Deployment

How long to build an AI data center in 2026

How long it takes to stand up an AI cluster yourself in 2026: two to four years for a greenfield data center, roughly 6 to 18 months in leased colo space, and what fills each month.

All notesPublished 10 min read
In this note
  1. AI data center timeline by route
  2. Why power takes the longest
  3. How long data center construction takes
  4. How long a leased colo deployment takes
  5. GPU and network hardware lead times
  6. How long burn-in and acceptance take
  7. What the timeline costs: numbers for the memo
  8. Which route fits

Building an AI data center from scratch in the US takes about two to four years before the first production job runs, and most of that time goes to getting power, not to construction. NVIDIA's Jensen Huang put it at "probably about three years" from breaking ground to standing up an AI supercomputer (Fortune (opens in a new tab)). Leasing liquid-cooled colo space and installing your own racks takes roughly 6 to 18 months, if you can find the space. Reserving capacity from a provider moves the date to whatever the provider commits to in the contract, counted from signing.

The timelines below assume a cluster of eight GB300 NVL72 racks, 576 GPUs. That's one NVIDIA DGX SuperPOD scalable unit, which NVIDIA sizes at 1.2 MW of thermal design power (NVIDIA reference architecture (opens in a new tab)). Bigger clusters take longer at every step.

AI data center timeline by route

These durations are our planning estimates, built from the public figures cited further down. They assume a US site and a team that has done this before.

PhaseGreenfield data centerLeased colo spaceReserved capacity
Find power and a site6 to 18 months to land a site and a utility commitment; the grid connection itself can take 4+ years in primary markets1 to 3 months to search; new halls signed today mostly deliver in 2028None: you choose a region
Design and permits6 to 12 months1 to 3 months for the fit-out designNone
Construction or fit-out12 to 24 months for shell, electrical and cooling0 to 2 months if the hall is liquid-ready, 4 to 9 months if notNone
Hardware order to deliveryIn parallel, on the OEM's ship dateIn parallel, on the OEM's ship dateNone: dates are set from signing
Install, burn-in, acceptance1 to 3 months1 to 3 monthsYour own acceptance testing, usually a few weeks
Total to first production job24 to 48 months6 to 18 months (longer if you wait for a new hall)about 10 to 14 weeks from signing for HGX B300; the proposal sets the date for other platforms

The ranges are wide because power decides them. A site that already has a utility commitment skips most of the first row. One without it can sit there for years.

Why power takes the longest

JLL's 2026 Global Data Center Outlook (opens in a new tab) puts the average wait for a grid connection in primary data center markets at more than four years, and says speed to power is now the first thing operators pick sites on. You can see it in where the building is happening. In JLL's midyear 2026 North America report (opens in a new tab), 77% of the more than 66 GW under construction is in "frontier" markets such as West Texas, Ohio and Louisiana, places that had almost no data center capacity ten years ago.

At 1.2 MW the fastest option is a site where the utility has already committed capacity, usually an existing building or campus with headroom. It's the only way a greenfield project lands near the low end of the range. The slow option is a new utility connection: a load study, an interconnection agreement and often a new substation, which is where JLL's four-year average comes from. Some projects bridge the gap with on-site gas turbines or engines plus batteries until the grid connection arrives. That works, but it brings its own air permits and fuel contracts.

Long-lead electrical gear belongs in this phase too, because it has to be ordered before construction starts. Wood Mackenzie reported in April 2024 that average transformer lead times had gone from about 50 weeks in 2021 to about 120 weeks, with substation and generator step-up units at 80 to 210 weeks (Wood Mackenzie (opens in a new tab)). In August 2025 it projected a 30% supply deficit for power transformers that year (Wood Mackenzie (opens in a new tab)). If your site needs a new substation transformer, that one order can take a year and a half to four years.

How long data center construction takes

Once power is committed, a purpose-built hall typically needs 6 to 12 months of design and permitting, then 12 to 24 months of construction and commissioning. Permitting is the less predictable part. Zoning, air permits for backup generators and water supply can each add months.

The record-setting projects are useful mainly for what they had in place before they started. NVIDIA says xAI's Colossus in Memphis, facility and 100,000 Hopper GPUs together, was built in 122 days, with 19 days from the first rack rolling onto the floor to the start of training (NVIDIA (opens in a new tab)). It went into an existing building, a former factory, and ran on temporary gas turbines and batteries while the grid connection caught up (Wikipedia (opens in a new tab))). The turbines later drew a notice of intent to sue under the Clean Air Act and needed permits of their own.

