---
title: 'Data center precon is not commercial precon with more power'
url: 'https://piper-ai.com/resources/data-center-preconstruction-vs-commercial'
description: 'On a data center the risk leaves the building and moves into owner-furnished gear, commissioning, and a schedule set by power. What changes for GC precon.'
contentTypes: [comparison, article]
publishedAt: '2026-09-26T12:00:00.000Z'
updatedAt: '2026-09-27T11:23:39.829Z'
author: 'Ido Gedanken, CEO'
readingTime: 10
---

# Data center precon is not commercial precon with more power

1. Price a data center per MW of IT load, not per square foot; the shell is the small part of the number
2. Redundancy multiplies electrical quantities, so count lineups by topology, not by megawatt
3. When the owner buys the big gear, the subcontract bid becomes a price for labor and interfaces
4. Commissioning is the real finish line: five levels, ending in a full-load test the contract dates are tied to
5. The critical path runs through the utility and the transformer factory, not through the building

**A commercial GC that treats its first data center as an office building with a bigger electrical package will price the building well and miss the job.** On a data center the risk leaves the building. It moves into three places the commercial playbook treats as footnotes: the interfaces around equipment the owner buys directly, the commissioning that has to prove the redundancy works, and a schedule set by the utility and the transformer factory rather than by the trades.

This comparison is for estimators, precon managers and directors at commercial GCs moving into data center work, most often as a negotiated GMP or design-build program for a hyperscale or colocation owner, and usually at the electrical and mechanical package level where most of the money sits. Delivery method still matters ([hard-bid vs CMAR vs design-build](https://piper-ai.com/resources/hard-bid-vs-cmar-vs-design-build)). This piece is about what the building type does to precon, whichever contract it sits under.

## What a data center is, in estimating terms

Strip out the servers and a data center is two machines in an enclosure. One brings in a very large amount of electricity and delivers it to every rack without interruption. The other removes the same energy again as heat. Almost every watt a rack draws leaves it as heat that has to be carried to the roof or the yard.

Three words organize the whole job:

- **White space** is the data halls where the racks sit.
- **Gray space** is everything that keeps them running: switchgear, UPS, batteries, cooling plant, controls.
- **The yard** holds the substation, the generators and outdoor mechanical equipment.

Owners size, sell and lease the building in **megawatts of IT load**, and the cost follows. Cushman & Wakefield's 2026 Data Center Development Cost Guide puts the US greenfield average at $17.6 million per MW, excluding the computing equipment, up 21 percent per MW since late 2024, with power infrastructure as the single largest line. A dollar-per-square-foot benchmark borrowed from office work will be wrong in both directions: too high for the enclosure and far too low for what goes inside it.

The density is what makes the systems dominate. A current AI rack such as NVIDIA's GB300 NVL72 is specified at about 135 kW, with a 155 kW peak, in the manufacturer documentation Lenovo publishes. An entire floor of a typical office building may draw less than a single row of those racks.

## Side by side: what changes for precon

| Dimension | Commercial office or lab | Data center |
| --- | --- | --- |
| Unit of value | Dollars per square foot of rentable area | Dollars per MW of IT load |
| Where the cost sits | Shell, envelope, interiors, MEP in proportion | Power and cooling systems; the enclosure is a small share |
| Major equipment | Contractor-furnished through subcontracts | Often purchased by the owner and installed by the GC's subs |
| Redundancy | One path, code minimum | N+1, 4-to-make-3, or 2N by design |
| What drives the schedule | Trade sequence, weather, inspections | Utility energization and electrical gear lead times |
| Commissioning | A closeout activity | Five levels, ending in a full-load integrated systems test |
| Repetition | Each building is a new design | Campuses of near-identical buildings built to an owner standard |
| Subcontractor market | Several bidders per trade | A few firms with the craft, bonding and capacity to perform |
| Design stability | Frozen at construction documents | Owner standards change with each generation of computing hardware |

Each row changes a specific part of the precon job. The rest of this piece takes them in order of how much money they move.

