How long does data center commissioning take? Scope-normalized benchmarks

How long does data center commissioning take?

Allow 8 to 16 weeks from first energization to integrated systems test completion for a single data hall, according to Project Completion Management's published planning guidance, checked November 5, 2026. Load-bank availability, fuel logistics and unfinished installation or startup work drive that allowance. source 1 For a separate scope of 270 equipment packages plus assigned integrated tests, using counts from CAI's published 36 MW project, Anvilfield estimates 107 working days, or about 22 five-day workweeks, from readiness review through selected-package closeout using the crews, effort rates and allowances below. source 2

The single-hall allowance is practitioner guidance; the worked calculation covers selected packages. CAI's 2024 Cheyenne case study reports a 36 MW facility, five turnover phases and a two-year commissioning engagement, rather than a separately measured two-year L4/L5 test campaign. source 2

Use the L1 to L5 release gates to identify the work, then attach a clock to that scope.

When does the commissioning clock start and stop?

Write the start and finish milestones before quoting a duration. BCxA's January 2024 New Construction Building Commissioning Best Practices covers commissioning through occupancy and operations, so the whole process has a longer clock than final acceptance testing. source 3

Keep baseline dates, actual dates and forecasts in separate fields. Record timestamp and timezone, elapsed calendar days, working days and crew-hours; each answers a different question.

Anvilfield evidence register v1.1 retains a paused project's elapsed interval when actual dates exist and tags the pause reason separately from active effort.

Which finish date belongs in the benchmark?

Use accepted test completion for a test-only cohort and signed acceptance-package completion for a closeout-inclusive cohort. WBDG's live Commissioning Documents guidance distinguishes report contents and owner acceptance, making the accepted record part of the endpoint definition. source 6 Beneficial occupancy, certification, first tenant load and certificate of occupancy need their own fields when they differ from that endpoint.

GSA's September 2020 Commissioning Guide places deferred and seasonal testing in occupancy and operations. source 5 Record initial handover and final seasonal closeout separately, with an owner, due date and required conditions for every deferred test.

How should L1 to L5 durations be scheduled?

Schedule each level from its actual tasks, holds and booked resources. BCxA's May 2, 2024 article places installation checks within L2 and equipment startup in L3; CXPlanner's live testing-levels page describes L2 as pre-installation receiving and L3 as startup and first functional checks. source 4 source 7 Crosswalk the tasks before comparing durations under the same level label.

Table 2’s L1 to L3 rows are separate Anvilfield estimates for the small scopes shown; they are outside the 270-package model. Replace their assigned tasks and rates with approved factory, receiving and OEM startup requirements.

Why does an eight-hour test need more than eight hours on the schedule?

Uptime Institute's May 4, 2015 technical article recommends at least eight hours of continuous load-bank runtime and describes up to 24 hours as best practice. source 8 That recommendation concerns a continuous test run, rather than the length of an L4 phase, an L5 campaign or UPS battery autonomy.

Reserve connection, permits, isolation, load staging, stabilization, hold, recovery, required recharge, evidence review and restoration as distinct activities. Staff continuous holds throughout their required windows under the approved test procedure. An eight-hour hold consumes an uninterrupted eight-hour window; schedule setup and restoration outside that hold, even when average daily productive capacity is six hours. Sharing an observation window across assets requires the approved plan, instruments, load capacity and witnesses to support that concurrency.

Build equipment-specific windows from electrical commissioning and power QA, UPS and STS commissioning hold points and generator acceptance and load-bank evidence. Build the integrated block from the approved integrated systems test sequence, including event recovery and restoration between scenarios.

Which projects qualify for a scope-normalized benchmark?

Anvilfield register v1.1 admits a project to a common empirical cohort only when its accepted endpoints and workload match the cohort definition. Table 3 retains five project reports, O1 to O5. G1 is planning guidance and X1 is an installation comparison; neither increases the project-report count.

Record new construction, expansion or retrofit; live operational adjacency; phase; shared plant; liquid-cooled scope; redundant paths; and owner acceptance coverage. A Tier classification describes resilience context without assigning the number of tests or days.

Keep design and procurement outside a site-testing cohort, and report delivery-only, installation-only, tenant IT burn-in and certification intervals separately. Deferred tests outside initial acceptance belong to the seasonal tail. Utility waiting outside the defined start is excluded; waiting inside the observed interval stays as a tagged delay. The grid and interconnection guide helps identify the separate utility milestone.

