AI Data Center PartsSearch
Home / How it fits together

How it fits together

An AI data center is four overlapping machines: one moves power in, one moves heat out, one moves data sideways, and one is the building that holds the other three. Follow any chain below — every node links to the part category and the companies on it.

1. Power

Electricity steps down from transmission voltage to the ~1 V a GPU core actually uses, losing a little at every conversion. 800 V DC architectures collapse the middle of this chain: the UPS, floor PDU and power-shelf stages merge into sidecar rectifiers and DC busbars.

Grid / generationHV substation & large transformerHV switchgearMV switchgearMV/LV transformerLV switchgearUPSBatteries / BESSFloor PDU / RPPBuswayRack PDUIn-rack busbarPower shelfVRMChip
  1. Grid / generation — Utilities, IPPs & power providers signing data-center deals
  2. HV substation & large transformer — Large power transformers (345–765 kV) & substations
  3. HV switchgear — HV switchgear, GIS, breakers, protection relays
  4. MV switchgear — MV switchgear, E-houses, power skids
  5. MV/LV transformer — MV/LV transformers
  6. LV switchgear — LV switchgear, switchboards, panelboards
  7. UPS — UPS (static, modular, rotary/DRUPS) & UPS power semis
  8. Batteries / BESS — Energy storage: UPS batteries & BESS
  9. Floor PDU / RPP — Floor PDUs, RPPs, static transfer switches, ATS
  10. Busway — Busway / busbar trunking
  11. Rack PDU — Rack PDUs
  12. In-rack busbar — In-rack busbar & power whips
  13. Power shelf — Power shelves & PSUs
  14. VRM — Board-level power: VRMs, power stages, PoL, power semis
  15. Chip — Merchant GPUs / accelerators

2. Heat

Every watt in becomes heat out. In a liquid-cooled rack the path runs from the die through cold plates and a coolant distribution unit to the facility water loop, and finally to the atmosphere — or to a district-heating network. The air-cooled variant swaps the first half for heatsinks, fans and CRAH units.

ChipLid / TIMCold plateRack manifoldQuick disconnectsCDUFacility water loopChillerDry cooler / towerAtmosphere or heat reuse
  1. Chip — Merchant GPUs / accelerators
  2. Lid / TIM — Chip-level thermal: lids, heat spreaders, TIMs
  3. Cold plate — Direct-to-chip cold plates
  4. Rack manifold — Rack manifolds
  5. Quick disconnects
  6. CDU — Coolant distribution units
  7. Facility water loop — Facility water loop hardware
  8. Chiller — Chillers & compressors
  9. Dry cooler / tower — Dry coolers, adiabatic/evaporative units, cooling towers, fans & motors
  10. Atmosphere or heat reuse — Heat reuse & district heating interfaces

Air-cooled variant: heatsink → fans → CRAH / fan wall → chiller.

3. Data

Two networks with different physics: the scale-up fabric inside the rack (copper, today) and the scale-out network between racks (optics). Inside the package sits a third: die to HBM across an interposer.

GPUNVLink / scale-up fabricNIC / SuperNICOptical transceiveror copper DAC/AECLeaf/spine switchDCI / coherent opticsNext campus
  1. GPU — Merchant GPUs / accelerators
  2. NVLink / scale-up fabric — Scale-up fabric
  3. NIC / SuperNIC — NICs / SuperNICs
  4. Optical transceiver — Optical transceivers
  5. or copper DAC/AEC — Copper: DAC / ACC / AEC cables
  6. Leaf/spine switch — Switch systems & white-box ODMs
  7. DCI / coherent optics
  8. Next campus

…and inside the package

Die→Interposer / bridge→HBM→ABF substrate→PCB

4. Shell & site

The building is a two-year project wrapped around a six-month fit-out. Land and power come first — interconnection queues, not concrete, set the schedule.

Land & power procurementDesign (A&E / MEP)Structure & envelopeGeneral contractorMEP fit-outCommissioningDeveloper / operatorTenant
  1. Land & power procurement — Site selection, power procurement, market intelligence, brokerage
  2. Design (A&E / MEP) — Architecture, engineering & MEP design
  3. Structure & envelope — Structure, steel, precast, envelope, site equipment
  4. General contractor — General contractors & design-build
  5. MEP fit-out — MEP & electrical contractors
  6. Commissioning — Testing, commissioning & certification
  7. Developer / operator — Data center developers, operators & REITs
  8. Tenant — Hyperscalers, neoclouds & model labs

One rack, exploded

The GB300 NVL72 teardown walks a single rack from the outside in — cabinet, busbar, power shelves, compute trays, switch trays, the copper spine, cold plates and the CDU — with the companies reported on every line.

Glossary

PUE
— Power usage effectiveness — total facility power divided by IT power. 1.0 is perfect; modern hyperscale runs near 1.1.
ORV3
— Open Rack V3, the OCP rack standard with a shared 48 V DC busbar fed by power shelves.
CoWoS
— TSMC's 2.5D advanced packaging (chip-on-wafer-on-substrate) joining GPU dies and HBM on an interposer — the historical supply bottleneck.
HBM
— High-bandwidth memory, DRAM dies stacked and placed next to the GPU die inside the package.
CDU
— Coolant distribution unit — the heat exchanger and pumps separating the IT liquid loop from the facility loop.
N+1 / 2N
— Redundancy schemes: one spare unit per group (N+1) versus a full duplicate system (2N).
UQD
— Universal quick disconnect, the OCP-standardized dripless coupling on liquid-cooled racks.
NVL72
— Nvidia's rack-scale system: 72 Blackwell-class GPUs and 36 Grace CPUs on one NVLink domain, ~120–140 kW per rack.
DCI
— Data center interconnect — coherent optics linking buildings and campuses.
GOES
— Grain-oriented electrical steel, the core material of large power transformers.
BESS
— Battery energy storage system, grid-scale batteries used for peak, backup and grid services.
DRUPS
— Diesel rotary UPS — a flywheel and diesel engine replacing batteries in some designs.