## Redundancy multiplies the equipment, not the building

A commercial building has one path for power. A data center is designed so a component can fail, or be shut down for maintenance, without the racks noticing. The design rule for that is its topology, and it is the largest single driver of electrical quantity.

Uptime Institute's Tier Standard names the outcome. Tier III is **concurrently maintainable**: any component can be taken out of service while the site keeps running. Tier IV is **fault tolerant**: a failure is stopped before it reaches the IT load. The topology is how the engineer gets there, and each one carries a different amount of gear for the same load:

- **N+1** adds one spare module to the modules the load needs.
- **Block redundant, "4 to make 3"** runs four independent power trains, each carrying about three quarters of its rating, so any one can fail and the other three absorb its load. Vertiv's published AI reference designs use this pattern.
- **2N** builds a complete A side and a complete B side, each able to carry the whole load alone.

Schneider Electric's white paper comparing UPS configurations makes the arithmetic concrete: for the same 3.2 MW site, a 2(N+1) design needs ten 800 kW UPS modules, while a distributed redundant design needs six. Two buildings with the same megawatt rating can carry very different amounts of switchgear, UPS, feeders and terminations. **Count lineups by topology, not by megawatt,** and check that every bidder counted the same ones.

Redundancy also changes what a substitution means. Uptime Institute has written up cases where post-design value engineering removed maintenance valves, combined transfer switches and swapped equipment, and in doing so quietly broke concurrent maintainability. On an office building a substitution is a price conversation. On a data center it can break the design rule the owner is paying for.

## The owner buys the big gear, so the bid becomes labor and interfaces

Hyperscale and many colocation owners buy generators, switchgear, UPS, transformers and cooling units directly, through their own vendor agreements, and hand them to the GC to install. This is owner-furnished, contractor-installed equipment (OFCI). It removes the GC's equipment margin and leaves what the GC can actually control: labor, coordination, installation quality and the evidence commissioning will ask for.

The consequence for leveling is direct. An electrical proposal on an OFCI job is mostly a price for a sequence of custody and responsibility steps: receive, store, rig, set, assemble, terminate, start up, support the manufacturer, support integrated testing. Two proposals can be close on the cover price and far apart on which of those steps they include. Contract advice published in Construction Dive puts the same point from the owner's side: the agreement has to say who will procure, transport, offload, install, integrate and test each system.

That is its own discipline, and it has its own guide: [leveling bids when the owner buys the equipment](https://piper-ai.com/resources/data-center-owner-furnished-equipment-bid-leveling).

## Commissioning is the real finish line

On a commercial job, commissioning is something that happens near closeout. On a data center it is how the owner proves the redundancy works, and the contract dates are increasingly tied to it. Commissioning runs in five levels:

1. **Level 1:** factory witness testing at the manufacturer.
2. **Level 2:** delivery inspection against the approved submittals.
3. **Level 3:** pre-functional checks and startup of installed equipment.
4. **Level 4:** functional testing of each system, including its failure modes.
5. **Level 5:** the integrated systems test (IST). Every system runs together at design load on load banks, with the utility deliberately dropped.

Each level has a different party in charge, and each has costs that belong somewhere in the number: load bank rental, generator fuel for testing, manufacturer field technicians billed by the day, temporary power, and re-test days when something fails. Uptime Institute's case notes describe what compression does. Test windows were so short that generators never ran long enough to drain their day tanks, so fuel-system faults went unseen.

Contract practice is moving the same way. Law firms advising on data center contracts, including Pillsbury and Bracewell, recommend tying liquidated damages to operational milestones such as energization, integrated testing and the handover of each data hall, rather than to a single substantial completion date. A precon team that prices commissioning as a closeout allowance is pricing the wrong finish line. For how to build it up level by level, see [commissioning is scope](https://piper-ai.com/resources/data-center-commissioning-scope-sub-bids).

