Aircloud Compute is now Lightyard. Aircloud continues as a private inference platformAircloud continues for inference

Sites

How to build a data center for GPU clusters

What it takes to build or fit out a data center for NVIDIA GPU clusters: what changes at 120 kW a rack, five ways to house a cluster, the stages of a build with public figures on time and cost per MW and who does what.

All notesPublished 14 min read
In this note
  1. What makes a GPU data center different
  2. Five ways to house a GPU cluster
  3. The stages of a data center build
  4. Site and power
  5. Permits, design and long-lead equipment
  6. Construction and commissioning
  7. What it costs to build a data center for GPUs
  8. Who does what on a data center build
  9. Where the building ends and the cluster begins
  10. Building with Lightyard

Building a data center for GPU clusters starts with the rack. NVIDIA puts a GB200 or GB300 NVL72 rack at about 120 kW (NVIDIA (opens in a new tab)), while the most common rack in Uptime Institute's 2026 survey still draws 4 to 5 kW (Uptime Institute (opens in a new tab)). In the US, a new hall takes about three years from breaking ground to a running AI supercomputer, by NVIDIA CEO Jensen Huang's estimate (Fortune (opens in a new tab)). JLL puts the average wait for a grid connection in primary markets at more than four years (JLL (opens in a new tab)). And the most advanced AI facilities in the US cost US$14 to US$16 per watt to build (Turner & Townsend (opens in a new tab)), before a single GPU is bought.

Most buyers don't build a new hall. They lease colocation space, upgrade a data center they already run or retrofit a powered building or mining site. Some set a prefabricated module on land with power. Below is what changes when the load is GPU racks, when each option fits, how a build runs and who does what.

What makes a GPU data center different

A data center for GPU clusters differs from a general IT hall in five places: power per rack, cooling, floor loading, network and the size of the power connection.

Typical data center rack8-GPU B300 systems, air-cooledGB300 NVL72, liquid-cooled
Power per rack4 to 5 kW is the most common band; 7.8 kW average without the densest sites14.5 kW per system, so 29 to 58 kW for two to four systems in a rack (our arithmetic)About 120 kW; about 150 kW with its share of networking, from 1.2 MW per eight racks (our arithmetic)
CoolingRoom airRoom air, inlet at 10 °C to 35 °CLiquid to the GPUs and CPUs, air for networking and storage
Arrives asServers racked on site10U systems racked on siteOne integrated rack, designed for rapid installation
SourceUptime Institute (opens in a new tab)NVIDIA DGX B300 (opens in a new tab)NVIDIA user guide (opens in a new tab), reference architecture (opens in a new tab)

Power per rack

Most halls weren't built for this. In Uptime's 2026 survey, 76% of operators had no rack at 30 kW or more, and about one in seven had racks above 50 kW. A room designed around 5 to 10 kW positions has busway, whips and UPS blocks sized for them.

An NVL72 rack needs about 120 kW at one position. NVIDIA's user guide describes eight power shelves, each needing 33 kW of input, arranged for N+N redundancy (NVIDIA (opens in a new tab)). Size the feeds to each position for that design, not for the average draw.

The ceiling keeps moving. NVIDIA says 54 V in-rack power distribution "begins to hit physical limits" as racks pass 200 kW, and plans 800 V DC distribution for 1 MW racks starting in 2027 (NVIDIA (opens in a new tab)). A hall designed now should leave room in its electrical rooms and pathways for that change.

Liquid cooling for NVL72-class racks

In an NVL72 rack, liquid runs through manifolds to the compute trays, while networking and storage stay on air (NVIDIA (opens in a new tab)). The room needs both systems. On the liquid side that means:

  • Coolant distribution units (CDUs) sized for the heat load plus redundancy
  • A facility water loop at the temperature and flow the CDUs need, with dry coolers or chillers to reject the heat
  • Secondary piping and manifolds to every rack position, with isolation valves
  • Leak detection in the racks and along the piping
  • Filtration and chemical treatment for the loop

Turner & Townsend reports that the cost premium for AI-ready facilities is concentrated in the mechanical package (Turner & Townsend (opens in a new tab)). In its 2026 survey, 52% of respondents expected liquid cooling in more than half of the data centers they work on (Turner & Townsend (opens in a new tab)).

