The Definitive Guide toAI Data Centers
Ask the Guide

Chapter 2.2

Delivery Models & the Owner's Organization

How you carve up the work and who carries each risk is not an org-chart formality — it is a single set of choices that buys (or burns) months of schedule, fixes (or floats) your price, and decides whether the GPU you cannot get late is the contractor's problem or yours.

POWER-BOUNDDENSITY-RAMPGOODPUT

What you'll decide here

  1. Which delivery model you contract under — EPC/EPCM, progressive design-build, traditional design-build, or design-bid-build — because that single fork sets who carries schedule and price risk, how early you can start construction against incomplete design, and whether you get a firm number or a target.
  2. Which owner-side roles you staff as independent (owner's representative, owner's engineer, commissioning agent) versus fold into the contractor — the cheapest line item in the project and the one whose absence shows up as a failed integrated systems test.
  3. Which equipment you furnish yourself (OFE) versus push to the contractor — GPUs and transformers almost always owner-furnished, the rest a deliberate trade of margin for schedule control and a new interface-warranty seam you now own.
  4. Whether you split the base building from the IT fit-out at a powered-shell handoff — and exactly what crosses that line, because an ambiguous shell boundary is the most expensive sentence in the lease.
  5. Whether you optimize for a one-off project or a repeatable portfolio — because design-assist, ECI, and a standardized reference design pay back only across the second, third, and tenth building, and the model that is right for one is wrong for ten.

Chapter 2.1 built the schedule and named the long poles. This chapter decides who is on the hook for each of them. An AI data center is delivered by a coalition — owner, designers, a general contractor or EPC, dozens of trade and equipment vendors, a utility, and the chip vendor sitting outside all of it on an allocation clock no contract can fully tame. The delivery model is the legal and organizational machine that assigns risk, sequence, and price across that coalition. It is chosen once, early, and it is expensive to change — a project re-papered from EPCM to lump-sum EPC mid-design has already lost the schedule it was trying to protect.

We work through the four delivery models and what each one does to schedule, price certainty, and the owner's exposure; the three independent owner-side roles that exist precisely because the contractor cannot mark its own homework; the owner-furnished-versus-contractor-furnished line and the interface warranties it creates; multi-prime versus single-GC and the early-contractor-involvement that bridges them; the powered-shell handoff that splits a building in two; and the portfolio question that quietly reframes every one of these answers. Throughout, the binding constraint is the one from Part 1: this is a power-bound, allocation-bound business, and the delivery model that wins is the one that protects time-to-power and absorbs the long-lead and GPU-allocation risk without pretending a contractor can guarantee what a chip vendor controls.

The four delivery models

The models differ along three axes that matter to an owner: how many points of responsibility you contract with (one throat to choke, or many to coordinate), when the price firms up relative to design maturity, and who carries the gap between the estimate and the actual cost. Read them as a spectrum from most owner-control/least-risk-transfer to least-control/most-transfer.

Design-bid-build (DBB) is the traditional sequence: the owner hires a designer, completes the design, then competitively bids construction to a general contractor on a lump sum. It maximizes price competition and design control and produces the firmest single number — but it is the slowest, because construction cannot start until design is essentially complete, and it is the most adversarial, because every gap in the documents becomes a change order the GC is incentivized to find. For an AI build racing a power clock, the linear schedule is usually disqualifying on its own.

Design-build (DB) collapses designer and builder into a single integrated entity under one contract. The owner gets one point of responsibility, fewer coordination disputes, and the ability to overlap design and construction. The trade is reduced design control — the design-builder optimizes for its own buildability and margin — and a price that is only as firm as the scope was when it was struck. Specialized systems (the IT fit-out, sometimes the GPUs and major electrical kit) are frequently carved out of the DB scope and kept in the owner's hands. (Bracewell / National Law Review, 2026.)