One developer started phase 1 of a 1.2 GW campus in Abilene, Texas, two buildings and more than 200 MW, in June 2024 and expected to energize it in the first half of 2025. By May 2025 about 3,000 people were working on the site every day (the developer (opens in a new tab)).

Both had power lined up before or alongside construction, a building or a prepared site, and the money to run trades around the clock. An enterprise putting up a 1 to 5 MW hall rarely has all of that, so Huang's three years is the better number to plan on.

How long a leased colo deployment takes

Leasing skips the building but not the shortage. JLL has North American vacancy at 1% for the third year running, with available capacity "limited to small, fragmented blocks" and most tenants signing today contracting for 2028 deliveries (JLL (opens in a new tab)). At 1.2 MW you're too small to get priority on a new hall, but small enough to fit into one of those fragmented blocks.

The harder question is whether the space can take the racks. GB300 NVL72 uses hybrid cooling, with the GPUs and CPUs on direct liquid cooling and everything else on air (NVIDIA (opens in a new tab)). At 1.2 MW per eight-rack unit, that's about 150 kW per compute rack once its share of networking is counted, far more than most existing colo space was designed for. NVIDIA also recommends a facility built to Uptime Institute Tier 3 or an equivalent standard, with concurrent maintainability and no single point of failure.

A hall sold as "liquid-ready" can still be missing some of this:

  • Coolant distribution units (CDUs), in-row or in-rack, sized for the full heat load plus redundancy
  • A facility water loop to the white space at the supply temperature and flow the CDUs need, with dry coolers or chillers to reject the heat
  • Secondary piping and manifolds to each rack position, with isolation valves and leak detection
  • Power distribution well above 100 kW per position, usually busway rather than whips
  • Floor loading, clearances and a delivery path from the dock for heavy pre-integrated racks

If everything is there, fit-out takes weeks. If the hall needs CDUs and secondary piping, our estimate is 4 to 9 months, set mostly by equipment lead times and the colo's own change process. If it needs a new water plant or more utility power, you're back on a construction schedule.

Month-by-month plan for a leased colo deployment

This is a planning template for eight GB300 NVL72 racks going into existing space that needs a liquid-cooling fit-out. Hardware and fit-out run in parallel, and the critical path is whichever finishes later.

MonthWork
0 to 1Size the cluster and its power budget; send RFPs to colos and OEMs; start internal capital approval
1 to 3Shortlist colos with liquid-cooled capacity; site visits; sign the lease; place the hardware order and get a written ship date
2 to 4Fit-out design with the colo: CDU count, loop temperatures, power per position, fabric layout, cable plant
3 to 9Fit-out: CDUs, secondary piping, busway, leak detection, commissioning the cooling loop under load
6 to 9Fabric and storage delivered; structured cabling and switches installed ahead of the racks
7 to 10Racks delivered and set; power, coolant and fabric connected; firmware and OS brought to a baseline
8 to 11Burn-in, part replacement and acceptance testing
10 to 12First production jobs, with a few weeks of elevated failure rates as the last weak parts show up

A slip in the colo's cooling commissioning moves everything behind it. So does hardware that arrives with nowhere to go. Racks sitting on a dock earn nothing, and their market value keeps falling before the depreciation clock has even started.

GPU and network hardware lead times

GPU systems usually aren't the longest item any more, but they're the least predictable one. Delivery dates for current-generation racks depend on allocation, so if you're buying a few racks, plan on the OEM's written ship date and ask before signing what happens if it slips.

NVL72 racks arrive integrated at the rack level, which NVIDIA says allows "rapid installation and deployment" of liquid-cooled, high-density racks (NVIDIA (opens in a new tab)). Setting a rack and connecting power, coolant and fabric should take days per rack. What tends to run late is what goes around the racks: switches, optics, the storage system and the management network.

There's also the product cycle. NVIDIA has said it ships new platforms on a "one-year rhythm" (NVIDIA (opens in a new tab)). On a three-year greenfield schedule, the generation you specify at the start will be one or two generations old when the hall opens, unless you order late and take the allocation risk that comes with it.

How long burn-in and acceptance take

New clusters fail early and often. When Imbue brought up 4,088 H100s, about 10% of the machines failed to boot, mostly on physical problems such as miswired cables, broken power supplies and bad drives. After a PCIe switch firmware update, around a quarter of the hosts needed their internal PCIe cables reseated (Imbue (opens in a new tab)). Meta's Llama 3 paper reports 466 job interruptions over a 54-day pre-training window on 16,384 H100s, and puts about 78% of the 419 unexpected ones down to confirmed or suspected hardware issues (Meta, Llama 3 paper (opens in a new tab)).

