Why AI Data Centers Are So Expensive

Data centres are now priced and managed in megawatts, not square feet. April 2026 benchmarks put a standard build at about $11.3M per MW, with AI-optimised facilities clearing $20M+ per MW at 40kW to 80kW rack densities. That shift forces a different commercial stance from day one, because most of the cost and programme risk sits in power delivery, heat rejection, and commissioning, not the shell.

Energisation is the date that matters, not Substantial Completion. A one-month slip on a 60MW facility costs the owner about $14.2M in lost revenue, around $3.55M per week. Owners react by tightening decision cycles, pulling long-lead lock-in forward, and demanding earlier commissioning access. Contractors get squeezed if they run RFIs, submittals, and changes on normal “commercial” timelines.

Square-foot budgets under-scope the MEP package, then punish the delivery team for the gap. A shell can look cheap on $/sq ft, but the job lives or dies on switchgear, generators, busway, cooling loops, controls, and test and commission. Most margin loss shows up late, when equipment is already bought and commissioning windows are blocked by access, sequencing, or incomplete prerequisites.

Run bids and live jobs on three numbers: $/MW, date-to-energisation, and delay cost per week. Put the delay cost next to every long-lead release and every commissioning gate, then push decisions upstream. Archdesk is built for that weekly control, linking packages, change, and cost-to-complete back to the energisation milestone so the team sees drift early enough to act.

The gap comes from a topology change in power and cooling, not from the shell. Move from 10kW to 60kW per rack and you stop building a “heavier” data hall. You start building a process plant with a live electrical environment wrapped around it. That shift drives bigger cost steps and longer test windows, and you can’t value engineer it late without breaking the operating model.

Thermal plant is the fastest cost escalator, and it shows up as soon as you move beyond air. A traditional air-cooled hall rejects heat at roughly $1.8M per MW. A liquid-cooled AI hall with CDUs, secondary pumping loops, and dry coolers sized for near 100% sensible load runs $4.5M to $5.2M per MW. That change also drags in water quality logs, flushing, and leak evidence that the operator won’t waive at handover.

Commissioning is where margins get written off. A standard facility runs Integrated Systems Testing (IST) in 4 to 6 weeks. Liquid-cooled AI halls stretch IST to 10 to 14 weeks because you have hydraulic balancing, leak detection on thousands of quick-connect points, and staged thermal load tests. Treat commissioning as “end of job” and you pay twice: once for rework, then again for lost windows and remobilisation.

Build your cost plan around the commissioning plan, not the drawings. Lock each long-lead item to a test script and a prerequisite list, then track it weekly. Archdesk teams do this by tying submittals, cost codes, and programme constraints to each system, so problems show up early enough to fix without blowing the test window.

MEP isn’t “behind the shell” on an AI data centre. It is the job. On a typical 60MW facility, electrical, mechanical, controls, life safety, and commissioning make up most of the installed cost. The GC can hit every pour and still miss the real finish line, power and cooling proven under test.

The critical path sits inside prerequisites, not partitions and doors. Protection coordination studies, points lists, labelling, water quality and flushing evidence, and access to live rooms decide whether IST passes. Most programme slips come from “no-fault” gaps between trades, like CT orientation, valve tagging, or trend logs that don’t meet the test script.

Buyout needs to match where the risk sits. Split the Schedule of Values into long-lead equipment, distribution, cooling, controls, and commissioning, then set clear release gates so you don’t argue later about what was “included”. Archdesk helps here because it links procurement, valuations, field evidence, and cost-to-complete, so commercial and delivery teams see the same impact when design, supply chain, or test readiness moves.

Long-lead procurement sets the energization date. Utility transformers and medium-voltage switchgear sit in factory queues for 12 to 22 months after submittals are approved. That kills the old assumption that steel and concrete drive the schedule.

A late submittal doesn't just push a delivery date. It bumps you into the next engineering review slot, the next Factory Acceptance Test (FAT, the witnessed factory test before shipping) slot, then the next logistics and utility outage window. Miss one window and you lose 6 to 8 weeks before a new slot opens. Those weeks don't compress later in the schedule.

Lead time isn't one number. It's separate clocks running across OEMs, utilities, and multiple trades, and all of them have to line up. The one most teams miss is FAT booking. Transformer OEMs schedule witnessed tests months in advance. If your submittal package isn't complete and approved before that booking window opens, the slot goes to another project.

The decisions that protect your schedule have to be made at 30% design, not at buyout. Transformer specifications, paralleling schemes for generators, and BMS point lists all require engineering sign-off before an OEM will accept a purchase order. Teams that wait for 60% or 90% design drawings before issuing RFPs are already 4 to 6 months behind on the critical path, even before a purchase order is cut.

Prefabrication won't save you if room readiness isn't locked to the delivery date. Skids arrive because the factory is done. They then sit because pads aren't signed off, access is blocked, or temporary power isn't available for pre-energization checks. The cost shows up as laydown storage, re-handling, warranty exposure from outdoor storage, and lost commissioning time. None of it appears as one clean change order, which is why it rarely gets captured accurately in post-mortems.

Build a procure-to-energize timeline, not a buyout log. Tie each critical asset to an activity in the master schedule, with one owner and one next action. Archdesk links submittals, purchase orders, FAT dates, delivery confirmations, and install-readiness checks in one live record, so slips surface in days rather than at the monthly schedule review when it's already too late to recover.