Progressive design-build (PDB) is the model the data center sector has gravitated to for fast-track AI work. The owner selects the design-builder early on qualifications — before a baseline design exists — then collaborates with that team to develop design and budget together, receiving a series of preliminary GMPs at milestones until a final GMP is set as the design nears completion. PDB front-loads the builder's constructability and procurement knowledge, lets long-lead orders and site work start against early design, and keeps the owner in the room for the decisions that matter — at the cost of deferring the firm number and demanding far more owner engagement than a hands-off lump sum. (AIA Contract Documents; FHWA, 2026.)

EPC / EPCM is the heavy-infrastructure model imported from power and process plants, and the one drawing growing appetite as owners look to push risk onto contractors for faster delivery. A lump-sum EPC wraps engineering, procurement, and construction under one contractor on a fixed price with a single point of responsibility — maximum risk transfer, the firmest wrap, and the model lenders like, but priced with a contingency the contractor owns and you pay for. EPCM (the 'M' is management) keeps the owner as the contracting party for the trade and equipment packages while the EPCM firm manages engineering, procurement, and construction as the owner's agent — the owner keeps the risk and the savings, and pays a fee for expertise. The EPC-vs-EPCM choice is, at bottom, the same price-certainty-vs-control fork in heavy-industry clothing. (Bracewell; King & Spalding, 2026.)

Delivery model → risk, speed & price-certainty tradeoffs
Delivery modelPoints of responsibilityWhen price firmsWho carries the cost gapSpeed (overlap)Best fit
Design-bid-build (DBB)Two+ (designer, then GC)At bid, on complete designOwner (via change orders)Slowest — linearMature, repeatable scope where price competition matters most
Design-build (DB)One (integrated DB entity)At contract, on the scope then definedMostly contractorFast — overlappedWell-defined scope; one throat to choke
Progressive DB (PDB)One (selected early on qualifications)Final GMP near design completionShared early, contractor at GMPFastest — collaborative fast-trackImmature scope + hard power clock; the AI fast-track default
Lump-sum EPCOne (EPC contractor)At contract, fixedContractor (you pay the contingency)Fast but needs scope to wrapLender-financed, risk-transfer-first; firm-wrap projects
EPCMMany (owner holds packages)Package by package, owner-managedOwner (keeps risk and savings)Fast — fast-trackableSophisticated owner with in-house controls; portfolio repeatability
Practitioner ranges, 2026. Points of responsibility, price-firming point, and risk allocation are the load-bearing columns; speed ranks are relative for an AI build racing a power clock. (Bracewell / NatLawReview; AIA Contract Documents; King & Spalding, 2026.)

One nuance the table flattens: project size sorts the models in practice. Smaller commercial cloud builds (construction cost under roughly $500M) tend to be procured on a GMP basis, where the contractor absorbs overruns beyond the agreed maximum; the largest hyperscale and self-build programs lean toward EPCM or progressive structures where the owner keeps control of the long-lead packages it cannot afford to hand off. (National Law Review, 2026.) The lesson from Chapter 2.1 carries straight through: the model exists to protect the critical path, and on an AI build the critical path runs through power and long-lead equipment — so the model that keeps the owner close to those packages usually beats the one that wraps them cleanly but slowly. → Chapter 2.1 on the integrated master schedule; the long-lead register itself in Chapter 2.3.

The owner's organization: three independent roles

Whatever the delivery model, the owner needs eyes that are not the contractor's. Three roles recur, and they are independent on purpose — each exists because the party doing the work cannot credibly certify its own work to the owner, the lender, or the future tenant.

The owner's representative is the owner's authority on the ground: the single point of decision who manages the delivery contract, approves changes and payments, holds the schedule and budget, and speaks for the owner in the field. In an EPCM or multi-prime job this role is large; under a lump-sum EPC it is leaner but never absent — someone has to enforce the wrap. The owner's engineer (OE) is the independent technical conscience: it reviews the contractor's design against the owner's project requirements and basis of design, witnesses factory acceptance tests, and protects the owner from a design optimized for the builder's margin rather than the owner's operating life. The commissioning agent (CxA) is the independent verifier that the built plant actually performs — it owns the Level 1–5 commissioning sequence and, critically, the integrated systems test (IST) that proves power, cooling, and controls work together under load before any GPU arrives. → commissioning levels and the IST in Chapter 5.11 and the go-live sequence in Chapter 6.6.