Burn-in is how you find those parts before your first real job does. A typical sequence:

  1. Node and rack health with NVIDIA's DCGM diagnostics at level 4, which covers GPU memory, PCIe and NVLink, targeted stress and power tests, and NCCL tests. It runs in under 2.25 hours on an 8-GPU system (NVIDIA DCGM (opens in a new tab)).
  2. Fabric checks: per-port bandwidth and error counters on every link, then nccl-tests (opens in a new tab) all-reduce and all-to-all across racks. Imbue's multi-node NCCL runs took 12 to 24 hours to surface problems.
  3. Full-system load: HPL or a representative training job across the whole cluster for one to several days, to catch thermal, power and straggler problems that only appear at full load.
  4. Replace failing GPUs, trays, optics and cables, then rerun the tests on the affected racks.
  5. Acceptance against a written test plan with pass criteria (bandwidth, all-reduce throughput, job completion over a fixed window), signed off before the cluster goes into service.

The tests take hours. The calendar time goes to waiting for replacement parts and rerunning, so two to six weeks is a reasonable plan for a cluster this size.

What the timeline costs: numbers for the memo

For a finance reader, the cost of the timeline is capital sitting idle while it runs. The building figures below are published averages. The 1.2 MW line is our arithmetic and leaves out the GPUs, network and storage.

ItemFigureSource
Average global construction cost, 2026 forecastUS$11.3M per MWJLL (opens in a new tab)
US greenfield development cost, most modern facilitiesUS$17.6M per MWCushman & Wakefield (opens in a new tab)
AI tech fit-out, on top of the buildingUp to US$25M per MWJLL (opens in a new tab)
Facility for one 1.2 MW scalable unit (estimate)About US$13.6M to US$21.1M before IT1.2 MW times the two averages above
Greenfield time to first job (estimate)24 to 48 monthsThis post
Leased colo time to first job (estimate)6 to 18 monthsThis post

On a greenfield project, most of the facility spend lands a year or more before the first GPU earns anything. Hardware bought early waits for power and space, and every month it waits comes out of its earning life before the next generation ships.

Schedule risk is the other line item. A slip in power or cooling commissioning delays everything behind it, and whoever owns the facility carries that risk.

Which route fits

Greenfield makes sense when you need tens of megawatts for a decade, already have a powered site or the patience to get one, and want to own the building. Leased colo fits when you need a few megawatts within a year and have a team that can run fit-out, procurement, burn-in and operations. Reserving capacity fits when the date matters more than owning the hardware, or when you need the cluster for a multi-year term but not for the life of a building.

If that last case is yours, Lightyard reserves single-tenant GB300 NVL72 racks and HGX B300 clusters in the region you choose, with dates counted from signing and the cluster run to an SLA. HGX B300 is about 10 to 14 weeks from signing. See how delivery works or send us your requirements.

Questions this note answers

Sources

  1. Fortune: Jensen Huang on US data center build times (Dec 2025) (opens in a new tab)
  2. NVIDIA: xAI Colossus built in 122 days (Oct 2024) (opens in a new tab)
  3. Wikipedia: Colossus (supercomputer), site and power history (opens in a new tab)
  4. JLL North America Data Center Report, Midyear 2026 (opens in a new tab)
  5. JLL 2026 Global Data Center Outlook (opens in a new tab)
  6. Cushman & Wakefield 2026 Data Center Development Cost Guide (opens in a new tab)
  7. Wood Mackenzie: power transformer lead times (Apr 2024) (opens in a new tab)
  8. Wood Mackenzie: transformer supply deficits in 2025 (Aug 2025) (opens in a new tab)
  9. Abilene campus joint venture and phases (May 2025) (opens in a new tab)
  10. NVIDIA DGX SuperPOD reference architecture, DGX GB300 (opens in a new tab)
  11. NVIDIA DCGM diagnostics user guide (opens in a new tab)
  12. NVIDIA nccl-tests (opens in a new tab)
  13. Imbue: from bare metal to a 70B model (infrastructure set-up) (opens in a new tab)
  14. Meta: The Llama 3 Herd of Models, reliability section (opens in a new tab)
  15. NVIDIA: Computex 2024 keynote, one-year rhythm (opens in a new tab)
Get started

Start with one rack.

Tell us the size and timing you have in mind. We usually reply within a few hours and come back with a written proposal.

sales@lightyard.ai