What goes into a Large colo / hyperscale building build
10–100 MW · typically: single large buildings, wholesale colocation, cloud regions · 119 of 130 catalog parts apply.
Compute
- Merchant GPUs / accelerators — Same die designs ship into an enterprise pilot rack and a gigawatt campus alike; only unit count and generation mix change with scale.
- Hyperscaler custom ASICs — A custom ASIc program only pencils out once a hyperscaler's own deployment volume is large enough to amortize the design cost, so presence skews toward hyperscale and campus-scale fleets even though the finished chips can run in smaller server pools too.
- ASIC design services, IP & EDA — Not tied to any single facility; revenue tracks the number of active chip design programs across the industry rather than deployed unit count at any one site.
- Startup accelerators — Deployment tends to concentrate with whichever customer or cluster has qualified a given architecture, rather than spreading evenly; presence at any tier depends on that customer's buying decision.
- Server CPUs — CPU count per server is roughly fixed regardless of facility size; total volume scales with server count across tiers, not with per-server intensity.
- NICs / SuperNICs — At least one NIC ships per server regardless of scale; port count and speed step up as cluster size and GPU generation move from enterprise toward campus- and gigawatt-class builds.
- DPUs — Attach rate depends on server architecture and hyperscaler infrastructure choices rather than facility size; not universal even in gigawatt-class fleets.
- FPGAs — Deployed in smaller, targeted volumes tied to specific use cases rather than scaling linearly with accelerator count as facility size increases.
- Leading-edge foundry — Wafer volume ordered scales with total accelerator unit count across the industry, not with the size of any single data center buildout.
- Silicon wafers — Wafer starts scale with total foundry capacity booked across all customers industry-wide, not with any single data center's size.
- 2.5D/3D advanced packaging — Every merchant and custom AI accelerator needs advanced packaging; volume tracks total accelerator unit count across all tiers rather than any single facility's footprint.
- Packaging & bonding equipment — Capacity investment here tracks industry-wide advanced-packaging volume rather than any single data center build.
- ABF / high-layer package substrates — Substrate volume scales with package count, which scales with total accelerator unit count across all tiers.
- Substrate materials — Material volume tracks substrate production industry-wide; a single input can gate an entire generation's packaging output regardless of which facility tier the finished accelerators end up in.
- Glass-core substrates — Still a qualification-stage technology; presence depends on which packaging generation a given accelerator design adopts rather than on data center site size.
- Interposers, silicon bridges, RDL — Volume follows advanced-packaging unit count industry-wide rather than the size of any one facility.
- Chip-level thermal: lids, heat spreaders, TIMs — One lid and one set of thermal interface materials per die regardless of facility tier; the specific material chosen can depend on whether the finished server is air- or liquid-cooled downstream.
- Semiconductor test & burn-in — Test and burn-in throughput scales with total die volume shipped industry-wide, not with the size of any single data center.
- Fab equipment — Tool capacity additions are a global, multi-year capital cycle independent of any single AI data center's footprint.
- Fab materials, masks, gases — Materials consumption tracks fab wafer starts industry-wide; not tied to the size of any single data center.
- HBM — Content is fixed per accelerator package; total HBM demand scales with the number of accelerators deployed rather than with facility size or cooling design.
- DRAM — Quantity tracks node count and the memory-per-socket ratio a platform is designed around, not the facility's physical scale.
- Memory modules — Quantity tracks socket and channel count per server rather than site scale.
- Memory interface / buffer chips — Presence depends on module design — registered and load-reduced modules carry these chips, simpler unbuffered modules do not — so it tracks platform choice as much as facility scale.
- CXL controllers & memory expanders — Presence depends entirely on whether a platform adopts CXL memory pooling; where it does, it appears from enterprise servers up through the largest clusters.
- PCIe/CXL switches & retimers — Content per server scales with endpoint count and trace length rather than facility size; it appears wherever PCIe fan-out exceeds what the host can drive directly.
- NAND flash — Quantity tracks the local storage footprint designed into each server, independent of facility scale or cooling approach.