Air-cooled systems avoid the liquid plant but not the heat. NVIDIA rates a DGX B300 at about 1,500 CFM of airflow at 70% fan speed (NVIDIA (opens in a new tab)). Four in one rack put close to 60 kW into the aisle, which calls for containment and more cooling per row than a general IT room has.

Floor loading and delivery

A DGX B300 weighs 168 kg, so four of them put about 670 kg of servers on one rack footprint before the rack, cabling and power gear (our arithmetic). An NVL72 rack arrives as one integrated unit with 18 compute trays, nine NVLink switch trays, power shelves and coolant manifolds (NVIDIA (opens in a new tab)). Get the loaded weight of the exact rack from its manufacturer and check it against the floor's rating for static and rolling loads, and against every ramp, lift, door and floor on the route from the dock.

Network

NVIDIA's GB300 reference architecture uses four physical networks: NVLink inside each rack, a compute fabric between racks, a storage and in-band management Ethernet fabric and an out-of-band management network (NVIDIA (opens in a new tab)). It also advises that if you start with fewer nodes than a full scalable unit, you build the fabric for the full unit, leaf switches and leaf-to-spine cables included. For the building, that means cable trays sized for the finished cluster, rack rows laid out so runs stay within reach of the optics and two separate fiber entries for the uplinks.

Power availability

One scalable unit of eight NVL72 racks needs 1.2 MW (NVIDIA (opens in a new tab)). Ten units need 12 MW, which can mean a utility load study, a new service and sometimes a substation. That's why power, not construction, sets the date for most new builds.

Five ways to house a GPU cluster

OptionFits whenWhat gets builtWhat sets the date
ColocationYou need capacity in a region within a year, without owning a buildingA fit-out in an existing hall: CDUs, piping, busway, cablingFinding liquid-ready space, then the fit-out
Your own data centerYou run a hall with spare power and want the cluster inside your perimeterUpgrades in a live roomThe room's power, cooling and floor
RetrofitYou hold a powered building or mining site that wasn't built for GPU racksPower distribution, liquid cooling, structure and networkThe building's condition
ModularYou hold land with power but no buildingPrefabricated modules set on a prepared padModule build, site work and the power connection
New purpose-built hallYou need tens of megawatts for a decade and want to own the buildingEverything, from the substation upPower, permits and long-lead electrical gear

Colocation is the fastest route if the space exists. JLL has North American vacancy at 1% for the third year running, with most tenants signing now contracting for 2028 deliveries (JLL (opens in a new tab)). CBRE says ready-for-service dates on fit-outs are stretching "from weeks to several months" (CBRE (opens in a new tab)). Check a "liquid-ready" hall against the list above before signing. Whether to own the cluster at all is a separate question, covered in buy vs rent a GPU cluster.

Your own data center works when the room has power to spare. More than a quarter of Uptime's 2026 respondents (27%) work in data centers over 15 years old (Uptime Institute (opens in a new tab)). A room like that can often take air-cooled 8-GPU systems at moderate density well before it can take NVL72 racks. Survey the power at each position, the cooling and the floor, then size the cluster to the room.

A retrofit starts from a power connection that already exists, which takes the longest item off the schedule. A GPU cluster has to stay up through maintenance, though, and NVIDIA recommends a facility built to Uptime Institute Tier 3 or an equivalent, with concurrent maintainability and no single point of failure (NVIDIA (opens in a new tab)). So a retrofit usually adds redundant power paths and backup power as well as liquid cooling, structural work and the network.

A modular build moves most of the construction into a factory. Modules are built while the pad, power and fiber are prepared on site, then set and connected. It still needs a power connection, site permits and commissioning.

A new hall is the slowest route, and the rest of this guide follows its stages. For month-by-month plans, see how long it takes to build an AI data center.