O2's three months remains the source's reported unit rather than an exact 12-week measurement. O3 includes removal and installation. O4's generating MW cannot become critical IT MW. O5's introductory and Vision descriptions aggregate floor area differently, so the register preserves both descriptions without floor-area normalization.

Count the WorkingBuildings web report and any duplicate PDF as one engagement. Likewise, five CAI turnover phases remain one project with five phases; without phase-level intervals and workload, splitting that engagement would manufacture independent observations.

What workload denominator makes duration comparable?

Anvilfield method M1 uses unique required test packages in a frozen scope as its principal workload count, with crew-hours per defined package as the effort rate. Count every required package once, including a package that fails its first attempt. Publish asset classes, test modes and integrated scenario counts alongside the total.

CAI's 2024 equipment list supplies 45 UPS, 44 automatic transfer switches, one diesel generator, 20 natural gas turbine generators and 160 COLO AHUs. source 2 M1 assigns Table 4’s rates to these 270 packages. Each asset package includes every required script mode and acceptance check.

The selected reference workload averages 3.17 first-pass crew-hours per package. One crew-hour is one complete crew working for one hour; M1 assigns two-person crews, so 856 crew-hours equals 1,712 person-hours. Match crew composition or convert to person-hours before comparing projects with different team sizes.

A transferable production rate needs matching test workload, crew composition and acceptance coverage. M1 assigns the same eight-hour rate to diesel and gas-turbine generator packages; use separate rates when their approved procedures require different effort.

What must a reusable benchmark record contain?

Retain record ID, source URL and date, project and phase, live/new status, MW and its basis, equipment counts, script revisions, coverage, crew composition, crew count, productive hours, calendar, actual endpoints, first-pass effort, repeats, correction waits, acceptance and exclusions. Freeze the register before execution and log additions or removals against a scope revision.

Track attempted, first-attempt-passed and finally accepted counts separately. Put repeated work in the retest numerator without removing failed packages from scope. For unlike equipment mixes, weight both scopes with the same versioned reference effort rates and divide actual first-pass crew-hours by reference crew-hours to obtain an effort index. M1 has assigned reference rates but no actual-effort observation, so its observed effort index is not reported. Integrated scenarios remain a separate workload track.

What do the observed durations actually show?

Anvilfield's November 5, 2026 review of five public project reports found four reported intervals but zero complete comparable whole-facility L4/L5 records, so it calculates no empirical median or P80 duration. The register is a bounded public-source sample.

Johnson Controls' 2019 case study, document code 100919, reports a controls-related test schedule changing from three months to four weeks and separately associates AHU factory pretesting with a four to six-week reduction. source 9 Air Water Power reports 2.5 weeks completed against four weeks allowed for a live CRAC replacement project. source 10 Aggreko reports 12 weeks from contracting through commissioning for a temporary power deployment. source 11 None supplies the complete normalized whole-facility record defined here.

Can these sources establish P50 or P80 commissioning duration?

For future cohorts, Anvilfield sorts completed durations, multiplies the percentile fraction by observation count, rounds upward and selects that ranked observation: rank = ceil(p × n). Publish count, missing-duration count, actual date span, calendar and method. Anvilfield’s editorial rule suppresses cross-project percentiles below five comparable observations; five does not establish precision or an on-time probability.

Keep ongoing projects open/right-censored, with start and last-observed dates, rather than inserting forecast finishes into completed observations. Report those open records beside the completed cohort to expose unfinished long-duration work. Missing-duration reports remain not reported; provider-selected cases and related campus phases also limit representativeness.

How do you calculate commissioning duration from test workload?

Multiply each equipment count by its first-pass crew-hour rate, add those products and divide by available crew capacity per workday. Calculate repeated effort and integrated work separately, round each activity upward, then place the blocks on their dependency paths.

In Anvilfield method M1, the first-pass effort is H4 = Σ(Ni × hi), where Ni is the package count in equipment class i and hi is crew-hours per package. First-pass days are D4 = ceil(H4 / (C4 × q4)); repeated-test days are DR = ceil(H4 × r / (CR × qR)); integrated days are D5 = ceil(N5 × h5 / (C5 × q5)). C means complete concurrent crews, q productive crew-hours per crew-day, r repeated effort divided by first-pass effort, and N5 and h5 integrated segment count and effort.

The quotient calculates crew-capacity duration; use the longest dependency path after assigning resources to determine project duration. For each activity, find the earliest start allowed by predecessor completion and resource availability, then reserve its complete execution window.

A late witness booking shifts the start; a continuous hold sets a minimum elapsed duration even when labor capacity is available. The critical path and project scheduling and look-ahead planning turn the calculated blocks into a dated forecast.