- Enterprise SSDs & controllers — Drive count and capacity per node scale with local storage design rather than facility size; present from a single edge server up through the largest clusters.
- HDD — Presence depends on whether a site runs a dedicated hot/cold storage split; where it does, capacity scales with total dataset and retention volume rather than rack count.
- AI storage systems — Presence and scale depend on cluster size and whether training data volume justifies a dedicated parallel file system versus general-purpose storage.
- Rack-scale systems — enterprise deployments typically use standalone or multi-GPU servers rather than a full rack-scale system; the fully wired cabinet becomes the norm once GPU counts justify NVLink or Infinity Fabric domains at hyperscale and above.
- EMS, system integration & rack-level test — enterprise servers get factory test at the unit level; hyperscale-and-up rack-scale systems require dedicated rack-level test cells and longer burn-in cycles.
- GPU baseboards & compute trays — present in every GPU server regardless of scale; what changes is tray count per rack and whether the tray is air-cooled or fitted with cold plates.
- Motherboards, CPU trays, switch trays — switch-tray count scales with the number of GPU trays and the scale-up fabric's radix, so it grows faster than headcount-style server metrics as racks get larger.
- High-layer-count PCBs — layer count and fabrication precision scale with board complexity rather than data center size, so the highest-layer-count boards appear wherever the highest-end GPU trays are built, at any deployment scale.
- Copper-clad laminate & PCB materials — an upstream material input present wherever high-layer-count boards are fabricated, independent of the data center's own scale tier.
- Board-level connectors & sockets — connector count per rack rises with tray count and interconnect density rather than with facility size, so it scales with rack generation more than with deployment tier.
- Board-level power: VRMs, power stages, PoL, power semis — present on every GPU and CPU board regardless of facility scale; what changes with rack density is the current and transient-response spec each VRM has to hit.
- Passives — passive count per board is roughly fixed by design regardless of facility scale; total demand instead scales with the number of boards deployed.
- BMC & platform firmware — present on every server board regardless of scale; its operational importance rises with density and liquid cooling rather than its physical presence changing.
- Timing, security & misc board silicon — present on every board regardless of facility scale; specification tightens with interconnect speed rather than with deployment size.
- Power shelves & PSUs — the Open Rack V3 shelf format is specific to rack-scale systems at hyperscale and above; smaller enterprise GPU servers typically use conventional internal PSUs rather than shared rack-level power shelves.
- Fans — fan count per tray declines as liquid cooling covers more of the heat load, but does not disappear even in the highest-density liquid-cooled racks.
- Chassis, rails, rack enclosures — mechanical demands rise with tray weight and plumbing complexity as density increases, even though the part category itself is present at every scale.
- In-rack busbar & power whips — the high-current ±48 V busbar architecture is specific to rack-scale liquid-cooled systems at hyperscale and above; smaller enterprise racks more often use conventional power whips.
- Rack PDUs — present at every scale; capacity per unit and the value of built-in metering both rise with rack power density.
- Air-cooled heatsinks & vapor chambers — shrinks in scope per tray as liquid cooling covers more of the heat load, but remains present for the components liquid cooling doesn't reach, even at the highest density tiers.
Data
- HBM — Content is fixed per accelerator package; total HBM demand scales with the number of accelerators deployed rather than with facility size or cooling design.
- DRAM — Quantity tracks node count and the memory-per-socket ratio a platform is designed around, not the facility's physical scale.
- Memory modules — Quantity tracks socket and channel count per server rather than site scale.
- Memory interface / buffer chips — Presence depends on module design — registered and load-reduced modules carry these chips, simpler unbuffered modules do not — so it tracks platform choice as much as facility scale.
- CXL controllers & memory expanders — Presence depends entirely on whether a platform adopts CXL memory pooling; where it does, it appears from enterprise servers up through the largest clusters.
- PCIe/CXL switches & retimers — Content per server scales with endpoint count and trace length rather than facility size; it appears wherever PCIe fan-out exceeds what the host can drive directly.
- NAND flash — Quantity tracks the local storage footprint designed into each server, independent of facility scale or cooling approach.
