The Definitive Guide toAI Data Centers
Ask the GuideAboutAccount

Chapter 6.6

In this chapter · 7 sections
Term help

Construction Execution, Sequencing & Phased Turnover

An AI data center earns nothing until megawatts are energized and accepted; the schedule that matters delivers turnover-sized blocks against a fast-depreciating GPU life, sequenced around civil work, equipment, utility power and acceptance evidence, whichever controls the finish.

POWER-BOUNDDENSITY-RAMP

What you'll decide here

  1. Whether you procure the build as a hard-bid lump sum, a CM-at-risk GMP, or an integrated design-build/EPC — and therefore who carries schedule and trade-coordination risk on a project where the long poles are owner-furnished electrical gear and a labor pool that is structurally short; name retained OFE custody and remedies under Chapter 2.4.
  2. The turnover block size — whole building, hall, or pod — because it sets how early you can energize and earn, and how much commissioning you must interleave with live, occupied space next door.
  3. Where the critical path actually runs: civil and shell readiness can control alongside the long-lead switchgear and transformers, the utility energization date, and the availability of qualified electricians and pipefitters during the MEP rough-in and equipment-set windows.
  4. How you weave Cx Level 1–5 into the construction schedule, with factory tests, surrogate and liquid-cooled/AI-emulating loads, injection/HIL, isolated scopes, and a staged product-representative workload assigned to the evidence each can validly prove.
  5. Your skilled-trades mitigation strategy — prefabrication, multi-shift work, travel premiums, and self-perform vs subcontract — decided at GC selection from package work hours and available qualified crew hours, before shortages emerge at peak manpower.

The design is frozen, the slab basis is set, the cooling plant is specified, the fire strategy approved — and none of it earns a dollar until megawatts are energized, accepted, and handed to the cluster team. The interval that decides whether that happens on time is the eighteen-to-thirty-month stretch between a signed GMP and a GPU drawing its first watt. Steel, slab, and skin are schedulable packages, but weather, ground conditions, inspections and access can put them on the critical path. What kills AI-data-center schedules is a short and repeating list — long-lead electrical gear that does not show up, a utility energization date the owner does not control, a skilled-trades labor pool short by hundreds of thousands of workers nationally, and, as of 2026, organized local opposition: Data Center Watch counted ~$130B of US projects blocked or delayed in Q1 2026 alone. That is a tracker's project-value sum, and "delayed" is not "dead" — but a 16-month modular schedule is fiction if rezoning takes 18 (→ Chapter 3.11). Pennsylvania made consent a permit gate outright: under Executive Order 2026-05 (2026-08-18), AI data-center proposals leave Fast Track and cannot use NDAs; DEP may review compliant applications from projects whose developers execute a GRID Consent Order and Agreement (COA) as they arrive but cannot issue permits before all required local approvals, while non-GRID projects do not enter DEP review until those approvals are secured and every required construction-permit application has been found compliant. Execution is the work of sequencing around those constraints so that the parts you can control are never what you are waiting on.

The delivery-model fork — lump-sum vs CM-at-risk vs design-build/EPC — decides who eats schedule risk. The canonical sequence (civil → shell → MEP rough-in → equipment set → fit-out) has predictable bottlenecks. Phased turnover is a revenue lever: the unit of turnover you choose determines how early the first block earns against the contested 2–3-year accelerated-life stress case. And the trades shortage is a first-order schedule risk to mitigate at GC selection, alongside the quality and inspection regime that carries a hall from substantial completion into commissioning readiness. The skilled-workforce program — recruiting, training, retention — lives in Chapter 14.11; here it is a construction-schedule input, and the owners have started paying to manufacture it: Meta committed $115M to train ~5,000 construction workers in year one, Google $50M into IBEW apprenticeships (annual intake 19,500 → 30,000 target), BlackRock $100M — $265M combined (NYT, Jul 2026) — against data-center trades pay running ~42% above comparable work. Training dollars are a mitigation, not a closed gap: the national shortfall figure has not moved. The construction-safety program lives in Chapter 6.9; here it is an interface, not the subject.