What is the model's calendar?

M1 selects five workdays per week, eight staffed hours per day and six productive hours per crew-day. It assigns two interchangeable L4/retest crews and one integrated-test crew.

Six productive hours is an average capacity input. Reserve full continuous hold windows with the required coverage and lengthen the schedule when those windows cannot be staffed. Assign holidays, outages and vendor availability before promising elapsed dates: 107 workdays converts to 21.4 five-day workweeks, rather than 107 calendar days.

How much time should commissioning retesting receive?

In Anvilfield's worked schedule, a 20% repeated L4 effort allowance adds 171.2 crew-hours and 15 working days, while five exposed correction days and two regression days are separate schedule blocks. First-pass yield counts successful first attempts; repeated effort measures rerun time, including multiple repeats.

BCxA's January 2024 best-practices guide calls for specifications to identify test scope, readiness and retesting responsibility, including acceptable failure and sampling provisions where sampling is used. source 3 It establishes no universal retest percentage. Uptime's 2015 mission-critical article favors full component testing and discusses repeating unsuccessful tests, including repeated attempts. source 8 Retain the owner’s complete mission-critical scope; reducing test coverage to meet a date changes the acceptance scope.

How do you choose a retest percentage?

Start with repeated crew-hours from completed packages that match the equipment, script and crew. Divide those hours by the same packages’ first-pass crew-hours, including successful and failed attempts, and separate rates by equipment class when the rerun burden differs. Early failures concentrated in one controls sequence justify revising that affected scope before increasing every package’s allowance.

Anvilfield estimates 5% repeated effort for 100 packages assigned two first-pass crew-hours each when 20 packages each need a half-hour rerun: ten repeated crew-hours divided by 200 first-pass crew-hours. That case has 80% first-pass yield. A 20% failure rate therefore need not mean 20% repeated effort.

Which failures add effort, and which add calendar delay?

A failed criterion opens a loop: record the failure and conditions, assign correction, confirm renewed readiness, rerun the affected test, complete dependent regression, restore settings and close the record with evidence. A sequence or firmware change can reopen downstream accepted tests; trace those dependencies in the issue register.

Exposed correction delay is the repair wait left on the controlling path after concurrent fixes. Several deficiencies repaired during other tests do not each add their full age to the finish. M1 carries five exposed correction days and two regression days, the latter providing 24 available crew-hours with two crews. Expand that regression block from the affected-test list when a change crosses electrical, cooling and controls interfaces.

How do you update the forecast after the first failures?

Recalculate untested first-pass work, add known reruns and apply fresh contingency only to remaining uncertain work. Remove completed work and consumed allowances. Recalculate outstanding correction waits and regression against the current path and issue a forecast with its data date and scope revision.

Anvilfield's remaining-work estimate assigns 240 first-pass crew-hours, 36 known rerun hours and 48 contingency hours, the last equal to 20% of remaining first-pass effort. Their 324 crew-hours require 27 workdays at two crews and six productive hours per crew-day. Add the outstanding correction, integrated and closeout blocks on the current path. Use the punch-list and turnover record to connect each remaining hour to an unresolved deliverable.

Worked example: a schedule from a published 36 MW equipment list

Anvilfield estimates 107 working days, or 21.4 five-day workweeks, for its selected 270-package L4/L5 model, using 856 first-pass crew-hours, two functional-test crews and a 20% repeated-effort allowance. Reserve 22 whole workweeks for Table 6’s blocks.

Allocate approved integrated events, holds, recovery and evidence across Table 6’s 24 assigned segments. Their 36 crew-hours require six working days with one crew at six productive hours/day.

Anvilfield estimates use CAI's 2024 published counts for four equipment categories and planning inputs selected on November 5, 2026: one aggregated package per asset, the stated crew-hour rates, two functional-test crews interchangeable across these package classes, six productive hours per crew-day, a 20% repeated-effort allowance and the fixed readiness, correction, integrated-testing, regression and closeout blocks. The calculation rounds each activity upward to a working day and schedules these blocks in series; actual overlap must be proved in the dependency and resource schedule. The estimate assumes readiness, qualified crews and continuous windows fit within those blocks; Table 8 restricts crews to separate disciplines.

The selected scope excludes other CAI-listed equipment, including transformers, switchboards, PDUs, STSs, control panels, fans and ancillary systems. It also excludes L1 to L3 execution, procurement, shared infrastructure outside the packages, project-specific fire/life-safety verification, tenant IT burn-in, certification and seasonal work. Add those scopes explicitly when estimating complete facility acceptance.

