The Physical Stack
← Full stack narrative
02 · Campusscroll = depth

Hyperscale Data Centers

AI training & inference campuses

100+ MW

per campus — some AI builds push toward gigawatt scale

A modern AI campus is not a building with servers. It is a power plant that happens to compute — electricity in, tokens out, heat rejected at industrial scale.

Depth strataL1/4
  1. Campus as plantL1

    Substations, halls, cooling yards as one thermal machine

    Megawatts enter at the fence; heat leaves through chillers

    Interconnection queues now gate AI capacity more than GPUs

  2. Aisle thermodynamicsL2

    Hot/cold corridors, liquid loops, rack as a furnace

  3. Package & fabricL3

    HBM stacks bonded to accelerator silicon

  4. Gate → tokenL4

    Nanometer switches flipping into model activations

Mechanism beats (4)
  1. 1
    Racks as thermal machines

    GPU racks draw 40–120+ kW each. The design problem is power delivery + heat extraction before it is FLOPS.

  2. 2
    Liquid cooling takes over

    Air hits a wall. Direct-to-chip liquid and rear-door heat exchangers move heat into facility water loops, then to chillers or free cooling.

  3. 3
    PUE is the scoreboard

    Power Usage Effectiveness tracks overhead. Best campuses approach ~1.1; every tenth of a point is megawatts of non-compute waste.

  4. 4
    Interconnect is the new chassis

    Training clusters live or die on NVLink / InfiniBand / Ethernet fabrics. The network is as capital-intensive as the GPUs.

Scale

One hyperscale campus can draw like a small city

≈ tens of thousands of homes continuous load

Stakes

Grid interconnection queues

Whoever secures power + land + fiber first locks in a decade of training and inference advantage.

spatial stageCampus → micro
loading depth
Depth · Campus

A hyperscale site is a power plant that happens to compute.

Halls, substations, and cooling yards as one thermal machine

5.0e+2 m
10²–10³ m footprint
scroll 0%
macroorders of magnitudemicro
  1. Campus
  2. Server hall
  3. GPU rack
  4. Accelerator
  5. Die / transistor
  6. Bit / token

Enterprise decision brief

From technical spectacle to an executable decision.

Designed for

AI infrastructure teams · campus developers · energy and grid partners

Can the site secure firm power, cooling, fiber, and an interconnection path before compute demand outruns the physical campus?

01 · Operating model

  • ↳Firm-power and interconnection sequence
  • ↳Rack density, liquid loops, and heat rejection
  • ↳Network fabric and failure-domain design
  • ↳Capacity phasing, telemetry, and operating ownership

02 · Decision artifacts

  • ↳Power-to-rack capacity model
  • ↳Thermal and water boundary map
  • ↳Campus dependency architecture
  • ↳Phased commissioning roadmap

03 · Diligence questions

  1. 01What is the first binding constraint: grid, cooling, fiber, or permits?
  2. 02Which load can be phased without stranding infrastructure?
  3. 03How are thermal, electrical, and compute incidents isolated?

Governance boundary

Illustrative architecture only. Final campus design requires utility studies, OEM specifications, local permitting, environmental review, and stamped engineering.

Scope a decision brief →

Grid interconnection queues

The constraint shifted from GPUs to substations and generation. Multi-year interconnection wait times now gate AI capacity.

Whoever secures power + land + fiber first locks in a decade of training and inference advantage.

Site selection is now a three-variable optimization: cheap firm power, fiber routes, and water/permitting — not cheap land.