LEVIATHAN SYSTEMS

Installation

Why Data Center Projects Take Time: Facility Readiness and GPU Deployment

Technical contentPublished Updated

Developing a hyperscale data-center facility can take much longer than installing its IT equipment because the project must deliver usable utility service, approved site and system designs, manufactured and integrated equipment, and an accepted facility outcome. These processes have their own dependencies and responsible parties. A prepared physical-deployment team can turn released work into installed equipment, cabling and usable records; its staffing cannot substitute for unfinished utility, approval or facility work.

Why developing the facility can take longer than installing the IT

Installing equipment in a released data hall begins with much of its supporting environment already decided. Developing that environment includes establishing how the site will receive power, what will be built, which equipment will provide the required functions, and how the owner will establish that the completed building meets its requirements. The shorter installation package can therefore sit near the end of a much larger development effort.

Elapsed time includes decisions, manufacturing and reviews as well as visible construction. Some activities can overlap; others need a specific earlier result before they can finish. Adding every department's estimate together exaggerates a sequence that may overlap, while counting only time spent mounting racks omits the development work. The useful comparison follows the actual outcomes needed for the particular phase.

The official examples below explain these mechanisms. They supply no national construction timetable. Once the facility dependencies are understood, the later sections show how to build the physical IT schedule around them.

References: [U1][A1][D1][C1]

Usable utility service requires more than a nearby power line

A large load may require infrastructure to be designed and constructed beyond the data-center building. In its Virginia data-center FAQ, Dominion Energy describes site-specific power plans and staged service as a facility's load grows. Its listed work includes infrastructure design, rights of way, permits, material procurement, possible regulatory filings and construction. It also says many steps can proceed concurrently.

The mechanism is straightforward: permission to use a route, an applicable infrastructure design and the equipment needed to build it are different inputs to delivered service. Adding installers inside the data hall does not produce those inputs. An initial supply stage and the eventual full-load arrangement can also support different project milestones.

This is one utility's process, not a promise for another territory or site. The relevant buyer question is which service stage supports the intended facility phase and what remains to deliver it. Use the related power-capacity guide for the detailed distinction between planned MW, technical delivery and customer allocation.

References: [U1]

Design and approval decisions define what can actually be built

A facility design must turn the intended equipment and operating requirements into a project that its appointed designers, owner and applicable authorities can review. Decisions about equipment and interfaces affect the information used by other disciplines. An unresolved choice can leave dependent drawings or supplier information unfinished; a later change can require the affected design to be reconsidered.

Loudoun County, Virginia, provides a concrete public-authority example. Its Building and Development department separately lists construction and site-plan review, permit issuance, construction inspections and final approval for occupancy or use. Those are different outcomes. A permit record does not by itself establish that the constructed work has received the later applicable approvals.

The actual jurisdiction and project determine which decisions apply. For a business schedule, identify the unresolved design or authority outcome and the phase it affects. A generic line labelled permits complete can hide unfinished review or inspection work. The IT installation team needs the released physical inputs that result from these decisions; it does not replace the facility's designers or public authorities.

References: [A1]

Manufacture, delivery and integration are different milestones

A project may need a selected transformer, switchgear, UPS or cooling assembly before a dependent facility package can be completed. The order, manufacture, delivery, installation and integration describe different states. The supplier and facility team need to connect their actual delivery commitments to the approved equipment and interfaces. A shipment forecast alone does not establish the installed result.

DOE's Office of Electricity offers a narrowly defined manufacturing example. Its distribution-transformer discussion describes COVID-era supply and labor constraints and material shortages. It also reports that participants in a March 2023 workshop identified inconsistent utility specifications as a contributor to longer production times. That history explains how component supply and specification choices can affect manufacturing. It supplies no current delivery estimate for a particular order, large power transformer, switchgear or cooling system.

For the actual project, ask what the supplier has committed to deliver, which technical choices remain open and what site work is still required around that assembly. Keep manufacturing progress, received condition and the completed facility interface distinguishable. A substitute also needs its actual design and interface consequences assessed before its earlier delivery can be credited toward the required finish.

References: [D1]

Commissioning establishes more than the presence of installed components

A building filled with equipment still needs evidence addressing the owner's intended building performance. DOE's Federal Energy Management Program describes commissioning as verification against design intent and owner needs. For new construction and major renovations, its discussion includes design review, functional testing, system documentation and operator training across the project. This is federal-building guidance, not a universal commercial requirement or a data-center test procedure.