The stages of a data center build

A purpose-built hall goes through six stages before the first rack arrives. They overlap: long-lead equipment is ordered during design, and permits run alongside it.

StageWhat happensPublic figuresOur planning range
Site and powerSite search, utility load study, interconnection agreement, sometimes a substationGrid connections average 4+ years in primary markets (JLL); 53% of respondents report 1 to 2 years for a utility connection, 32% more than 2 (Turner & Townsend)6 to 18 months to a utility commitment; the connection can take years
PermitsZoning, building permits, environmental review, air permits for generatorsLocal approval as critical to site selection as power and fiber (CBRE)6 to 12 months, alongside design
DesignBasis of design, MEP engineering, construction drawingsLong-lead orders start before design is fixed (Turner & Townsend)Alongside permits
Long-lead procurementTransformers, switchgear, generators, UPS, chillers, CDUsPower transformers 128 weeks, switchgear 44 weeks (Wood Mackenzie)Ordered during design
ConstructionCivil works, shell, electrical and mechanical plant, white spaceAbout three years from breaking ground to an AI supercomputer (Huang)12 to 24 months
CommissioningEquipment tests, start-up, functional tests, integrated systems testLonger with liquid cooling: loops need flushing, filtration and chemical baselining (Turner & Townsend)Inside the construction schedule

The planning ranges are our estimates for a US site and a team that has built before, the same ones behind our deployment timeline. Add one to three months after commissioning to install, burn in and accept the cluster.

Site and power

Power decides most schedules now. JLL counts more than 66 GW under construction in North America, 77% of it in what JLL calls frontier markets (JLL (opens in a new tab)). Turner & Townsend's 2026 survey found 56% of respondents naming access to power as a major obstacle to delivery, and 42% using or exploring alternative power sources, up from 14% who included them in 2025 (Turner & Townsend (opens in a new tab)).

Getting power also means signing for it. Under AEP Ohio's data center tariff, adopted in July 2025, new data center loads of 25 MW or more pay a minimum demand charge whether they use the power or not, on a contract of up to four years of ramp plus eight years. The exit fee is 36 months of minimum charges (AEP Ohio (opens in a new tab)). Finance should read a power contract like that the way it reads a lease.

Before committing to a site, get these in writing:

  • The capacity the utility will deliver and the date, plus what you pay if you take less
  • The upgrades the connection needs and who pays for them
  • Zoning that allows a data center and its backup generators

Permits, design and long-lead equipment

Most builds need land-use approval, building and electrical permits and environmental permits for backup generators. In the US, EPA rules limit how backup generators run outside emergencies, with up to 50 hours a year allowed for local grid support under set conditions (EPA (opens in a new tab)). CBRE reports that local opposition "has also become a serious obstacle" (CBRE (opens in a new tab)).

Design should start from the hardware. The basis of design fixes power per rack position, redundancy, cooling type and water temperatures, floor loading and the fabric layout, so the GPU platform has to be chosen, or at least bounded, first. NVIDIA's reference architecture asks for Uptime Institute Tier 3, TIA-942-B Rated 3 or EN 50600 Availability Class 3 (NVIDIA (opens in a new tab)). Uptime says a Tier III facility "requires no shutdowns for equipment replacement and maintenance," with a redundant delivery path for power and cooling (Uptime Institute (opens in a new tab)).

Long-lead equipment has to be ordered before the design is finished. Wood Mackenzie's October 2025 update put average lead times at 128 weeks for power transformers, 143 weeks for generator step-up units and 44 weeks for switchgear (Wood Mackenzie (opens in a new tab)). Turner & Townsend notes that on AI projects, "contract awards and long-lead procurement have to begin before design is fixed" (Turner & Townsend (opens in a new tab)). That puts deposits on transformers and switchgear at risk before the permits are in hand, which makes it a financing question as much as an engineering one.

Construction and commissioning

Huang's three years counts from breaking ground, so it covers construction and commissioning together, not the search for a site and power before it. CBRE reports that power constraints are extending completion timelines (CBRE (opens in a new tab)), and Cushman & Wakefield that labor shortages are raising costs and timelines (Cushman & Wakefield (opens in a new tab)).