The delivery-model fork: who carries schedule risk

Before a shovel moves, the owner picks how the build is contracted, and that single choice allocates the two risks that dominate an AI build: schedule and trade coordination. The fork is rarely about lowest first cost — at $12–13B of annual revenue per energized gigawatt, the value of getting 200 MW online six months early dwarfs the contractor's fee delta. It is about who is on the hook when the switchgear slips and the electricians are double-booked across three concurrent campuses.

Lump-sum / hard-bid transfers price risk to the GC but assumes a complete, frozen design at bid time — an assumption that is almost never true for a fast-tracked AI build where the rack generation is still moving under the design. Change orders become the battleground, and the owner discovers that a fixed price bought on incomplete documents is not actually fixed. Construction-management-at-risk (CMAR) with a guaranteed maximum price (GMP) is the workhorse model for hyperscale: the GC joins during design, prices a GMP against a still-developing set, and included-scope overruns follow the agreed GMP terms while exclusions, allowances, owner changes and any savings share remain contract-specific. It buys early constructability input and preserves the ability to start civil while the fit-out design matures. Integrated design-build / EPC collapses designer and builder into one contract and one schedule — the fastest path to a single throat to choke, favored when speed-to-power dominates and the owner will trade some price transparency for it. Choose wrong — a hard-bid contract on a moving design — and the change-order war costs more schedule than the fixed price ever saved on fee.

Build-delivery model → risk allocation and fit
Delivery modelDesign completeness at startWho carries schedule riskTrade-coordination ownerBest fit
Lump-sum / hard-bidMust be ~complete & frozenGC (in theory); owner via change orders (in practice)GCStable, repeatable, fully-designed shells
CMAR + GMPDeveloping; priced at GMP setShared above/below GMP; owner on OFEGC, with early owner inputMost hyperscale AI builds; fast-track with cost control
Design-build / EPCPerformance-spec; design proceeds in parallelSingle entity (designer+builder)EPC integratorSpeed-to-power-dominated; single-throat accountability
Owner-led multi-primeVaries by packageOwner (the integrator role)Owner / owner's CMSophisticated self-perform owners; max control, max risk
Practitioner framing for AI/hyperscale builds, 2026. "Owner-furnished equipment" (OFE) — switchgear, transformers, CDUs, GPUs procured directly by the owner to protect lead time — shifts schedule risk back to the owner under every model.

The construction sequence and where it bottlenecks

The canonical sequence is linear on paper and overlapped in reality. Civil and earthwork — mass grading, stormwater, deep utilities, foundations — is weather-exposed and front-loaded. Shell — structural steel or tilt-up, roof, skin — gets the building watertight so interior trades can work year-round. MEP rough-in — the electrical raceway, bus duct, mechanical piping, the liquid-cooling distribution backbone — is the labor-intensive heart of the schedule and the first place the trades shortage bites. Equipment set — landing the switchgear, UPS/BESS, CDUs, chillers, generators, and the heavy mechanical plant — is rigging-and-crane choreography gated by when the owner-furnished gear arrives. Fit-out — busway runs to the rack, manifold and quick-disconnect plumbing, containment, terminations, and the thousand small completions — is where substantial completion is won or lost.

What separates a 16-month delivery from a 26-month one is how you overlap the phases — and the overlap is bounded by the critical path, which can run through the concrete when civil readiness slips past equipment delivery. Civil and shell durations depend on weather, ground, curing, inspections and available crews; money cannot remove every predecessor. Other binding constraints sit downstream and outside the GC's direct control: the date the utility energizes the service, the date the owner-furnished switchgear and transformers land, and the headcount of qualified electricians and pipefitters available during rough-in and fit-out. Sequence to keep those three off the path of the things you control, and you protect the schedule. Let the trades become the constraint at peak manpower and the whole back end of the project slides — fit-out and commissioning are exactly where the shortage concentrates.