The scope implication is useful beyond that example: evidence about one component cannot alone establish the complete building outcome. A factory report can address its stated assembly and conditions while leaving site interfaces and the building's combined behavior to be established. Completing installation is consequently a different milestone from completing the owner's defined facility acceptance.

Where required evidence or an interface remains unresolved, the appointed facility and commissioning parties must resolve that work before their acceptance milestone can close. Their project defines methods, criteria, participating specialists and authority. The related acceptance guide explains how to read the boundaries of returned evidence; the physical IT schedule must retain these separately owned facility outcomes.

References: [C1]

Where a prepared physical-deployment team makes a useful contribution

The physical team's useful contribution starts before the field window: reconcile the identified deliveries and accessories with the released rack package, review the applicable equipment and cable inputs, and return missing or conflicting information to the owner who can resolve it. When the work is released, the team performs its agreed installation and IT cabling, records the checks actually completed and transfers usable equipment, endpoint and exception records to the next team.

Relevant experience matters at those specific tasks. Ask how the proposed people and supervision fit the actual equipment, permitted work and required records, then confirm their availability for the window. The related partner-selection guide covers the evidence to request. Extended hours are useful only where access, personnel, supervision and specialist interfaces support them. Where shifts are arranged, a handover should preserve the current document issues, performed work and open decisions so the next shift can continue on the same basis.

Consider an illustrative comparison with the same delivered IT package and proposed physical crew. In one plan, the intended facility phase has its required service and facility acceptance. In the other, only the building shell is handed over, with the relevant service stage and facility acceptance still outstanding. The same installation effort cannot answer both end-to-end schedules. In the second plan, separately permitted preparation may still be useful, but it cannot supply the missing facility outcomes. This is an explanatory comparison, with no durations or actual project results.

Leviathan's rack-and-stack service is the route for discussing that defined physical package and its records. Facility engineering and construction, electrical energization, rigging, liquid-side work and full software/system commissioning remain with separately appointed parties. Use the related physical-project inquiry with a nonconfidential equipment, site, target-window and handoff outline to discuss scope and availability. The schedule guidance that follows applies these boundaries to actual work packages.

References: [L1][G1]

Define what the finish date means

A delivery date, an installed-rack date and an operational-cluster date describe different milestones. State which one the schedule is meant to achieve. If the buyer needs a usable compute service, identify the separate facility, OEM, network, software and application work packages as well as physical installation.

For the physical package, describe the completion evidence: installed equipment locations, agreed cable records, inspection results and exceptions. Identify the receiving team and its review window. A contractor should not be scheduled to certify work outside its scope merely because the overall project has one target go-live date.

References: [L1]

Build a release-gate schedule before estimating labor

List each workfront and its predecessor. A workfront might be a cabinet group, route segment or documented installation package. Record the evidence that permits it to start and the person who can release it. Keep estimated task effort separate from elapsed time waiting for access, decisions or materials.

Use the matrix as a planning structure, then replace it with the actual project dependencies. OEM instructions and facility requirements may impose additional gates or a different sequence.

Scroll the table horizontally on smaller screens.

Illustrative schedule gates, without invented durations
Work packagePrerequisite to resolveCompletion evidenceSource
Receiving and stagingFacility delivery arrangement; receiving authorityIdentified equipment and unresolved-condition registerPlanning analysis
Physical installationApproved scope, locations, handling arrangements and complete accessoriesEquipment-location and inspection recordsPlanning analysis
CablingReleased endpoint schedule, route and customer-furnished materialsAs-installed endpoint record and scoped verificationPlanning analysis
Specialist facility/OEM workSeparate authorized work package and prerequisitesSpecialist owner’s own release recordPlanning analysis
Physical handoffAgreed deliverable format and exception decisionsAccepted physical records, with remaining work visiblePlanning analysis
Network/software acceptanceRelevant technical owner’s approved test planThat owner’s system/application evidencePlanning analysis

Original planning worksheet. Agree the fields and responsibilities for the project; this is not a safety procedure or acceptance standard.