Grid interconnection is the new master programme, and it behaves like a long-lead package with heavy volatility. California’s interconnection queues averaged 46 months in 2023, according to Lawrence Berkeley National Laboratory’s “Queued Up” (2024). Treating “utility power” as a permit milestone leaves site teams blind to the real risk. Treat it like switchgear. Demand dated milestones, named scope, and signed interfaces.

Developers are moving out of legacy hubs because “date-certain megawatts” beat “best historic location”. Dominion Energy filings in 2025 show Northern Virginia interconnection queues at 48 to 60 months for large loads. AEP and AES filings in 2025 put central Ohio and Indiana at 18 to 24 months. That shift changes who wins work. Contractors that can plan phased handovers and manage utility interfaces like subcontractors get repeat programmes.

Fast interconnection often comes with reliability and curtailment exposure, and that lands in construction scope. ERCOT issued 14 conservation alerts in summer 2025. Owners react by pulling more resilience forward. They ask contractors to build more on-site generation and controls earlier, so factory testing and commissioning can run without waiting on a stable grid window.

Phased energisation is the practical response, and it changes how you run the site. Developers are splitting large campuses into 20 to 30MW blocks, each with its own tie-in date and test pack. That creates overlapping programmes in one footprint, with shared cranes, corridors, and shutdown windows. The practical move is to run each phase as its own critical path, then manage shared logistics as a separate workstream.

Archdesk helps teams run this as a control problem, not a hope-and-pray programme. Link utility dates, OEM submittals, and commissioning prerequisites in one plan, then measure slippage weekly. Treat the utility scope like a subcontract package, and manage evidence the same way you manage FAT results and test sign-offs.

The cost story from 2020 to 2026 is an inflection, not a straight line. Standard builds crept up year by year. AI-tier builds broke away after 2023, and that’s where tender risk sits. If you’re still pricing off “last job plus inflation”, you miss the step-change and you buy the problem at GMP.

MEP-plus-commissioning now carries the commercial job. Draft package splits show M&E at 82% of value on an AI-optimised build, versus 59% on a standard build. That changes how you manage risk. The package that holds the money should also hold the weekly forecast, the change log, and the test evidence.

Grid capacity now decides where work starts, and it shifts your labour and logistics plan. Public filings and LBNL queue data show multi-year waits in some established markets. Owners respond by phasing campuses into 20MW to 30MW blocks with separate tie-ins. That creates more mobilisation and more interface risk, even when the headline MW stays the same.

Practical takeaway: run your board pack in three lines. Track package-level forecast weekly for MEP and commissioning. Track long-lead equipment by test slot and utility window, not by PO date. Archdesk is built to tie budgets, procurement, and closeout evidence together at package level, so the job doesn’t drift while teams wait on the next prerequisite.

Energization is the milestone that gets paid attention to, because it proves the site can run safely and hold load. Substantial Completion only says the building is “done enough” under the contract. The schedule risk sits in the gap between “MEP looks complete” and “operator accepts performance”, and that gap is usually paperwork and test readiness, not missing hardware.

Documentation is part of the build on these projects. Pre-functional checklists stall because submittals are not tied to asset tags and serial numbers. Functional performance testing then slips because the controls point list is still moving. Integrated Systems Testing (IST) turns into a design meeting when sequence of operations changes have not been pushed through every vendor, especially MV gear, UPS, generators, and cooling plant.

Load bank work fails for boring reasons that still cost weeks. Banks and fuel get booked, then temp works, cable routes, and protection settings are not frozen. Controls tuning becomes re-test churn because trend logs are missing or not time-synced. MOP/SOP readiness, the Method of Procedure and Standard Operating Procedure the operator uses to run live power, is the last gate and it blocks turnover if it does not match the latest as-builts.

The practical move is to manage commissioning evidence like cost. Archdesk teams link each system, tag, submittal revision, test script, and punch item to a gate on the programme. Run a weekly question that forces the truth: “How many systems can pass IST this week with a complete turnover pack that the operator will accept?”

Archdesk treats procurement, cost, and commissioning as one control loop, not three separate admin tasks. A 2% overrun on a $500M MEP package is $10M. That wipes out most GC fee positions. It happens fast once installs start, because recovery options narrow by the week.

This risk is sharpest for GCs now bidding Ohio, Indiana, and Texas giga-campuses, where package values are running $400M to $700M MEP on a single site. Monthly cost reporting doesn't catch drift early enough. By the time the lag clears, the commercial team has lost its window to act. Archdesk runs cost-to-complete weekly, against SOV (Schedule of Values) rules that are set at award, not invented mid-job.

Archdesk also protects the program by managing long-lead plant by gates, not by "PO sent" or a single lead time. Submittals, relay settings, witness points, FAT (Factory Acceptance Test) slots, and ship windows are schedule events. Archdesk ties each event to dated evidence, an owner, and a next action. "Bought" stops being a false sense of safety.

Archdesk wins at handover because quality traceability is built into day-to-day delivery, not reconstructed at the end. Test results that live in emails and shared drives are a liability. Owners in Ohio and Texas are now requiring digital as-built records at energization, not 90 days after. Archdesk links each test result to an asset tag, location, and approved document set. The handover pack builds itself as the work progresses.