Commissioning proves the building works before any IT goes in. ASHRAE Guideline 0 describes a process that runs from predesign through warranty expiration, to verify that the facility meets the owner's project requirements (ASHRAE (opens in a new tab)). The testing usually runs in this order:

  1. Factory witness tests of major electrical and mechanical equipment
  2. Installation checks and start-up of each system
  3. Functional tests of each system on its own
  4. An integrated systems test: load banks stand in for the IT load while the team simulates utility failures and switches between redundant paths
  5. For liquid cooling, flushing, filtration and a chemical baseline of the loop

That ends the building's work. Racking, burn-in and acceptance of the cluster come next, as covered in the deployment timeline.

What it costs to build a data center for GPUs

Published averages for the building, before any IT equipment:

MeasureFigureSource
Global average construction cost, shell and core, 2026 forecastUS$11.3M per MWJLL (opens in a new tab)
US average construction cost, typical 60 MW designUS$12.08 per wattTurner & Townsend (opens in a new tab)
Most advanced AI-specific facilities, USUS$14 to US$16 per wattTurner & Townsend (opens in a new tab)
Premium for AI-ready facilities over traditional ones, USAbout 10%Turner & Townsend (opens in a new tab)
US greenfield, most modern facilitiesUS$17.6M per MWCushman & Wakefield (opens in a new tab)
AI tech fit-outUp to US$25M per MWJLL (opens in a new tab)

Applied to NVIDIA's 1.2 MW per scalable unit of eight NVL72 racks, those averages give the building costs below. This is our arithmetic, and it leaves out land, utility charges, GPUs, network and storage.

Rate per MW of IT loadOne unit, 1.2 MW (8 NVL72 racks)Eight units, 9.6 MW (64 NVL72 racks)
US$11.3M, JLL shell and coreUS$13.6MUS$108.5M
US$12.08M, Turner & Townsend US averageUS$14.5MUS$116.0M
US$14M to US$16M, Turner & Townsend AI-specificUS$16.8M to US$19.2MUS$134.4M to US$153.6M
US$17.6M, Cushman & Wakefield greenfieldUS$21.1MUS$169.0M

The averages come from large projects, so in our view a hall of a few megawatts will cost more per MW, because its fixed costs are spread over less load.

For GPU clusters, the IT usually costs more than the building around it. Our model in buy vs rent a GPU cluster puts a 32-node HGX B300 cluster at about US$19 million to US$24 million of capex for about 415 kW of IT load. At the Turner & Townsend US average, the building for that load is about US$5 million (our arithmetic).

Costs are rising. Turner & Townsend has average cost per watt up 13% year on year in North America and Europe (Turner & Townsend (opens in a new tab)), and Cushman & Wakefield reports development costs up 21% per MW since its previous guide (Cushman & Wakefield (opens in a new tab)).

Who does what on a data center build

RoleWhat they doWhat they hand over
Developer or ownerFinds land and power, gets approvals, raises capital, signs the utility and construction contractsA site with power committed
UtilityStudies the load, builds the connection, sets the tariffEnergized service at the agreed capacity
Architect and MEP engineersDesign the building and its mechanical, electrical and plumbing systemsA basis of design and stamped drawings
EPC or design-build contractorDesigns, buys the equipment and builds, under one contractA commissioned facility
Commissioning agentTests every system against the owner's requirements, for the ownerCommissioning reports and test results
Cluster integratorRacks, cables and configures the GPU systems and fabrics, then burns them inAn accepted cluster
OperatorRuns the facility, the cluster or bothService levels against the contract

EPC (engineering, procurement and construction) comes from energy and industrial projects: one contractor takes all three and hands over a working plant. In US building, the closer term is design-build, where "the single contract for both design and construction is the fundamental difference" from other methods (DBIA (opens in a new tab)). Under design-bid-build, the owner hires a designer, then bids the construction to a general contractor.

The labels matter less than three questions. Who carries the risk if the hall can't cool the racks? Who buys and expedites the long-lead equipment? Who fixes a failed integrated systems test? Put the answers in the contracts.