- Enterprise SSDs & controllers — Drive count and capacity per node scale with local storage design rather than facility size; present from a single edge server up through the largest clusters.
- HDD — Presence depends on whether a site runs a dedicated hot/cold storage split; where it does, capacity scales with total dataset and retention volume rather than rack count.
- AI storage systems — Presence and scale depend on cluster size and whether training data volume justifies a dedicated parallel file system versus general-purpose storage.
- Scale-up fabric — Present wherever multi-GPU racks or pods exist; the fabric's reach and topology (single rack vs. multi-rack pod) scale with cluster size rather than being an enterprise feature.
- Scale-out switch ASICs — Scales with cluster size: enterprise and small deployments may use merchant switch systems without needing the highest-radix ASICs that only very large clusters require.
- Switch systems & white-box ODMs — Present at every scale; the mix shifts from branded systems at enterprise scale toward white-box, merchant-silicon designs as deployments approach hyperscale and campus scale.
- Optical transceivers — Volume scales directly with GPU and switch port count; negligible at enterprise scale, the dominant network cost line at hyperscale and above.
- Optics contract manufacturing — Volume follows transceiver demand (3.04); relevant wherever high-speed optics are being built at scale, immaterial below hyperscale-level deployments.
- Lasers — Volume follows transceiver and co-packaged optics demand; immaterial outside hyperscale- and campus-scale optical deployments.
- Optical DSPs, drivers, TIAs — Volume follows transceiver and co-packaged optics demand; immaterial outside hyperscale- and campus-scale deployments.
- Silicon photonics & co-packaged optics — An emerging alternative to pluggable optics rather than a universal requirement; presence depends on a design choice by the switch or rack vendor, concentrated at the highest-speed, highest-density deployments where it exists at all.
- Copper: DAC / ACC / AEC cables — Volume scales with GPU and port count; the dominant short-reach interconnect at hyperscale and above, negligible at enterprise scale.
- NVLink copper spine / backplane cartridges — Present wherever a scale-up rack architecture uses a copper backplane; cable count scales with rack GPU density and could shrink or disappear in racks that adopt co-packaged optics.
- Fiber, connectors, structured cabling — Volume scales with optical port count and data hall size; minor at enterprise scale, a substantial physical-infrastructure line at hyperscale and above.
- Timing & telemetry silicon — Present wherever high-speed switching and optics exist; content per port grows with speed grade and cluster complexity rather than with physical size alone.
Power
- Rack-scale systems — enterprise deployments typically use standalone or multi-GPU servers rather than a full rack-scale system; the fully wired cabinet becomes the norm once GPU counts justify NVLink or Infinity Fabric domains at hyperscale and above.
- EMS, system integration & rack-level test — enterprise servers get factory test at the unit level; hyperscale-and-up rack-scale systems require dedicated rack-level test cells and longer burn-in cycles.
- GPU baseboards & compute trays — present in every GPU server regardless of scale; what changes is tray count per rack and whether the tray is air-cooled or fitted with cold plates.
- Motherboards, CPU trays, switch trays — switch-tray count scales with the number of GPU trays and the scale-up fabric's radix, so it grows faster than headcount-style server metrics as racks get larger.
- High-layer-count PCBs — layer count and fabrication precision scale with board complexity rather than data center size, so the highest-layer-count boards appear wherever the highest-end GPU trays are built, at any deployment scale.
- Copper-clad laminate & PCB materials — an upstream material input present wherever high-layer-count boards are fabricated, independent of the data center's own scale tier.
- Board-level connectors & sockets — connector count per rack rises with tray count and interconnect density rather than with facility size, so it scales with rack generation more than with deployment tier.
- Board-level power: VRMs, power stages, PoL, power semis — present on every GPU and CPU board regardless of facility scale; what changes with rack density is the current and transient-response spec each VRM has to hit.
- Passives — passive count per board is roughly fixed by design regardless of facility scale; total demand instead scales with the number of boards deployed.
- BMC & platform firmware — present on every server board regardless of scale; its operational importance rises with density and liquid cooling rather than its physical presence changing.
