2026 US Data Center Construction: $85B Market Breakdown

Data centers are now big enough to move MEP pricing outside the sector. US data center put-in-place spend is about $85.3B in 2026, up from about $45B in 2024 (US Census put-in-place categories, with "office/industrial" reclassified by industry trackers). That step-change is why switchgear quotes are being held for days, not weeks, and why good MEP firms are selective about what they take on.

The cost spike is sitting in electrical distribution, cooling plant, and controls, not the frame. Across Archdesk-managed digital infrastructure jobs, the fastest cost growth is in LV and HV distribution, mechanical plant, and BMS and controls. Treat these packages like long-lead, not "later-stage fit-out", or you'll buy the risk you didn't price.

Geography is now a procurement decision, not just a business development one. Northern Virginia and Texas still set a lot of "national" pricing, but grid and interconnect queues are pushing work into Indiana, Ohio, Iowa, and Wisconsin. If you carry one allowance for freight, sales tax, and AHJ inspection time across all states, you're donating margin on the markets doing the growing.

Project controls capacity is the bottleneck, not forecasting. Owners pay for certainty, so the best planners, BIM coordinators, and procurement QSs are being pulled onto data halls and away from hospitals, labs, and airports. Teams that protect margin run weekly cost-to-complete and committed-cost reviews by package, tied to submittals and long-lead dates.

Owner concentration has turned data centres into a “rules of engagement” market. A small group now sets the specs, the programme gates, and the contract posture that everyone else copies. That matters because you’re no longer just pricing a building. You’re pricing an owner playbook, including how hard they drive commissioning, how much risk they push down, and how much scope they keep for themselves.

Two owner models now drive how you make or lose margin. Hyperscalers repeat a tight kit of parts and enforce hard design freeze and turnover gates. AI-native owners build denser compute, move faster, and push more “unknowns” into the MEP package. You can’t bid both with the same prelims, procurement plan, or commissioning allowance.

OFE changes your commercial plan more than MW does. On hyperscale work, owners often buy major electrical and mechanical kit directly, then issue it to the project as free-issue. That removes equipment margin and shifts your exposure into labour productivity, coordination, install quality, and clean commissioning evidence. Colo developers more often keep procurement in contractor scope, so the same sized hall can carry very different cash risk and change control.

Price the owner, then price the job. Build your tender pack around three items before you touch rates: who owns the equipment (OFE vs contractor buy), what the commissioning gate looks like (what must be proven, and when), and how change is agreed (who signs, and how fast). Archdesk helps teams keep committed cost and progress at cost-code level tied to turnover and commissioning tasks, so scope drift shows up in days, not at month-end.

Speed on these jobs is won or lost before you mobilize. Design freeze now sits three to five months earlier than most teams are used to, and any change after freeze triggers a procurement cascade that adds 6 to 10 weeks to the critical path. The contractors hitting 16-to-20-month delivery targets are the ones holding design freeze discipline, not the ones pouring concrete fastest.

Prefab buys you field time, but it makes long-lead slippage more expensive. Modular electrical rooms and cooling skids arrive as factory-tested assemblies, cutting on-site termination work by roughly 44%. But the sequencing logic is unforgiving. One late switchgear lineup blocks energization and idles every downstream trade simultaneously. Commissioned cooling skids sitting on a slab waiting for a missing transformer are not a schedule problem — they are a $2.8M-per-month carrying cost on a 50 MW site.

Procurement hasn't compressed. It has moved earlier and become binary. Miss a submittal approval window and you don't just lose a week on site — you lose the factory slot that set your entire sequence. On medium-voltage switchgear at 22-month lead times, a missed slot means slipping the energization date by a full production cycle. That is not recoverable through overtime or re-sequencing.

Commissioning is the phase that exposes every earlier compromise. Modular assemblies arrive pre-tested to factory standards, but integrated systems commissioning — where cooling controls talk to power management, and both talk to the BMS — cannot be prefabricated. That sequencing still runs on site, in order, and it compresses badly when upstream packages are late. A two-week delay to switchgear energization doesn't cost two weeks on the commissioning program. It costs four to six, because the integrated testing sequence restarts from the delayed milestone.