The consequence of skimping here is predictable. These three roles together are a low single-digit percentage of project cost; the failures they catch — a mis-specified CDU loop, an undersized neutral, a controls sequence that oscillates under transient load — are the ones that strand capacity or cause a 1.5 GW load-loss event the grid operator never forgives. The independence is the point: a CxA that reports to the EPC has no leverage when the EPC's schedule is slipping, which is exactly when you need it most.

Owner-furnished vs contractor-furnished equipment

The next fork is what crosses the owner's procurement desk versus the contractor's. Owner-furnished equipment (OFE) is bought by the owner directly and free-issued to the project for the contractor to install; contractor-furnished equipment is inside the construction scope, marked up, and warranted as part of the wrap. The choice is a deliberate trade: OFE strips the equipment margin out of the GC's number and gives the owner direct control of the order date — but it transfers the schedule and interface risk for that item onto the owner, who now owns the gap if it arrives late or does not integrate.

Two categories are almost always owner-furnished regardless of model. GPUs and AI servers are bought directly because the owner holds the vendor allocation, the chips are a financeable asset on the owner's balance sheet, and no general contractor will warrant a delivery date a chip vendor controls. Power transformers and GSUs are owner-furnished because their 2–5 year lead times put them on the critical path before a GC is even selected — the order must be placed against a reservation deposit long before construction contracting closes. On hyperscale work the OFE line extends further: owners routinely buy switchgear, generators, and the cooling plant (chillers/CDUs) directly to capture the margin and control the schedule, then issue them free to the job — which removes equipment margin from the GC and shifts the GC's exposure into labor productivity, install quality, and clean commissioning evidence. (Industry practice, 2026.)

OFE vs contractor-furnished: the equipment-by-equipment decision
EquipmentTypical sourceWhyLead timeInterface-warranty seam the owner owns
GPUs / AI serversOwner (OFE)Owner holds allocation; financeable asset; no GC warrants chip datesAllocation-gated (months)Server-to-rack-power and server-to-cooling fit; rack-level integration
HV / GSU & main transformersOwner (OFE)2–5 yr lead time is on critical path before GC selection~128 wk std, ~144 wk GSU, up to ~60 moTransformer-to-switchgear protection coordination
MV switchgear / generatorsOwner (OFE) on hyperscale; else contractorCaptures margin + schedule control on big programsLong; surgingSwitchgear-to-genset and switchgear-to-UPS interface
Chillers / CDUsOwner (OFE) increasingly; else contractorCooling plant is critical-path and capacity-definingLong; surgingCDU-to-rack loop and CDU-to-facility-water boundary
Building MEP / fit-outContractor-furnishedStandard trades; GC carries productivity riskShort to moderateFolded into the GC wrap
Default 2026 practice for an owner-led AI build; lead times per the long-lead register. The interface-warranty column is the seam OFE creates and the owner must then manage. → lead times in Chapter 2.3, contract seams in Chapter 2.4.

The hidden cost of OFE is the interface warranty. When the owner free-issues a transformer and the EPC installs it, neither party fully warrants the seam between them — if the transformer's protection settings do not coordinate with the contractor-supplied switchgear, the owner is in the middle of a two-vendor finger-point with no single wrap to fall back on. The more OFE you take, the more of these seams you own, and the more the owner's representative and owner's engineer earn their keep stitching them together. This is why OFE is a schedule-and-control decision, not a savings decision: you take it to control the order date on a long-lead item, and you accept the integration burden as the price. → the OFE interface warranties as a contract problem in Chapter 2.4; the procurement mechanics of slot-reservation deposits in Chapter 2.3.