- Timing, security & misc board silicon — present on every board regardless of facility scale; specification tightens with interconnect speed rather than with deployment size.
- Power shelves & PSUs — the Open Rack V3 shelf format is specific to rack-scale systems at hyperscale and above; smaller enterprise GPU servers typically use conventional internal PSUs rather than shared rack-level power shelves.
- Fans — fan count per tray declines as liquid cooling covers more of the heat load, but does not disappear even in the highest-density liquid-cooled racks.
- Chassis, rails, rack enclosures — mechanical demands rise with tray weight and plumbing complexity as density increases, even though the part category itself is present at every scale.
- In-rack busbar & power whips — the high-current ±48 V busbar architecture is specific to rack-scale liquid-cooled systems at hyperscale and above; smaller enterprise racks more often use conventional power whips.
- Rack PDUs — present at every scale; capacity per unit and the value of built-in metering both rise with rack power density.
- Air-cooled heatsinks & vapor chambers — shrinks in scope per tray as liquid cooling covers more of the heat load, but remains present for the components liquid cooling doesn't reach, even at the highest density tiers.
- MV switchgear, E-houses, power skids — Enterprise sites typically use simpler, non-modular MV switchgear rooms rather than E-houses; the pre-fabricated form factor becomes standard once a campus is adding capacity in repeatable blocks.
- MV/LV transformers — Present at every scale; only the count and size of units per site changes.
- LV switchgear, switchboards, panelboards — Present at every scale; board size and ampacity rating scale with hall density rather than campus size alone.
- Busway / busbar trunking — Common wherever hall layouts are standardized and repeated across many rows; smaller enterprise rooms more often use cable and tray instead.
- Floor PDUs, RPPs, static transfer switches, ATS — Present at every scale; unit count and ampacity per zone scale with hall density.
- Circuit protection, fuses, surge, grounding — Present at every scale as an embedded component of the switchgear, transformers, and distribution equipment around it rather than a standalone system.
- LV cable, wire, tray, raceway — Present at every scale; cable gauge and tray capacity scale with rack density.
- UPS (static, modular, rotary/DRUPS) & UPS power semis — Present at every scale; the architecture shifts from a handful of large static UPS at hyperscale and above toward more distributed or rotary designs depending on site design choices.
- Energy storage: UPS batteries & BESS — UPS batteries are present at every scale; grid-scale BESS is largely an ai_campus and gw_class phenomenon tied to interconnection and load-smoothing needs at that size.
- Power smoothing / supercapacitors — Relevant wherever GPU clusters are large enough to synchronize load swings across many racks; less applicable to enterprise-scale deployments with smaller, less correlated workloads.
- Backup generation: gensets, fuel systems, HVO — Present at every scale, from a single genset at enterprise up to large paralleled fleets at ai_campus and gw_class sites.
- Paralleling switchgear & generator controls — Scales with the number of generators being paralleled; a single-genset enterprise site typically has little need for this layer.
- Metering, power quality, monitoring — Present at every scale; the number of meter and monitoring points scales with the number of tenants and the granularity of billing required.
- Reciprocating engines & engine microgrids — Scales from a handful of modular units for bridge power at a single building up to large microgrid arrays at campus scale; less common at enterprise-scale sites that rely on grid power.
- Fuel cells — Deployed at hyperscale_building and ai_campus scale as onsite or behind-the-meter generation; less common at smaller sites given per-megawatt capital cost.
- Utilities, IPPs & power providers signing data-center deals — Relevant at every scale tier, since even an enterprise-scale site ultimately buys power from a utility or IPP; the size and structure of the deal (standard tariff versus negotiated large-load contract) scales with site size.
Heat
- Direct-to-chip cold plates — Present at any scale that runs liquid-cooled compute; count per rack tracks GPU, CPU, and switch die count rather than facility size.
- Rack manifolds — Sized to rack tray count and GPU density rather than facility scale; higher-density racks need more ports and higher flow rates per manifold.