The cooling package changes critical path when civil readiness slips

22 weeks · assumedmodeled
Cooling package: assumed accepted-service milestone measured from release on the shared calendar
Scope & caveats

Assumed cooling-package sequence on one calendar from release. A separate self-powered rig flushes isolated piping; installed CDU startup and tests await utility. Chapter 6.6’s worked case states every input, civil/utility crossover and release owner; Chapter 2.1 owns the network method. No supplier duration range is established.

Use the network from Chapter 2.1. Installation finishes at max(12, 14) + 3 = week 17. The independent rig completes fluid acceptance at 17 + 2 = week 19. Energized startup finishes at max(19, 18) + 1 = week 20; functional and integrated acceptance then finish at 20 + 2 = week 22. Accepted service is max(22, 20) = week 22. Delivery is the controlling predecessor to installation; civil has two weeks before it reaches that gate. Select the baseline package sequence and protect its cleaning and test windows.

The civil flip is explicit. Site readiness at week 18 gives installation at 21, readiness at 24 and accepted service at week 26. Civil becomes controlling above week 14, so paying to expedite an already-delivered CDU does not recover this slip. In the utility flip, utility at week 24 delays startup until then: max(19, 24) + 1 = week 25, followed by two test weeks, gives accepted service at week 27. Utility becomes controlling above week 19, when fluid acceptance is otherwise ready. Fix the delayed gate, not the test duration. This is why the concrete can own the finish even on a project famous for electrical lead times.

Release each step from its actual record: the GC confirms slab/access and supports; the installing contractor closes joints and installation inspection; the cooling specialist supplies the flush and fluid-quality evidence; the commissioning authority accepts the test scope; the authority grants the required use/occupancy permission. The owner accepts the block only after those gates close. GAO supplies the schedule method, Chapter 2.1 owns the integrated network, Chapter 5.13 owns hydraulic qualification, and Chapter 13.5 owns cooling acceptance.

Deep dive: the liquid-cooling backbone changes the sequence — pipe before you can pull rack

A traditional air-cooled hall could defer most of its mechanical scope to fit-out. A direct-to-chip liquid hall cannot. The facility-water and technology-cooling-loop distribution — risers, headers, the CDU gallery, in-rack manifolds, and the ~150–200 quick-disconnects per rack — is heavy, code-governed charged piping (ASME B31.x / EN 13480) whose material and jointing method determine installation, weld/NDE scope, flushing and pressure testing before the racks it serves can be energized. That work lands squarely in the MEP rough-in and fit-out windows and competes for the same pipefitter and welder labor that is already short. It also introduces a hard sequencing dependency the air world never had: multi-stage flushing and fluid-quality acceptance (with a duration derived from loop volume, method, acceptance samples and retest allowance) gate the cooling acceptance that gates the GPU burn-in.

In a liquid hall, then, the cooling distribution sits on the critical path to first-watt in a way it never did for air. Operators that prefabricate skidded CDU and manifold assemblies off-site (→ Chapter 6.4) pull this scope out of the congested on-site fit-out window and into a controlled factory environment — as much a labor-availability strategy as a quality one. The piping mechanical engineering itself is owned in Chapter 5.4 (DLC) and its pressure-system code basis in the facility-piping chapter; here it is a schedule dependency that reorders the sequence.

Phased turnover: the unit you energize is the unit you earn on

The biggest schedule gain in the whole construction phase comes from not waiting for the whole building. AI campuses are built as repeated, near-identical units — a baseline pod or hall design stamped out across phases — precisely so that the first block can be energized, commissioned, and producing while later blocks are still being poured. Multi-building hyperscale campuses now hit 18–24-month physical build times — a single hall or colo-scale build runs 12–18 — specifically through phased delivery (Data Center Knowledge / Mastt, 2026). The decision is what size the turnover block should be: whole building, data hall, or pod/row. Smaller blocks energize earlier and start the revenue clock sooner, but they multiply the number of times you must commission and partially energize next to occupied, live space — which brings concurrent-maintainability constraints, arc-flash boundaries around energized gear, and life-safety separation between a construction zone and a running hall into the schedule as hard interfaces.