## The schedule is set outside the building

On a commercial job the schedule risk sits inside the building. On a data center the building is often the fastest thing on the project. Wood Mackenzie's 2025 survey, reported by POWER, put average lead times at 128 weeks for power transformers. Its 2026 update, reported by Data Center Knowledge, put substation transformers above 160 weeks and switchgear close to a year. JLL reports grid-connection waits of more than four years in primary markets.

Three things follow for precon:

- **Gear is bought before the design is finished.** Long-lead packages are released early, often as limited notices to proceed, and the estimate has to carry escalation to each item's actual purchase date, not to the bid date.
- **Early gear is not good news either.** Equipment that arrives before its building is ready becomes a storage, climate-control, battery-maintenance and re-inspection job that someone has to price.
- **The utility controls a date the contract may hold the GC to.** When power is late, owners bridge it with temporary generation. Whether that is scope, and whose, is a question to settle before award.

Turner & Townsend's 2025 data center cost index found that only 19 percent of respondents were confident suppliers would meet delivery dates. On this building type, [the schedule is part of the estimate](https://piper-ai.com/resources/data-center-schedule-risk-estimate).

## Repetition changes what history is worth

Hyperscale campuses are rows of near-identical buildings built to an owner design standard, bought under a master agreement with a separate work authorization for each phase. That makes the second building different from the first in a way commercial work rarely offers. Phase 2 can be priced from phase 1's actual productivity, change log and commissioning findings, and each work authorization is a chance to reprice.

The catch is that the standard moves. Owners revise their designs as computing hardware changes, and a change from air cooling to [liquid cooling](https://piper-ai.com/resources/data-center-liquid-cooling-scope-boundaries) in the middle of a program brings new steel, new piping and new electrical loads. Repetition is only an advantage to a team that can see exactly what changed between the last building and this one.

## What does not change

Most of the discipline carries over. The shell, site and civil packages behave much like commercial work, which is why [commercial vs industrial estimating](https://piper-ai.com/resources/commercial-vs-industrial-construction-estimating) treats data centers as often hybrid. Proposals still need to be [leveled against the project, not against each other](https://piper-ai.com/resources/construction-bid-leveling-guide). Exclusions and qualifications still need to be read line by line. And [estimating is still risk management](https://piper-ai.com/resources/estimating-is-risk-management-not-cost-calculation): the number is still the output of judgments about uncertainty.

What changes is where the uncertainty lives. It moves from quantities in the building to responsibilities at the handoffs, dates outside the site, and tests at the end.

## What this means for the precon team

A GC entering the sector does not need a new estimating method. It needs its existing method pointed at different targets:

- Build the scope basis from the owner's equipment responsibility matrix and the redundancy topology, not from a square-foot benchmark.
- Level electrical and mechanical packages by responsibility step for every owner-furnished item, not by cover price.
- Price commissioning by level, with named owners for load banks, fuel, manufacturer support and re-tests.
- Tie escalation, storage and general conditions to the procurement schedule, and read which milestones carry liquidated damages.
- Treat each new building in a program as a revision of the last one, and trace what the owner's standard changed.
- Measure subcontractor coverage as crews that can staff the peak, not proposals received ([bid coverage when only a few subs can do the work](https://piper-ai.com/resources/data-center-subcontractor-bid-coverage)).

Every one of those depends on connecting information that usually lives in separate places: the equipment list, the one-line, the commissioning plan, the procurement log, the subcontract exhibits and the estimate.

## Where Piper fits

The difficulty on a data center is rarely a single hard calculation. It is that a change in one place, a new rack generation, a late switchgear release, an owner-furnished unit that arrives without its connecting bus, lands in several others at once: the scope of two subcontracts, the commissioning plan, the procurement dates and the estimate.

Piper is the AI operating system for preconstruction. It builds an understanding of the project from the drawings, specifications, owner standards, equipment lists, proposals and addenda, and uses that same understanding across scope, leveling and review. When a revision arrives, the question it works on is what the change affects, not only what changed, and every finding links back to the sheet or clause behind it. Estimators still decide what to carry and how to price the risk.