- Quick disconnects — Count per rack scales directly with the number of liquid-cooled trays and cold plates.
- Hoses, tubing, fittings, valves — Volume tracks total loop length installed, which scales with the count of liquid-cooled racks and rows rather than facility footprint alone.
- Coolant distribution units — Form factor is a design choice, not a strict scale tier: in-rack and in-row units suit smaller or mixed deployments, while facility-class (1–2.5 MW-plus) units suit large, liquid-dense halls.
- CDU internals: plate heat exchangers, pumps, filters — Component sizing and redundancy (single pump versus N+1) scale with the CDU's rated capacity, from in-rack units to facility-class systems.
- Rear-door heat exchangers — Used as a hybrid or bridging measure in mixed air/liquid racks or facilities not built for full direct-to-chip cooling; less common at the highest-density AI campus and GW-class builds, which favor direct-to-chip.
- Immersion cooling systems — Deployment depends on a design choice to build for immersion rather than direct-to-chip; where used, tank or chamber count scales with server count rather than facility power tier.
- Fluids & coolant treatment — Present wherever liquid cooling is deployed; fluid family depends on whether the design uses direct-to-chip cold plates or immersion tanks.
- Leak detection, flow & pressure sensing — Monitored point count scales with the number of quick disconnects and manifold branches in the rack, which itself tracks liquid-cooled GPU density.
- Rack/row cooling controls & software — Complexity rises with the number of CDUs, racks, and sensor points coordinated together rather than with facility size alone; larger halls push controls toward row- and facility-level coordination rather than per-rack management.
- Chillers & compressors — N+1 chiller redundancy at enterprise scale, shifting to 2N or fully parallel chiller blocks with independent compressor trains at hyperscale and gw-class campuses.
- Dry coolers, adiabatic/evaporative units, cooling towers, fans & motors — Enterprise sites often run a handful of packaged units; hyperscale and campus-scale builds move to arrays of large dry coolers or cooling-tower cells sized in parallel blocks.
- CRAH / CRAC / fan walls / in-row units — Sole cooling method at air-cooled enterprise and hyperscale halls; shrinks to a residual and backup role per rack once a site is liquid-cooling-first at ai_campus and gw_class scale.
- Air handling, economizers, filtration — Sized to the full IT heat load at air-cooled enterprise and hyperscale sites; shrinks to a shell-conditioning and residual-load role once liquid cooling dominates at ai_campus and gw_class.
- Facility water loop hardware — Single loop with N+1 pumps at enterprise scale; parallel or 2N loops, with a dedicated liquid-cooling-side loop layered on, at ai_campus and gw_class.
- Water treatment, reuse, consumption — Basic loop water treatment exists at every scale; dedicated reuse and reclaim infrastructure, and consumption permitting, become material design and site-selection factors at ai_campus and gw_class.
- Heat reuse & district heating interfaces — Presence depends on site-specific access to a district heating network or heat customer rather than on facility size alone; more thermodynamically attractive at ai_campus and gw_class sites running hotter liquid-cooled loops.
- Building management, controls, environmental sensors — Present at every scale; sophistication rises from a basic BMS at enterprise to AI-driven, multi-loop optimization layered over PLCs at ai_campus and gw_class.
- Fire detection & suppression — Present at every scale as a code requirement; zoning and agent quantity scale with hall size and, for liquid-cooled sites, with added coverage for manifold and CDU areas.
Shell & site
- Data center developers, operators & REITs — Present at every scale from a single edge cabinet to a multi-gigawatt campus; what changes is ownership structure — enterprise and edge sites are more often owned outright by the tenant, while hyperscale and larger sites increasingly involve a developer/REIT as landlord.
- Hyperscalers, neoclouds & model labs — Concentrated at hyperscale-building scale and above; individual labs and smaller neoclouds may lease enterprise-scale capacity as well, but the volume that defines this category sits at campus and gigawatt scale.
- Architecture, engineering & MEP design — Engaged on essentially every purpose-built data center above the smallest edge deployments; the scope of work (and the premium commanded by firms with AI-campus experience) grows with scale and with liquid-cooling complexity.