The economics are what make phasing dominate. At roughly $12–13B of revenue per gigawatt per year (SemiAnalysis, Jun 2026), bringing a 200 MW block online six months early is on the order of $1.2–1.3B of incremental revenue — against a GPU asset whose useful economic life is tested at 2–3 years in the guide's contested bear case, versus published 4–6-year estimates and 5–6-year book policies. Every month a finished, un-energized hall sits dark burns a month of that short asset life at zero return. So the schedule gets measured in energized, accepted, GPU-ready megawatts per quarter, not in percent-complete.

Commissioning woven into the schedule, not bolted on at the end

The most expensive commissioning mistake is treating Cx as a phase that begins only when construction ends. The Cx Level ladder (L1–L5) is interleaved through the entire build, and construction execution either creates the evidence trail as it goes or pays to reconstruct it later. Level 1 (factory witness / FAT) happens at the manufacturer before equipment ships — witnessing switchgear, UPS, and CDU testing in the factory is a construction-schedule activity, and skipping it imports defects to the critical-path fit-out window where they cost ten times as much to fix. Level 2 (site acceptance / installation verification) is the pre-energization inspection work that overlaps directly with equipment set and fit-out. Level 3 (pre-functional, one system standalone and energized) and Level 4 (functional performance across the operating envelope, per discipline) — energization, load-bank acceptance, redundancy-topology validation — begin the moment a block's gear is set and its service is energized, well before the building as a whole is done. Level 5 (integrated systems testing, IST) is the approved integrated-scenario evidence gate before handover.

Build the commissioning windows into the construction logic as named, duration-bearing activities, and protect them. Facility air-rejecting load banks do not cover every TCS or workload dynamic. Liquid-cooled and AI-emulating test loads prove declared surrogate envelopes, while a staged product-representative workload closes the remaining named normal-operation evidence. Protective functions stay in injection/HIL or an isolated controlled scope unless the approved risk plan explicitly requires otherwise. The full Cx program — levels, scripts, acceptance gates, and the load-realism problem — is owned in Chapter 13.1 and the cooling-acceptance specifics in Chapter 13.5; the construction job is to make the schedule honor them.

The skilled-trades shortage as a first-order schedule risk

Of all the constraints on an AI build, the one most likely to be underestimated at GMP signing and most painful at peak manpower is labor. The numbers are structural, not cyclical. ABC estimated that the US construction industry as a whole needed to attract 349,000 net new workers in 2026 to meet demand (a one-year net-hiring figure — distinct from the ~499k cumulative data-center construction shortfall projected on this page); the electrical trade alone needs hundreds of thousands more bodies, and IEC's 2023 Emerging Leaders survey of member contractors put retirements at ~10,000 a year against ~7,000 new entrants (IEC, 2023) — a survey-derived industry picture rather than a national entry/exit census, so read it as direction and not as a deficit you can compound year on year. Electricians, MEP engineers, commissioning specialists, and project managers are consistently named the hardest roles to fill. And because electrical systems are 45–70% of total data-center construction cost (iRecruit, 2025), the shortage hits the most cost-and-schedule-dense scope in the project precisely during the MEP rough-in and fit-out windows that already sit on the critical path.

Two models price the delay differently. The cited ~$14M per month source estimate for a delayed 60 MW block belongs to a colocation-rent model, implying ~$233/kW-month in 2026. At ~$12–13B/GW/year of AI-capacity revenue, the same block represents roughly $60–65M per month in delayed revenue. The trades shortage is therefore a site-selection constraint — operators choose markets partly on the depth of the local skilled-labor pool, because a site you cannot staff is a site you cannot energize on schedule. Mitigation belongs at GC selection, not in a back-office HR plan. The strongest levers, in rough order of impact:

  • Prefabrication and modularization (→ Chapter 6.4) — shifting skidded power and cooling assembly into a factory cuts on-site labor hours 30–50% and moves the work to a controlled labor pool, away from the congested site. This is the single largest labor mitigation available.
  • Self-perform vs subcontract — GCs that self-perform electrical and mechanical control their own crews instead of competing for them on the spot market; the delivery-model choice and the labor strategy are coupled.
  • Multi-shift and travel-crew premiums — running second shifts and importing traveling trades closes the local gap at a cost premium (data-center work already commands a wage premium over general construction — reported at ~30% in trade surveys and ~42% above comparable work by mid-2026).
  • Early trade-partner lock-in and stacked starts — committing subcontractors during design and staggering trade peaks across a phased campus so the same crews flow from block to block rather than all peaking at once.