<$500M
build-cost threshold below which GMP procurement dominates; larger programs lean EPCM/progressive
2026National Law Review (Bracewell), DC delivery models
~$11.3M/MW
global average data-center construction cost (base building); tenant AI fit-out adds up to ~$25M/MW
2026JLL via ConstructElements
$10–12M/MW
mainstream hyperscale / build-to-suit construction benchmark before buyer-specific scope
2026JLL / industry synthesis
~128 wk
HV/substation transformer lead time (≈144 wk GSU; up to ~60 mo constrained) — OFE-by-default, pre-GC
2025-Q2Wood Mackenzie / pv magazine
6–12 mo
H100/H200 GPU lead times through 2025–2026 — why GPUs are owner-furnished and on allocation
2026Industry synthesis (Buildermuse / vendors)
~$10–12B
revenue per GW of AI capacity per year — the prize the delivery model protects by saving months (contested — single-source)
2025SemiAnalysis (onsite gas)
Tier IV ~26 min/yr
downtime the commissioning agent's IST exists to prove (vs Tier III ~1.6 hr/yr)
2025Uptime Institute
~1.5 GW in 82 s
data-center load dropped on a single Virginia fault — the failure independent Cx and OE exist to prevent
2026NERC Level 3 Alert / Utility Dive

Multi-prime vs single-GC, and the design-assist bridge

A parallel structural choice runs alongside the delivery model: do you contract one general contractor who subcontracts everything (single-GC), or hold several prime trade contracts yourself (multi-prime)? Single-GC gives you one schedule, one point of coordination, and one party to hold accountable — at the price of the GC's overhead and markup on every package and a layer of distance from the trades. Multi-prime strips that markup and lets you fast-track by awarding the long-lead and critical packages (electrical, mechanical) directly and early — but you, or your EPCM agent, now own the coordination risk that the single GC was being paid to absorb. Multi-prime works for a sophisticated owner with real project controls; it is a trap for an owner without them, because the coordination failures land on the party with no one to subcontract them to.

The bridge between 'design first, build later' and 'build fast against immature design' is early contractor involvement (ECI) and design-assist. Both bring the builder and key trade contractors into the design phase — not to design, but to price, sequence, and pre-order against an evolving design. Design-assist on the mechanical and electrical packages lets the contractor lock the cooling plant and power-chain vendors and place long-lead orders while the architect is still drawing, which is the only way to honor a power clock that started before the design did. ECI is what makes progressive design-build progressive: the contractor is in the room early enough to convert constructability and procurement knowledge into a schedule, not just a price. The consequence of skipping it is the classic fast-track failure — design and construction proceed in parallel but uncoordinated, and the rework at the seam costs more than the schedule the parallelism saved.

Splitting the base building from the IT fit-out: the powered-shell handoff

The largest scope split in the whole delivery question is whether to deliver the base building and the IT fit-out as one continuous scope or as two, joined at a powered-shell handoff. A powered shell is a secure structure with power brought to the site, the building envelope, structural readiness for high-density loads, and cooling pathways stubbed in — but typically without the UPS, generators, cooling plant, and IT infrastructure. (Build.inc, 2026.) The developer builds and delivers the shell; the tenant or operator fits out the inside. This is the same powered-shell instrument that Part 1 framed as a way to preserve IT-fit-out optionality and keep an irreversible decision cheap. → Chapter 1.6.

The split buys two things: parallelism (the shell rises while the fit-out design and GPU allocation are still maturing) and optionality (the operator commits to the inside only when the chips and the workload are real). The cost is a new, sharp interface: exactly what crosses the handoff line, and in what condition. Where does the developer's power delivery stop and the operator's electrical fit-out start? Are cooling pathways merely stubbed or actually capable of the density the operator intends? Who owns the structural rating for wet racks the developer never saw? An ambiguous shell boundary is costly, because every gap becomes a dispute exactly when the operator is racing to energize. The fix is a meticulously specified delivery condition — power capacity and voltage at the boundary, structural loading, cooling pathway sizing, and the conditions that trigger additional landlord obligations — written before the shell is designed, not discovered at handoff. (Build.inc, 2026.)

