Working exhibit / energy path for AI racks

The AI data center energy stack, from utility grid to GPU board

The energy problem has two different questions. First: are enough megawatts available at the site? Second: can those megawatts be transformed, protected, distributed, and converted into stable power for dense NVIDIA racks without crashing the workload?

MW first If Taipower or another utility cannot allocate power, the site is dead before design starts.
kW/rack AI racks can exceed 100 kW, so old data center assumptions break quickly.
Clean power Training jobs hate interruptions, voltage instability, and messy failover events.
Long leads Transformers, switchgear, UPS, PDUs, and busbars can become schedule killers.
01

Utility grid

Plain English: The source of the electricity.

Think

  • Taipower
  • Generation mix
  • Transmission capacity
  • Renewable PPAs

Why it matters

This decides whether a site can get 5 MW, 10 MW, 20 MW, or nothing useful.

Failure mode

The buyer finds land, but the utility cannot allocate enough power on the timeline.

Transmission towers and substation equipment

generation + grid

capacity gate
downstream to the site
02

Substation

Plain English: The grid node that brings high-voltage power closer to usable facility power.

Think

  • Distance to site
  • Available capacity
  • New feeder path
  • Utility approval

Why it matters

Close substation access can make a site realistic. A far or congested substation can ruin the economics.

Failure mode

There is grid power in the region, but no practical way to deliver it to this parcel.

Electrical substation equipment

substation

capacity gate
voltage steps down
03

Transformers

Plain English: The equipment that converts high voltage into lower voltage the facility can use.

Think

  • Step-down transformers
  • Lead times
  • Redundancy
  • Spare capacity

Why it matters

Transformers are the bridge between utility-scale electricity and facility-scale electricity.

Failure mode

The site has a theoretical power path, but the transformer lead time blows up the deployment date.

transformer

delivery bottleneck
controlled inside the facility
04

Switchgear

Plain English: Giant industrial circuit breakers that control and route electricity safely.

Think

  • Breakers
  • Protection relays
  • Isolation
  • Safe maintenance

Why it matters

At AI power densities, a fault is not a small inconvenience. It is a safety, uptime, and equipment risk.

Failure mode

The electrical room cannot safely route the power density the racks demand.

Electrical switchgear cabinets

switchgear

delivery bottleneck
stabilized and buffered
05

UPS

Plain English: Battery-backed systems that keep power stable during outages or transitions.

Think

  • Batteries
  • Inverters
  • Ride-through time
  • Generator handoff

Why it matters

Even tiny interruptions can crash AI training jobs and waste expensive GPU time.

Failure mode

The grid blips, generator transition is messy, and the cluster drops a job.

UPS batteries

reliability gate
distributed to rows
06

PDUs

Plain English: Power Distribution Units that send electricity to rows of racks.

Think

  • Row-level distribution
  • Monitoring
  • Breaker panels
  • Load balancing

Why it matters

PDUs translate facility power into rack-row power with measurement, protection, and control.

Failure mode

The room has total MW, but cannot distribute it cleanly across the rack layout.

PDU

delivery bottleneck
final high-current rack feed
07

Busbars and rack power shelves

Plain English: Final high-current delivery into AI racks.

Think

  • Copper busbars
  • Rack power shelves
  • High-current connectors
  • Thermal monitoring

Why it matters

This is where abstract MW becomes very real current entering a dense rack.

Failure mode

Connectors, busbars, or shelves cannot handle the amperage or heat safely.

AI server rack with power delivery hardware

rack feed

delivery bottleneck
converted on the board
08

GPU boards

Plain English: Voltage regulators turn rack power into the exact levels GPUs need.

Think

  • VRMs
  • Power stages
  • Telemetry
  • Thermal safety

Why it matters

GPUs need precise, stable power at massive current. Bad conversion creates crashes, heat, or hardware failure.

Failure mode

The facility works, but board-level power instability causes throttling, failures, or reliability problems.

Server board with voltage regulators and capacitors

GPU board power

final conversion

The simple version

Power is not one thing. It is allocation, transformation, protection, distribution, and board-level conversion.

Why this matters to the wedge

  • 1 It makes site scoring realDo not ask only "does the site have power?" Ask where in the chain the power is confirmed.
  • 2 It exposes fake supplyA landlord may have land, but no substation path, transformer, switchgear, or AI-rack-ready distribution.
  • 3 It gives Lawrence a buyer-facing checklistThe buyer wants a credible power path, not vague assurances about Taiwan relationships.