The owner-side workforce program — apprenticeships, pipeline development, retention — is the canonical subject of Chapter 14.11. The construction-execution point is narrower and sharper: if the trades strategy is not decided at GC selection, it will be discovered at peak manpower, and by then the only levers left are the expensive ones.

18–24 mo
physical build time for a hyperscale campus (construction only; grid energization is separate — a single hall or colo-scale build runs 12–18 mo)
~499kforecast
projected 2026 US data-center construction worker shortfall; context, not a project crew plan
~10k vs ~7kestimate
electricians retiring vs entering the US trade per year — IEC 2023 member-survey observation, not a national entry/exit census
Scope & caveats

Survey-derived observation from the IEC 2023 Emerging Leaders report's 'current state of industry' findings (Independent Electrical Contractors member survey, US): 7,000 new electricians join each year while 10,000 retire. Not a national entry/exit census, and the entrant, completion, retirement and exit definitions do not match — do not compound it into a running national deficit.

45–70%
electrical systems as a share of total data-center construction cost
~$14M/moderived
source-modeled colocation rent delayed by a 60 MW schedule slip — distinct from AI-capacity revenue
Scope & caveats

colocation-rent economics for 60 MW of delayed capacity

Colocation-rent model (~$233/kW-month in 2026), not AI-capacity revenue — at ~$12–13B/GW/yr the same 60 MW block is ~$60–65M/month; the chapter works both models.

30–50%
on-site labor and schedule reduction from prefabrication / modular construction (McKinsey: up to ~40% less on-site manpower; 30–50% schedule compression)
Scope & caveats

prefab labor reduction

~128–208 wk
standard load-serving power-transformer lead time (to ~60 months in constrained markets) — the OFE long pole
Scope & caveats

Load-serving substation/power transformers only. Generator step-up (GSU) transformers are carried as a separate register entry (~144–208 wk). The upper bound comes from large-unit and constrained-market quotes, not from GSU indices.

Indices diverge in mid-2026 for large power transformers generally; the GSU-specific divergence (VAWN 144 wk vs SemiAnalysis 3–4 yr) is recorded on the GSU claim.

~$12–13B/GW/yrestimate
AI revenue per energized gigawatt — why six months early is worth ~$1.2–1.3B gross on 200 MW (contested — single-source)
Scope & caveats

This is the rental/IaaS denominator (SemiAnalysis, contested). Distinct and much larger is the lab token-revenue side: SemiAnalysis's Tokenomics model (Aug 2026) puts OpenAI/Anthropic API inference at >$100B/GW/year on a GB300 cluster against ~$12B/GW/year of rental cost — a model-derived figure sensitive to utilization and price mix, not an audited disclosure. Do not conflate lab API revenue with IaaS rental in one number.

Quality, inspections, and the path to commissioning readiness

Between "the building is done" and "the cluster team can take it" sits a quality gauntlet that, run poorly, becomes the longest unplanned slip in the project. The path runs through inspections and authority-having-jurisdiction (AHJ) sign-offs — electrical, mechanical, fire, life-safety, occupancy — each of which can stop energization cold; through quality-control verification of the trades' work (weld NDE on charged piping, megger and torque verification on electrical terminations, fluid-quality acceptance on the cooling loops); and into substantial completion, the contractual milestone at which the owner can beneficially occupy. Substantial completion and commissioning readiness are separate gates whose order follows the contracted work and intended use. The gap between them is the punch list and the deficiency-tracking discipline that closes it.