Deep dive: where the powered-shell line should fall — and the density-ramp trap inside it

The temptation is to draw the shell line at the cheapest place for the developer: deliver a structure and a power feed, and let the operator do everything else. That minimizes the developer's spend and maximizes the operator's integration burden — and, worse, it bakes in a density ceiling the operator discovers too late. The shell's structural rating and cooling pathways are irreversible substrate (Chapter 1.1's reversible-vs-irreversible logic): a shell poured for 40 kW air-cooled racks cannot later absorb a 132 kW liquid-cooled NVL72 generation, let alone a ~600 kW Kyber-class rack, no matter what the operator wants to install. If the developer sizes the shell to today's density to save capital, the operator inherits a building that cannot ride the density ramp — the same density-ramp trap from Part 1, now hiding inside a lease boundary.

The disciplined handoff specification therefore pins the irreversible attributes generously and leaves the reversible ones to the fit-out: the shell guarantees floor loading, structural readiness, water/cooling-pathway capacity, and electrical headroom sized to the operator's ramp, not its day-one load; the operator brings the UPS topology, the specific CDUs, the rack layout, and the accelerator generation. Get this division right and the powered shell is a genuine optionality machine — the developer carries the long-pole power and structure, the operator commits capital only when the workload and the GPUs are real. Get it wrong and you have either a developer who over-built a shell no tenant wants or an operator who leased a building that physically cannot host the workload it was signed for. → density wall in Chapter 5.1; structural design for dense liquid-cooled halls in Chapter 6.2; the lease terms that paper this boundary in Chapter 2.4.

The decision drivers — and the portfolio reframe

Six drivers decide which way each fork above breaks, and they do not point the same direction. Schedule pushes toward progressive design-build, fast-track, ECI, and OFE on the long poles — anything that overlaps work and protects time-to-power. Cost certainty pushes the opposite way, toward a complete design and a lump-sum wrap a lender can underwrite. Scope maturity is the tiebreaker: a mature, well-understood scope can be bid lump-sum cleanly; an immature one cannot be wrapped honestly and must be delivered progressively. In-house capability decides how much risk the owner can actually hold — EPCM and multi-prime hand the owner the savings and the risk, and only an owner with real project controls should take them. The fifth and sixth drivers are where the analysis changes shape.

Repeatability across a portfolio reframes everything. The model that is right for one building is wrong for ten. A single project optimizes for the certainty and speed of that one asset; a portfolio optimizes for a standardized reference design built repeatedly with the same delivery team, where design-assist, ECI, and OFE master agreements amortize across buildings and the learning curve drives cost and schedule down with each repetition. A hyperscaler running a multi-GW program will invest in an EPCM-style structure with in-house engineering and a templated design precisely because the per-project overhead is repaid across the fleet — the same investment is dead weight on a one-off colo shell. The honest first question is therefore not 'which model for this building' but 'am I building one or building a program' — because that answer reorders the other five drivers before you weigh them.

This chapter assigns the risks that Chapter 2.1 scheduled. The long-lead equipment whose lead times force the OFE decision is registered in Chapter 2.3; the contract stack that papers every seam here — EPC terms, OFE interface warranties, the powered-shell lease, liquidated damages — is built in Chapter 2.4; the project finance and SPV structures that constrain which delivery model a lender will accept live in Chapter 2.5; and the builder's-risk and delay-in-startup insurance that overlays the whole delivery in Chapter 2.6. The commissioning agent's integrated systems test is engineered in Chapter 5.11 and sequenced in Chapter 6.6; the powered-shell optionality thesis traces back to procurement in Chapter 1.6 and the density-ramp substrate to Chapter 5.1 and Chapter 6.2.