Grid interconnection now sets the schedule, even when the build itself is straightforward. A 2024 Lawrence Berkeley National Laboratory study found the median time from interconnection request to commercial operation was about 5 years for projects that came online in 2023. That is not a planning assumption. It is a commercial risk you need to price before you commit to a guaranteed date.

PJM is the worst-case zone right now. Its queue carries over 260 GW of requests against a grid that serves Northern Virginia, the Mid-Atlantic, and the upper Midwest. Average interconnection lead times in PJM have stretched to 60 months as of 2025 queue data. That is the region hosting the largest concentration of hyperscale data center capacity in the country, and it has the longest queue of any ISO/RTO. Projects entering PJM's queue today should not assume power before 2030. CAISO, which serves Silicon Valley and the broader California market, runs at 30 months average, but that number masks local transmission constraints that push individual project timelines well past 48 months in congested areas. Both regions fail the 48-month threshold test for any project bidding 2026 or 2027 completion.

Queue position does not equal an in-service date. Utility restudies and network upgrade scope changes push dates back even for projects already deep in the queue. Every restudy triggered by a withdrawal ahead of you resets your cost allocation. PJM's own process reform documentation acknowledges that withdrawals cascading through the queue are one of the primary drivers of study delays. Build a probability plan, not a single date. Tie procurement gates and commissioning milestones to a P50 and P90 energization date, with cost exposure shown against each scenario.

Phased energization is the margin killer most electrical subcontractors fail to price correctly at bid stage. Teams bring halls online on temporary generation, then return for permanent power changeover, protection studies, witness testing, and contract close-out paperwork. Each return is a mobilization cost. A 2025 industry benchmark by a major cost consultancy put the uplift at 15% to 20% on electrical subcontract costs on affected sites. That is not a contingency figure. It is a baseline on any project in a gridlock zone, and it needs to sit in your Schedule of Values from day one, not surface as a change order after the fact.

Build a one-page power readiness forecast and review it weekly with ops and commercial. Tie utility milestones, procurement status, and commissioning sequence to committed cost so the exposure is visible before it becomes a loss. If you can't explain the energization date, the fallback plan, and the cost at risk on one page, you're carrying grid exposure you haven't priced.

Time-to-power now decides where projects go, and it changes the risk you carry. Constrained metros can give you a site and a permit, but they can’t give you an energisation window you can build a programme around. Inland states win work because utilities will commit to a scope and a sequence, even if the location is worse for latency.

Inland work doesn’t fail on price, it fails on labour logistics. You import the talent, then you pay for travel, per diem, and the productivity dip while crews learn the site, local rules, and the commissioning sequence. Treat that dip as a named line item, or it will show up later as “lost hours” you can’t recover.

Cost per delivered megawatt is the board KPI that moves fastest in MEP and commissioning. Elemental plans on 2026 AI-focused bids put MEP and controls at about 75% of the GMP, with electrical distribution alone at 32–38%. Structure your estimate and your weekly reporting around a power-and-cooling programme, not a shell-and-core build. If you don't, your cost model is tracking the wrong 25% of the job.

The jump from $7M/MW in 2020 to $11M–$20M/MW in 2026 isn't one cost driver. It's four. Electrical density has moved from 6–10 kW per rack on standard colocation builds to 30–80 kW per rack on AI training facilities. That means heavier busbars, larger UPS blocks, and redundant power trains that didn't exist on the previous generation of bids. Liquid cooling adds a second mechanical plant with its own commissioning regime. And scripted functional testing at 60MW scale now takes four to eight weeks of dedicated engineering time. Each of those four factors has a separate bid tab line. If your estimate treats them as one blended MEP rate per square foot, you'll bid short before you've signed the contract.

Most overruns start as unit-rate drift you can see early if you track the right measures. Change orders on one-line diagrams drive termination counts, containment rework, and arc-flash protection revisions that compound fast. Liquid cooling loop reroutes add weld count and pressure test cycles. The miss is measuring progress by area completed while the money is sitting in switchgear line-ups, coolant distribution units, controls points, and scripted testing sequences. By the time area-based progress flags a problem, the commissioning slip is already three weeks in.