Where the building ends and the cluster begins

The building and the cluster are usually bought from different parties, and with liquid cooling the line between them runs through pipes. Agree each interface in writing before either side orders equipment.

InterfaceBuilding sideCluster sideAgree in writing
PowerUtility service, transformers, switchgear, UPS, generators, busway to each rack positionPower shelves or PDUs and the busbar inside the rackPower per position, redundancy, connectors
Liquid coolingHeat rejection, facility water loop, in-row CDUs, secondary piping to the rackRack manifolds, cold plates and in-rack CDUsWho owns each CDU, water temperature and flow, water quality, leak response
NetworkFiber entries, meet-me room, cable traysSwitches, optics, cabling and fabric configurationWho installs structured cabling, and when
MonitoringBuilding and electrical monitoring systemsGPU, node and fabric telemetryWhich alarms cross the line, and who responds
HandoverIntegrated systems test passedRack installation, burn-in and acceptanceThe date the hall is ready for racks

The handover date is the one finance should watch. Hardware that arrives before the hall is ready sits on the dock earning nothing, and a hall that's ready before the hardware sits empty. Tie the hardware ship dates to the commissioning schedule, with a margin on the cluster side.

Building with Lightyard

Lightyard designs, sources, builds and operates NVIDIA GPU clusters on its own software, for neoclouds, asset owners, investors and enterprises. We work on the cluster side of the line above, and on the site work a cluster needs. The cluster can sit in a Tier III colocation facility, in whichever region you need, or in your own data center. It can also go in a powered building or mining site we upgrade for it, or in a prefabricated module on land with power. When you don't have a site, finding one is part of the work.

The cluster is designed against the site before any hardware is ordered: power per rack, cooling and floor loading first, then the network (RoCE v2 Ethernet, spine-leaf), built for the size you plan to reach. For a large cluster, the work runs in monthly waves of 32 nodes. Then your team runs it, or we operate it for you. See GPU data centers for the four site types and GPU cluster deployment for what a deployment covers, or send us your site and power numbers.

Questions this note answers

Sources

  1. NVIDIA DGX SuperPOD reference architecture, DGX GB300 (opens in a new tab)
  2. NVIDIA DGX GB rack scale systems user guide: hardware (opens in a new tab)
  3. NVIDIA DGX B300 user guide: system specifications (opens in a new tab)
  4. NVIDIA technical blog: 800 VDC power architecture (May 2025) (opens in a new tab)
  5. Uptime Institute Global Data Center Survey 2026 (opens in a new tab)
  6. Uptime Institute: the Tier Classification System explained (opens in a new tab)
  7. JLL 2026 Global Data Center Outlook (opens in a new tab)
  8. JLL North America Data Center Report, Midyear 2026 (opens in a new tab)
  9. CBRE North America Data Center Trends H1 2026 (opens in a new tab)
  10. Cushman & Wakefield 2026 Data Center Development Cost Guide (opens in a new tab)
  11. Turner & Townsend Data Centre Construction Cost Index 2026-27 (October 2026) (opens in a new tab)
  12. Turner & Townsend 2026 US index: AI data centers (opens in a new tab)
  13. Turner & Townsend 2026 US index: industry challenges (opens in a new tab)
  14. Wood Mackenzie: supply-chain challenges in electric T&D, lead times (October 2025) (opens in a new tab)
  15. AEP Ohio data center tariff (opens in a new tab)
  16. US EPA fact sheet: use of backup generators to maintain grid reliability (May 2025) (opens in a new tab)
  17. ASHRAE Guideline 0-2019, addendum a: the commissioning process (opens in a new tab)
  18. Design-Build Institute of America: what is design-build (opens in a new tab)
  19. Fortune: Jensen Huang on US data center build times (December 2025) (opens in a new tab)
Get started

Tell us where you're starting.

Capital, hardware, a site or a team: tell us what you have and when you need it running. We usually reply within a few hours and come back with a written plan.

sales@lightyard.ai