Scope the Blackwell work package before assigning labor

A request for a Blackwell installation needs the exact platform and delivery configuration before it can support a labor estimate. NVIDIA's DGX B200 quickstart asks for relevant site information and states that installation must be performed by NVIDIA partner-network personnel or NVIDIA field-service engineers. That is a specific prerequisite for that system, not a qualification supplied by a generic installation crew line. Confirm the actual authorized party and its scope before making its work a scheduled resource. This article does not establish that Leviathan has that status.

The GB200 NVL72 product description names a rack-scale, liquid-cooled system containing 72 Blackwell GPUs. The rack is the documented system context; dividing that GPU count into assumed server-installation hours does not estimate its deployment work. Establish what is already assembled and verified before delivery, what remains at the destination, and which separately authorized parties own facility interfaces and system acceptance. The public architecture description does not tell the estimator the delivered condition, actual site releases, personnel or remaining task quantities.

Compare estimates at the same boundary. An estimate for receiving and placing supplied equipment is not interchangeable with one that includes site assembly, cabling records or separately owned system checks. Count the actual remaining tasks in the agreed package, then apply their resource and predecessor assumptions. Keep work effort, waiting time and the final milestone separate as the following schedule example does. A different platform or delivery configuration changes those inputs; neither a GPU-family label nor a published GPU count supplies a crew size, production rate or delivery duration.

References: [N1][N2]

Scroll the table horizontally on smaller screens.

Named Blackwell context changes the estimate inputs, not a universal labor rate
Documented contextLabor or schedule question to resolveWhat the public description does not supplySource
DGX B200 quickstart: site information and specifically authorized installation personnelWho is authorized and available for that package, and which agreed tasks are included?Leviathan authorization, an available crew, project hours or an installation method.[N1]
GB200 NVL72: rack-scale, liquid-cooled system with 72 Blackwell GPUsWhat arrives assembled, which destination tasks remain, and which facility/system-owner releases precede them?Delivered condition, on-site task counts, cooling/electrical/handling approval or a delivery duration.[N2]

Original estimate-boundary comparison of the cited platform descriptions. No actual customer configuration, site method, system commissioning, work authorization or calculated productivity is represented.

Separate crew effort, working duration and the site calendar

Ask for three different quantities: the estimated labor effort, the time an activity occupies an available workfront, and the calendar dates on which that workfront can be used. A person-hour estimate cannot become a delivery date until the available crew, working calendar and predecessors are specified. GAO’s Schedule Assessment Guide distinguishes effort from duration and treats working calendars as part of the schedule model. [G1]

For example, a fictional task estimated at 32 person-hours would occupy two eight-hour working days for two people only if that division of work is feasible and all 32 hours can be used for that task. It is arithmetic, not a GPU-installation production rate. An escort window, shared loading dock, inspection appointment or unavailable accessory can make the elapsed interval longer without changing the original labor estimate.

Record the estimator’s basis beside the number: quantity and configuration, included activities, crew roles, available hours, review time and unresolved assumptions. A second shift changes the estimate only when the facility and relevant contractors have confirmed access, supervision and resources. Do not silently turn eight-hour days into twelve-hour days to make a target date fit.

References: [G1]

Show which dependency actually moves the finish date

The following is a deliberately fictional scheduling calculation for two independent work areas. T0 is the start of the planning window; all figures are elapsed working-day offsets on one shared calendar. Durations are invented solely to explain dependencies. They are not installation productivity estimates, a recommended technical sequence or a Leviathan delivery offer.

In this simplified example, each area has separately available installation and cabling resources. The area-specific work is already authorized, and no unlisted facility, safety or technical dependency is assumed away in a real project. The common handoff begins only after both areas’ records are ready.

Scroll the table horizontally on smaller screens.

Fictional two-area schedule: changing a non-driving task does not necessarily change delivery
ActivityMust wait forIllustrative durationStart → finishSource
A installationArea A release at T02 working daysT0 → T2Planning analysis
A cabling packageA installation3 working daysT2 → T5Planning analysis
A records reviewA cabling package1 working dayT5 → T6Planning analysis
B installationArea B release at T32 working daysT3 → T5Planning analysis
B cabling packageB installation2 working daysT5 → T7Planning analysis
B records reviewB cabling package1 working dayT7 → T8Planning analysis
Common physical handoffBoth A and B records reviews1 working dayT8 → T9Planning analysis

Original planning example. T0–T9 are working-time offsets, not dates or promised durations. Resources are independently available for each area; replace every assumption with the actual authorized project.

