 30 years. Most of the jump came in the last 5.

A 100 MW campus housing GB200 NVL72 racks holds roughly 800 racks. The same 100 MW running 7 kW traditional racks would hold 14,000 racks spread across 4–5× the floor space — but accomplish a tiny fraction of the AI work.

4 · The grid problem

Modern AI buildouts are bumping into hard limits: utilities can't deliver power fast enough. Northern Virginia, Dublin, Singapore, and Frankfurt all have multi-year interconnect queues. Hyperscalers are now doing things that would have been unthinkable in 2020:

Co-locating with nuclear (Microsoft–Three Mile Island restart, AWS–Talen Susquehanna)
Long-term geothermal PPAs (Google–Fervo Energy, Nevada)
On-site gas turbines (xAI Memphis: Solar Mobile turbines pending permit)
Buying entire substations and behind-the-meter generation
Fast-tracking sites where stranded power exists (Crusoe in West Texas oilfields)

Source: FERC interconnection queue reports; company press releases (Microsoft Sep 2024 Three Mile Island deal, AWS Mar 2024 Talen $650M deal, Google Nov 2023 Fervo PPA).

Lesson 02 — TL;DR

• 6 stages: grid → substation → switchgear → transformer → UPS → PDU.
• UPS bridges the <10-second gap; generators take over for hours-to-days.
• Power density jumped 10–20× when GPUs arrived. Everything downstream has to keep up.
• Grid availability is now the primary site-selection constraint for AI campuses.
• Hyperscalers are buying nuclear, geothermal, and stranded-gas to bypass the queue.

Useful? Share so the next engineer learns this faster.