- General contractors & design-build — Present at every scale above the smallest edge site; project size and the premium for proven AI-campus delivery experience both scale up with site size.
- MEP & electrical contractors — Scope per square foot grows with density — a liquid-cooled AI campus requires substantially more electrical and piping labor hours than an air-cooled enterprise hall of the same footprint.
- Structure, steel, precast, envelope, site equipment — Volume of structural material scales roughly with building footprint and floor loading requirements, which both grow from enterprise through gigawatt-class campuses.
- Prefab / modular data centers & power/cooling skids — Adoption is a design choice rather than a function of size alone — used where speed of deployment and repeatability matter most, which skews toward hyperscale and larger AI campuses, but not universal even there.
- Site utilities, water & fiber entrance — Water demand scales with cooling load and with the share of liquid or evaporative cooling in the design; fiber entrance requirements scale with site criticality rather than size alone.
- Physical security & access — Present at nearly every scale; scope (number of mantraps, perimeter length, guard headcount) scales with site footprint and tenant compliance requirements rather than with compute density.
- Testing, commissioning & certification — Scope and duration of the commissioning cycle grow with facility complexity and density; load bank capacity required scales directly with the IT load being tested.
Upstream supply
- Copper — Present at every scale as wire and busbar; the cold-plate use is limited to liquid and hybrid deployments, while busbars, cable and transformer windings apply regardless of cooling approach.
- Aluminum — Structural and cabling uses scale with total footprint at every tier; the heat-sink role is most prominent in air-cooled and hybrid deployments and recedes on components that move to direct liquid cooling.
- Grain-oriented electrical steel & amorphous cores — Applies from enterprise scale up, wherever a dedicated power or distribution transformer is used; material draw grows sharply with transformer size and count toward hyperscale, campus and gigawatt-class sites.
- Rare earths & permanent magnets — Present at every scale through fans, pumps and gensets; motor count and magnet draw per site rise with fan-wall and pump density in liquid-cooled hyperscale, campus and gigawatt-class builds.
- Specialty glass & glass fiber — Optical-fiber demand scales with network fabric density at every tier and is highest in hyperscale, campus and gigawatt-class clusters; the glass-core substrate use applies only to the largest chip packages, independent of site size.
- Refrigerants & insulating gases — Insulating-gas demand tracks switchgear count and voltage class, heaviest at campus and gigawatt-class sites and negligible below enterprise scale; refrigerant demand tracks mechanical cooling plant size and is present in hybrid designs too, since chillers still handle facility heat rejection alongside a liquid-cooled loop.
- Industrial & specialty gases — The fab-supply use is independent of data-center site size, since it tracks wafer capacity rather than IT load; the fire-suppression use is present at every tier and scales with the number and size of protected spaces.
- Critical minerals: gallium, germanium, indium, tungsten, beryllium — These are chip- and component-level inputs rather than site infrastructure, so presence is uniform across site sizes; total draw scales with GPU, power-semiconductor and optics volume rather than with facility footprint.
- DCIM, digital twins, capacity planning — Present at every scale; at hyperscale and above it typically expands from asset tracking into full digital-twin modeling used during design and commissioning, not just live monitoring.
- Cluster management & GPU orchestration — Present wherever GPU clusters run; complexity and the case for dedicated orchestration software grow with cluster size rather than with any single facility's density mode.
- Site selection, power procurement, market intelligence, brokerage — Relevant at every scale a new site is being evaluated; the power-procurement piece becomes the dominant consideration at ai_campus and gw_class scale.
- Capital & financing — Relevant across all scales as a funding source; asset-backed structures like GPU-backed lending become more prominent at ai_campus and gw_class scale where compute purchases are largest.
- Test & measurement instruments — Used at every scale during manufacturing, integration, and commissioning; intensity of use scales with the pace and volume of hardware being qualified and deployed rather than with facility size alone.
- Robotics, inspection & operations hardware — Adoption is a facility-operations choice rather than a scale requirement; more common at larger sites where the inspection burden across a bigger physical footprint is greater.