Keep the four-day closeout block tied to completed, accepted records using the proof-packet assembly guide.

Which assumptions move the finish date most?

Table 7 varies L4 capacity, productive hours, repeats and exposed correction delay with scope fixed. Integrated testing stays at six days; readiness, regression and closeout stay at three, two and four days.

The combined lower and higher selections produce 87 to 137 working days, or 17.4 to 27.4 five-day workweeks. Reserve 18 to 28 whole planning weeks for those selected cases. They are input scenarios, with no assigned probabilities.

Anvilfield's discipline-restricted variant assigns one crew to the 616 electrical crew-hours and one to the 240 AHU crew-hours. Their first-pass tracks run in parallel, followed by a common finish gate and a separate parallel retest block. The electrical track controls both blocks, producing 144 workdays, 37 more than the interchangeable-crew estimate.

Reserve 29 whole workweeks; transferring mechanical capacity requires electrical qualifications and support. Releasing retests asset by asset before the common gate or transferring qualified people changes the network; recalculate it rather than retaining the same result.

How do phasing, live sites and prefabrication change the schedule?

Phase-level schedules need shared-system dependencies. A common generator plant, cooling header or control sequence can hold several halls behind the same readiness gate. Record shared work once and identify dependent hall milestones.

Assign equipment and scenarios to each phase and retain regression of shared systems after later connections. Schedule each phase from its own register; CAI’s five-phase count supplies no per-phase workload allocation.

Air Water Power's live-site project includes removal, installation, integration and commissioning within its reported interval. source 10 Live operations also require the permitted outage and restoration windows to fit the work. Treat liquid cooling loop commissioning as explicit added scope where flushing, pressure tests, balancing and interface checks enter the package list.

Schneider Electric WP238 v1.0, dated October 16, 2026, reports 15 onsite installation hours for a prefabricated pod against a 75.5-hour traditional comparison, approximately an 80% reduction. source 12 Schneider's October 20 explanation distinguishes observed prefab time from a traditional comparison combining observations and specialist schedules, and excludes later commissioning and offsite manufacture. source 13

Keep that installation result outside commissioning-duration observations. Factory validation completes selected work before delivery, while installed interfaces and site-wide sequences retain their specified tests. JCI's factory controls work and Schneider's pod installation study cover different tasks; use each against its matching scope instead of combining their reductions into one commissioning acceleration percentage.

How do you price commissioning time and retest allowances?

Price the work package and resource calendar together. Identify crews, witnesses, OEM support, load banks, connections, temporary power, fuel, instruments, travel, standby, repeats, reporting and mobilizations. Tie inclusions to the scope and schedule revision.

For public capital work, separate commissioning support from installation and general construction in the bid inclusions. State the test packages, included repeat visits and responsibility for supplied resources. Estimators pricing commissioning support on public capital work use PorchRocket Bid Control to see who won past bids and for how much, alongside their own defined test and retest scope.

M1 assigns two people at $160 per person-hour, giving $320 per crew-hour. Its 1,087.2 selected productive crew-hours produce an Anvilfield estimate of $347,904; calculate from the unrounded quantities rather than adding the rounded table rows. CAI’s broader $1,820,000 engagement lacks a detailed cost allocation. source 2

The subtotal excludes markup, fuel, rentals, standby, travel, OEM charges, reporting, readiness/closeout labor and unproductive paid attendance. Price these separately.

What changes when a test window slips?

Separate the extra test effort from paid standby, reserved vendor days, rental extensions and remobilization. A displaced appointment can increase cost without changing the test count. Apply contracted notice periods and included repeat visits, recording authorization for added work.

Commissioning schedule readiness checklist

Anvilfield's schedule-release rule requires zero unresolved issues that block the scheduled test's acceptance criteria. Record owner-permitted nonblocking issues with their agreed disposition. BCxA's January 2024 best-practices guide distinguishes contractor correction, commissioning-provider verification and owner acceptance. source 3

Use the commissioning roles and turnover guide to name the required signatures and the look-ahead to confirm booked support. Resolve missing checklist inputs before releasing the affected activity.

GSA's September 2020 guide describes a warranty review at 10 months into a 12-month warranty. source 5 Schedule that GSA program milestone separately and retain the project’s contracted follow-up obligations.