Long-lead procurement is now a change-control problem, not just a scheduling one. High-capacity transformers and liquid-cooling components are running 12–18 months from order to delivery. If a change order revises the one-line diagram after those items are committed, you're not just managing a cost adjustment. You're managing a potential re-order, a redesign, and a schedule hit that compounds through commissioning. Your change-control process needs a procurement impact check built in before any design change is approved, not after. The firms getting this right are flagging committed cost against long-lead packages in real time, so the commercial team sees the exposure the same day the change order is opened, not at month-end close.

Run cost and progress at system level every week. Tie committed cost and earned value to the same work package codes in your Schedule of Values. Add a commissioning-readiness KPI that counts scripted tests passed, not punch list items closed. Those are different things. A closed punch list item means someone signed off a defect. A passed scripted test means a system is ready to carry load. Archdesk tracks procurement status, subcontractor valuations, and earned value in one live view, so the gap between "installed" and "ready to test" is visible before it becomes dead time on the commissioning program.

Delay stops being "schedule risk" the moment the owner has phased handover, availability credits, or a ramp-to-load commitment. Cash starts leaking before Substantial Completion because the owner's revenue clock is tied to energization, IST sign-off, and usable white space, not your punch list. A 60MW facility running at $0.24/kWh colocation rates at 85% utilization generates roughly $14.2M per month in revenue. Every month that capacity sits dark, that money is gone. Put the commercial exposure next to the lookahead plan, or the site team will treat float as optional.

Construction-side delay costs stack quietly, then hit your margin all at once. Extended general conditions, re-procurement when quotes expire, storage and re-inspection on held equipment, and resequencing costs in commissioning weeks usually do more damage than the headline liquidated damages exposure. A single month of extended general conditions on a large MEP package typically runs $800k to $1.4M before you add re-procurement escalation on long-lead gear. Track the triggers weekly, not at month-end, because most of these costs start building before the schedule shows a clear slip.

Most preventable slips happen at the back end, inside MEP integration and commissioning. Late design changes hurt most after freeze because every change forces new submittals, rework, and re-test. Failed Integrated Systems Testing (IST), where all mechanical, electrical, and controls systems are run together under simulated load, is worse. You don't just fix one issue. You re-run the integrated scripts across multiple systems and rooms, and every failure resets the clock. IST failures on power date slips, the single biggest driver on the chart above, average 3.8 months of slip and sit closest to the $14.2M/month exposure line for a reason: they occur after all other costs have already been committed.

Practical takeaway: build a delay cost model at bid stage, then run it weekly in delivery. Tie long-lead release dates, access dates, and commissioning readiness to a live cost-to-complete view. Archdesk tracks earned value against committed costs in real time, so the gap between what's been spent and what's been built surfaces every week, not at month-end. "We're 80% done" can't hide the last 20% that blocks energization and handover when the numbers update daily.

Most margin leaks happen between the day you commit and the day finance sees it. Expedite fees, temporary power, idle crews, and commissioning overtime get agreed fast to protect programme. They often hit the job as scattered costs with no clear owner. If you only reforecast at month-end, you find out after the damage is locked in.

Control Tower stops that drift by running three numbers together, every day: earned value from installed quantities, committed cost from every PO and subcontract, and forecast-at-completion against the Schedule of Values. That matters because data centre work is system-led, not area-led. “Percent complete by room” can look fine while switchgear release dates, controls scripting, and test gear are already pushing you into overtime and rework.

Long-lead control fails when “ordered” gets treated as a safe status. Real risk sits in submittals, fabrication slots, factory testing, delivery windows, and room access. Control Tower links each long-lead package to the lookahead plan with a constraint tag. Site and commercial teams see the same gate before it turns into idle labour, double handling, or commissioning stack.

Run Control Tower on one live job first. Pick five system measures you can count each week and tie them to valuation, labour hours, and committed cost. If the view doesn’t tell you where margin moves next week, not last month, change the measures. Archdesk is built to make that weekly control routine stick across every site.