Compare recovery options before adding crew

Under those exact fictional assumptions, reducing A installation from two days to one moves A’s records-ready point from T6 to T5. It leaves the common finish at T9 because B still releases the handoff at T8. The buyer would pay for faster work on A without moving the milestone being purchased.

Moving B’s release from T3 to T1 has a different effect. B installation then ends at T3, its cabling at T5 and records review at T6. With A also ready at T6, the common handoff can finish at T7. That two-day improvement exists only in this simplified model. If the areas share a cabling crew, an access window or a receiving reviewer, their resource conflict must be resolved before using the T7 forecast.

For each proposed recovery action, ask which named predecessor changes, what authority permits the change, which resource is available and how the finish calculation changes. Compare the earliest plausible finish with the risks introduced by resequencing. A chart with shorter bars is not enough. GAO’s guidance distinguishes the driving path from activities that merely appear critical because of imposed date constraints. [G1]

References: [G1]

Expose the constraints that crew count cannot solve

Confirm usable work hours, access restrictions, escort needs, shared staging areas and the number of workfronts the site can release. More people do not resolve a missing approved route or unavailable customer-supplied rail kit. Additional shifts also require explicit site and staffing arrangements; they are not a default service promise.

Group shortages by the tasks they block. A missing item may stop one cabinet while leaving another independent workfront available. State the re-planning decision and who approves it. Avoid presenting unfinished fragments as completed racks to make daily quantities look better.

Turn each hold point into an owned decision

A useful hold entry identifies a blocked activity and the evidence needed to release it. “Waiting on customer” is too broad to support a forecast. Name the requested item, its decision owner, the date by which the scheduler needs it and the next activity affected. Keep the actual authorization with the responsible party; a scheduler can record a release but cannot create technical permission by moving a bar.

Use one status for the prerequisite and another for the downstream work. “Rail kit delivered” does not say that the correct kit is accepted for the actual server and rack. Likewise, “cable schedule issued” does not say that its endpoints are the revision released for this work area. The project’s qualified owners settle applicability; the schedule records their result.

Scroll the table horizontally on smaller screens.

An actionable hold register for a physical installation schedule
Blocked activityEvidence to requestDecision owner to nameForecast treatmentSource
Cabinet group cannot startActual equipment/accessory list and outstanding discrepancy dispositionAppointed equipment or purchasing owner; applicable OEM/configuration authorityKeep the start conditional; show the decision needed before the proposed startPlanning analysis
Cable package cannot startReleased endpoint schedule and permitted work areaDesign owner and facility/work-area authorityIdentify the specific affected package; keep independent authorized work visiblePlanning analysis
Records review cannot finishMissing record IDs, required correction and receiving reviewerContracted receiving/acceptance ownerForecast the outstanding review and correction work, not only field installationPlanning analysis
Next specialist package cannot startThe specialist’s stated predecessor evidence and releaseRelevant facility/OEM/network/system ownerCarry the interface milestone; do not claim the physical installer has certified itPlanning analysis

Original coordination worksheet, not an approval procedure, technical acceptance standard or contractual allocation of liability.

Handle a late change as a dependency change

Illustrative example: the network design owner changes the destination switch ports after one cabinet group is installed. Identify which cable records and physical links are affected, hold that affected package for instructions and assess the work needed to revise it. Preserve the completed records that remain valid.

Update the dependency, resource and acceptance implications together. Moving only the end date hides the cause and can repeat the same problem at the next workfront. Commercial treatment of a scope change belongs to the agreed project process; this guide does not establish contract remedies or pricing terms.

Update the forecast from remaining work, not a completion percentage

Use a stated status date for every schedule update. Record what actually started or finished, which gate evidence was received, and the current estimate of remaining work. Keep the approved baseline available for comparison; do not overwrite it merely to make the current forecast appear on time. GAO separates updating actual progress from revising the future plan and warns that time consumed is not the same as work accomplished. [G1]

A cabinet can be physically populated while its cable records, exceptions or agreed inspection remain open. Report those outputs separately. A sentence such as “installation complete; two endpoint discrepancies await design disposition; receiving review has not started” tells the buyer more than “90 percent complete.” The actual forecast should identify who must resolve the discrepancies and what review time remains.

At the coordination meeting, address the next blocked milestone first: what changed since the previous status date, which finish it affects, which owner has the next action, and when the forecast will be updated again. Escalation is useful when it requests a specific decision. It is not evidence that the decision has been made.