• Identify the hall, phase and shared infrastructure on the marked-up scope drawing; split unlike phases into separate records.
• Write the exact start and finish milestones; replace undated schedule labels before comparing elapsed durations.
• Mark reported MW as IT, facility, generation or unspecified; retain unspecified when the source does not define it.
• Reconcile equipment count with the register and script list; list excluded categories beside the estimate.
• Crosswalk each project's L2/L3 labels to receiving, installation and startup tasks; price the tasks actually required.
• Obtain signed readiness evidence and verify zero acceptance-blocking unresolved issues for the next scheduled test; retain permitted nonblocking items in the issue register.
• Confirm each crew's discipline qualifications, OEM attendance, required witnesses and their actual notice windows; rebook the activity if any required role is unavailable.
• Confirm load-bank capacity, connections, thermal rejection, fuel and instrumentation against the approved test plan; constrain concurrency to the limiting resource.
• Reserve uninterrupted event, hold, recovery and restoration windows; extend the calendar when these exceed ordinary shifts.
• Carry repeated effort, exposed correction delay and dependent regression separately; update remaining work, known reruns and unused contingency after a failed criterion or sequence change.
• Name the acceptance authority, require signed results and retain script revisions, as-left settings and issue closure evidence before closing the selected package.
• Assign dates and owners for deferred/seasonal tests, final records and warranty review; retain initial handover and final commissioning closeout as separate milestones.

Review date: 2026-11-05. Update log: verified all 14 sources and calculations; clarified continuous test windows and retest denominators; corrected table order. Evidence register v1.1; Anvilfield estimate method M1 v1.0.

How to cite this page: Anvilfield, How long does data center commissioning take? Scope-normalized benchmarks, 2026, https://anvilfield.com/field-guides/datacenter/data-center-commissioning-duration-benchmark/

Related tools

FAQ

People also ask

• What is data center commissioning?
• What is an integrated systems test?
• What is commissioning in construction?
• What is a commissioning hold point on a UPS?
• What is an NFPA 110 acceptance test?

Sources checked

• S1. Project Completion Management, Data Center Commissioning Levels: L1-L5 Explained Undated guidance; single hall, first energization to IST completion. Reopened with curl and checked 2026-11-05.
• S2. CAI, Commissioning Services for LEED Gold 36 MW Data Center, 2024 Pages 2 to 3: engagement, contract and equipment inventory. Reopened with curl and checked 2026-11-05.
• S3. BCxA, New Construction Building Commissioning Best Practices, January 2024 Process scope, readiness, specifications and retesting responsibility. Reopened with curl and checked 2026-11-05.
• S4. BCxA, Commissioning Levels ≠ Commissioning Process, May 2, 2024 Level/task crosswalk and distinction from the full process. Reopened with curl and checked 2026-11-05.
• S5. GSA Commissioning Guide, September 2020 Pages 32 to 33: deferred/seasonal testing and GSA warranty review. Reopened with curl and checked 2026-11-05.
• S6. WBDG, Commissioning Documents: Process, Contents, and Acceptance Undated live guidance: reporting and owner acceptance; used for process guidance, not current standards editions. Reopened with curl and checked 2026-11-05.
• S7. CXPlanner, Data center commissioning: The 5 levels of testing Undated live page: L2/L3 terminology. Reopened with curl and checked 2026-11-05.
• S8. Uptime Institute Journal, Improve Project Success Through Mission Critical Commissioning, May 4, 2015 Dated technical recommendations: test coverage, resources, runtime and repeat testing. Reopened with curl and checked 2026-11-05.
• S9. Johnson Controls, Data Center Solutions: Large Enterprise Cloud Provider, 2019, 100919 Vendor-reported controls test schedule and separate factory pretest saving. Reopened with curl and checked 2026-11-05.
• S10. Air Water Power, Live Site CRAC Replacement in Higher Education Facility Data Centre Undated provider report: retrofit project interval, including installation. Reopened with curl and checked 2026-11-05.
• S11. Aggreko, Gas power and battery storage for a data center Undated provider report: generation scope and contracting-to-commissioning interval. Reopened with curl and checked 2026-11-05.
• S12. Schneider Electric, WP238, Prefabricated White Space Module Time Study: Lessons from a High-Density Liquid-Cooled Deployment, v1.0, October 16, 2026 Manufacturer document page: installation comparison and version metadata. Reopened with curl and checked 2026-11-05.
• S13. Schneider Electric, Measuring the impact of prefabricated IT pods, October 20, 2026 Study explanation: observed and modeled comparison boundaries, commissioning exclusions. Reopened with curl and checked 2026-11-05.
• S14. Trinity Consultants/WorkingBuildings, Trinity Ensures 24/7 Reliability for Data Center Operations Undated report: L4/L5 scope, inconsistent floor-area aggregation, no reported duration. Reopened with curl and checked 2026-11-05.