The decision here is whether quality is verified continuously as the work is installed or audited at the end. Continuous in-line QA — using the digital Cx platform to capture installation verification and baseline "fingerprint" data as each subsystem is built (→ Chapter 13.2) — front-loads defect discovery into the construction window where it is cheap to fix. End-of-line auditing concentrates discovery into the handover window where every defect is on the critical path to first-watt and every fix competes with the same short trades for attention. A clean transition between contract completion and Cx readiness is won months earlier, by the QC regime written into the GC's execution plan, not by heroics at the finish. Run it sloppily and deficiencies surface during L4/L5 testing, where they read as commissioning failures, blow the IST window, and push the turnover block — and its revenue — into the next quarter.

Deep dive: substantial completion, the punch list, and why "done" is a defined term

"Done" is contractual, not intuitive. Substantial completion is the milestone at which the work is sufficiently complete that the owner can use the facility for its intended purpose — it typically triggers the start of warranty periods, the transfer of certain risk and insurance, and the release of much of the retainage. It does not mean every item is finished; it means the remaining items (the punch list) do not prevent beneficial use. Final completion comes later, when the punch list is closed.

For an AI build the seam that matters is the one between substantial completion and commissioning readiness — the point at which a block has enough verified, energized, accepted infrastructure for L3/L4/L5 testing to proceed. These are not the same milestone, and conflating them is a classic schedule trap: a hall can be "substantially complete" with a punch list that still contains items fatal to commissioning (an un-flushed cooling loop, an unproven protective-relay setting, an incomplete BMS point-list). The disciplined execution plan defines commissioning readiness as its own gate with its own checklist, sequenced against the contract package rather than by one universal order: on a shell-or-package contract, readiness can legitimately follow substantial completion, but where the accepted work is the operating hall itself an un-flushed loop defeats the intended use the milestone certifies. Name construction completion, ready-for-startup, ready-for-FPT/IST, contract-specific substantial completion and acceptance as separate gates, and put commissioning readiness before the cluster team's GPU burn-in. The commissioning program that consumes that readiness is owned in Chapter 13.1; the construction job is to deliver a block that is genuinely ready, not merely substantially complete.

Release the block whose installation, cleaning, tests and authority permissions have all closed. Expedite the gate controlling that date and protect the test windows; squeezing commissioning to preserve a calendar promise moves defects into the occupied hall.

Construction execution is the hinge between the building and the cluster. The long-lead equipment it sequences around is registered in Chapter 2.3; the structural and slab basis it builds to is in Chapter 6.2; prefabrication and its project-specific labor and schedule tradeoff is in Chapter 6.4; the fire and life-safety regime that gates partial-energization separation is in Chapter 6.5; and the construction-safety / arc-flash program it interfaces with is in Chapter 6.9. The liquid distribution that reorders the sequence is engineered in Chapter 5.4. The commissioning program it weaves in — levels, scripts, electrical and cooling acceptance, and Level-5 IST — is the whole of Chapter 13.1, Chapter 13.2, Chapter 13.5, and Chapter 13.6. And the workforce program behind the trades it depends on is in Chapter 14.11.
Cite this chapter
Fehn, J. (2026). Construction Execution, Sequencing & Phased Turnover (Chapter 6.6). The Definitive Guide to AI Data Centers. https://aidatacenterguide.com/part-6-the-building-civil-structural-fire-life-safety-and-construction-execution/6-6-construction-execution-sequencing-and-phased-turnover (accessed 2026-09-29).
@misc{aidc-6-6,
  author       = {Fehn, Jacob},
  title        = {Construction Execution, Sequencing & Phased Turnover (Chapter 6.6)},
  howpublished = {The Definitive Guide to AI Data Centers},
  year         = {2026},
  url          = {https://aidatacenterguide.com/part-6-the-building-civil-structural-fire-life-safety-and-construction-execution/6-6-construction-execution-sequencing-and-phased-turnover},
  note         = {Accessed 2026-09-29}
}
Spotted an error? Suggest an edit