References: [G1]

Request a schedule with stated assumptions

Provide the equipment and rack list, site conditions, current drawings, endpoint schedule, delivery plan, access windows and required handoff. Ask for the estimate’s assumptions, customer dependencies and exclusions alongside the proposed field effort. Ask how a blocked task will be reported and who can release it.

Leviathan can discuss a schedule for the agreed physical scope once these inputs are known. No generic week count, percentage saving or guaranteed mobilization date is implied. Keep software health checks and system acceptance with their technical owners; NVIDIA’s DGX B200 guidance is one example of a separate system operation and diagnostic workflow.

References: [L1][N1]

Common planning questions

How long does a particular number of GPU racks take?

A rack count is not enough to estimate elapsed time. Request a scope-specific schedule with configuration, site, material, access and acceptance assumptions.

Can cabling and equipment installation run in parallel?

Only where the approved work packages, access and site arrangements permit independent work. The project leads must resolve shared routes, handling areas and acceptance dependencies first.

Why can more installers fail to shorten the deployment date?

Extra labor helps only if it shortens work that drives the required finish and the work can actually be divided. A delayed release, shared access window, specialist dependency or receiving review may still control delivery. Calculate the effect on the linked schedule before treating additional crew as recovery.

How should a schedule show an approval that has not arrived?

Keep the downstream start conditional and name the specific evidence, decision owner and date required. Show the forecast assumption separately from the actual release. The schedule does not grant technical or facility authorization.

Does a physical handoff date mean the GPU cluster is ready for production?

Only if the separately owned facility, network, system and application acceptance requirements for that outcome have also been met. The physical package’s completion evidence and the compute service’s readiness criteria need distinct owners and milestones.

Can I use the two-area example as a delivery benchmark?

No. Every duration is fictional and exists to demonstrate dependency arithmetic. A real estimate needs the actual configuration, approved scope, material and site releases, access calendar, resources and acceptance requirements.

Primary sources

  1. [L1] Leviathan rack-and-stack service boundariesChecked 2026-10-14. Current public scope, conditional handling and customer-furnished materials; each engagement is separately agreed.
  2. [N1] NVIDIA DGX B200 User Guide: installation and basic operationChecked 2026-10-14. DGX B200 only; installation authorization and software/health workflows, not a generic procedure for other systems.
  3. [G1] U.S. GAO Schedule Assessment Guide, GAO-16-89G (December 2015)Checked 2026-10-18. General schedule-management principles: effort/duration and calendars (printed pp.63–64), driving paths (pp.75–79), progress updates and status date (pp.121–124). Not a GPU-installation method, productivity benchmark, technical release or Leviathan service guarantee.
  4. [N2] NVIDIA GB200 NVL72: rack-scale system contextChecked 2026-10-19. Named rack-scale, liquid-cooled72-GPU system context only. No performance/TCO claim, installation sequence, delivered condition, site interface, availability, productivity or engineering approval inferred.
  5. [U1] Dominion Energy Virginia — Data Center RequestsChecked 2026-10-22. Virginia utility example of site-specific plans, staged service and overlapping infrastructure, land-right, approval and material dependencies. The actual site's service plan determines applicability; no general connection duration or capacity commitment is inferred.
  6. [A1] Loudoun County, Virginia — Building & DevelopmentChecked 2026-10-22. Department overview distinguishes plan review, permits, construction inspections and occupancy/use approval. Useful for separating approval outcomes; it supplies no project-specific permit decision or universal jurisdictional sequence.
  7. [D1] U.S. DOE Office of Electricity — Supply Chain and Market AnalysisChecked 2026-10-22. Distribution-transformer discussion describes historical supply, labor and material constraints and findings from a March 2023 workshop about specification differences. It provides manufacturing context, not a current equipment-order forecast or lead time for other equipment classes.
  8. [C1] U.S. DOE FEMP — Commissioning in Federal BuildingsChecked 2026-10-22. Federal-building guidance on design intent, owner needs, design reviews, functional evidence, documentation and operator preparation. Useful for explaining the building-outcome boundary; actual commercial project requirements and specialist acceptance remain separately defined.

Build a scope-based installation schedule

Share the equipment list, site window, current plans and target handoff so we can discuss dependencies and availability.

Discuss a ProjectEmail Project Scope