### Is data center construction commercial or industrial for estimating?

Usually a hybrid. The shell, site and civil packages behave like commercial work. The power, cooling and controls systems behave more like industrial work, with equipment-driven scope, long lead times and heavy testing. Pick the estimating approach from the cost drivers of each package, not from the label on the pursuit.

### How much does a data center cost to build per MW?

Cushman & Wakefield's 2026 cost guide puts the US greenfield average at $17.6 million per MW of IT load, excluding servers and chips. Treat it as an owner-level benchmark. A GC's own number depends on which equipment the owner furnishes, the redundancy topology and the market.

### What is OFCI on a data center project?

Owner-furnished, contractor-installed equipment. The owner buys major gear such as generators, switchgear, UPS and cooling units directly, and the GC's subcontractors receive, set, connect and help test it. It shifts the GC's bid from equipment pricing to labor, coordination and responsibility at each handoff.

### What are the five levels of data center commissioning?

Level 1 is factory witness testing, Level 2 is delivery inspection, Level 3 is pre-functional checks and startup, Level 4 is functional testing of each system, and Level 5 is the integrated systems test at full design load. Contract milestones and liquidated damages are increasingly tied to Level 5 and to each data hall's handover.

## Sources

- Cushman & Wakefield, [2026 Data Center Development Cost Guide](https://ir.cushmanwakefield.com/news/press-release-details/2026/Cushman--Wakefield-Releases-2026-Data-Center-Development-Cost-Guide-Citing-21-Rise-in-Per-MW-Construction-Costs/default.aspx) (US average cost per MW and cost categories).
- Lenovo Press, [Lenovo NVIDIA GB300 NVL72 rack-scale AI](https://lenovopress.lenovo.com/lp2357-lenovo-nvidia-gb300-nvl72-rack-scale-ai) (rack power and liquid cooling requirements).
- Uptime Institute, [Explaining the Uptime Institute's Tier Classification System](https://journal.uptimeinstitute.com/explaining-uptime-institutes-tier-classification-system/) and [Avoiding data center construction problems](https://journal.uptimeinstitute.com/avoiding-data-center-construction-problems/) (concurrent maintainability, fault tolerance, and construction-phase failures).
- Schneider Electric, [White Paper 75, Comparing UPS System Design Configurations](https://it-resource.schneider-electric.com/white-papers/wp-75-comparing-ups-system-design-configurations) (redundancy topologies and module counts).
- Vertiv, [360AI reference design 015](https://www.vertiv.com/48ecdc/globalassets/documents/ai-hub-reference-design/vertiv-360ai-reference-design-015.pdf) (a block-redundant power architecture for AI racks).
- Construction Dive, [Tips for owners and contractors on data center contracts](https://www.constructiondive.com/news/owners-contractors-tips-data-center-contracts/817235/) (equipment responsibility, master agreements, liquidated damages).
- Pillsbury, [Data center construction contracting](https://www.gravel2gavel.com/data-center-construction/), and Bracewell, [Managing contractual risk in data center construction](https://www.bracewell.com/resources/managing-contractual-risk-in-data-center-construction/) (milestone-based liquidated damages and owner-furnished equipment clauses).
- POWER, [Transformers in 2026](https://www.powermag.com/transformers-in-2026-shortage-scramble-or-self-inflicted-crisis/), and Data Center Knowledge, [AI data center boom rewires US power supply chain](https://www.datacenterknowledge.com/build-design/ai-data-center-boom-rewires-us-power-supply-chain) (Wood Mackenzie lead-time surveys).
- JLL, [Data Center Outlook](https://www.jll.com/en-us/insights/market-outlook/data-center-outlook) (grid connection timelines).
- Turner & Townsend, [Data Centre Construction Cost Index 2025](https://reports.turnerandtownsend.com/data-centre-construction-cost-index-2025/industry-challenges) (supplier delivery confidence).