The Data Center Boom: What It Means for Commercial Construction

Part 1Who’s Paying for the Buildout

Three groups pay for the buildout: big cloud companies (“hyperscalers”), AI labs, and the developers and lenders who build and finance the buildings.

Hyperscalers: Four Companies Are Funding Most of the Boom

Amazon, Microsoft, Google and Meta guided to more than $735 billion in combined capital spending for 2026. That’s up from about $640 billion in January. In late April, Evercore and Bank of America projected more than $1 trillion in 2027. Add Oracle, which is building heavily for OpenAI, and the five biggest spenders are projected to top $800 billion this year, more than those companies’ combined operating cash flow for the first time (Brookings). The gap between what they spend and the cash their businesses generate has to come from somewhere, and that somewhere is debt.

Foundation Labs: Major New Buyers That Lean on Big Tech’s Credit

The AI labs are the newest big buyers, and most of them can’t finance a data center on their own. For its four-year, $500 billion Stargate data center program, OpenAI is relying on SoftBank and Oracle to supply most of the money and the buildings. Its flagship campus in Abilene, Texas, is partly running. Last November, Anthropic committed $50 billion with developer Fluidstack for data centers in Texas and New York. xAI lined up a reported $20 billion last fall, much of it debt, to buy chips for its Colossus 2 campus in Memphis.

Lab projects are less certain because the labs depend on raising new money to pay for them. In March, OpenAI and Oracle dropped a planned 600-megawatt expansion at Abilene after financing talks dragged, and OpenAI said it would put that capacity at other sites (Bloomberg). On a lab-driven job, the lab’s name matters less than who is financing and leasing the building. A project backed by a profitable company like Microsoft or Meta, which can keep paying even if one AI bet disappoints, is on firmer ground.

Developers and Debt: Much of the Buildout Runs on Borrowed Money

A developer builds a data center and rents it out instead of using it, much like a landlord. Firms like Vantage, CyrusOne, Stack and Crusoe buy the land, secure the power and pay for construction, mostly with borrowed money, then sign a tenant such as Microsoft to a long lease.

Most of what developers build is still for big tech. About 40% of hyperscaler capacity worldwide is leased rather than owned (Synergy Research Group), and five big tech companies have about $1.16 trillion in signed leases that haven’t started yet (Reuters). Rent on most leases doesn’t count as capital spending, so the capex numbers above understate how much construction big tech is driving.

Neoclouds, companies like CoreWeave that rent out AI computing power, made up about 11% of North American data center demand in the first half of 2026 (JLL). Even that often traces back to big tech: CoreWeave’s biggest customers are Microsoft, OpenAI and Meta.

Much of this runs on debt. Goldman Sachs Research estimates about a third of big tech’s 2026 capital spending is debt-financed. Construction is even more leverage-driven, because hyperscalers tend to pay cash for chips and finance the shells, power and real estate with project debt and leases (Brookings). Meta’s Hyperion campus in Louisiana, a joint venture with Blue Owl, was about 90% funded with debt: roughly $27 billion, against only about $2.5 billion of equity.

CoreWeave showed this fall what happens when lenders get nervous. Carrying about $52 billion of debt and lease obligations, it saw one of its bonds briefly yield about 15%, a level investors treat as distressed, and its stock fell about 15% in a single day.

If you’re building for a developer, you’d rather that developer’s tenant be a hyperscaler than a neocloud or AI lab. The developer signs your contract and pays your invoices, but its lenders are really betting on the tenant. A long-term lease with Microsoft or Meta keeps that financing solid. A lease with a neocloud that borrows to pay its own rent is much shakier. Before mobilizing, find out who the tenant is and who stands behind the lease.

Part 2Constraints and Risks in the Process

Six things decide how fast the buildout moves: power, gear, chips, labor, permitting and debt. Power usually sets the critical path, and the rest pile on delays of their own. The contractor controls none of them.

Power: The Grid Can’t Keep Up

Power is the biggest constraint, because a data center can’t open until the utility can serve it. Data center developers have asked U.S. utilities for about 1,066 GW of power, equal to about four-fifths of all the power plant capacity in the country, and Wood Mackenzie expects utilities to commit to only about 28% of it. Utilities’ own reports show the scale:

• AEP: 69 GW of contracted load growth through 2030, about 90% of it data centers.
• Dominion (Virginia): more than 53 GW in various stages of contracting, but only about 12 GW under signed electric service agreements.
• ERCOT (Texas): about 474 GW of new load requests, more than five times the state’s record peak. In August the governor ordered an audit of every data center seeking a grid connection.

Getting connected is the slow part. The average grid wait in major markets is four years or more, and as long as seven years for a 50 MW data center in Northern Virginia (JLL), while the building takes about 18 months. Northern Virginia still has about 2.4 GW under construction (CBRE) because those projects got in line years ago. On most of these jobs it’s the utility, not the building, that holds things up.

So the work follows the power. Through July, Indiana, Illinois, Michigan and North Carolina each topped $10 billion in 2026 data center starts, with Texas fifth at $9.9 billion (ConstructConnect). JLL says frontier markets like West Texas, Ohio, Louisiana and the Carolinas now hold 77% of the capacity under development. Labor pools in those markets aren’t sized for gigawatt campuses, so some contractors build worker camps to house traveling crews (Bloomberg).

Even with a utility deal, a building can finish before its power is ready. Then commissioning waits or runs on temporary generation, equipment sits with warranties running, crews demobilize, and closeout and final payment stall on milestones that depend on the utility.

Gear: Owners Are Locking Up Scarce Equipment

Electrical gear is scarce, and owners are locking it up before contractors can get to it. Transformer lead times went from about 50 weeks in 2021 to about 128 on average by mid-2025 (Wood Mackenzie). NEMA says transformer backlogs have eased some, but switchgear is still tight (about 44 weeks on average in Wood Mackenzie’s mid-2025 survey).

Owners are reserving factory capacity directly. Last fall, Digital Realty signed a $373 million deal with Schneider Electric for switchgear, UPS systems and prefabricated power skids, including a dedicated production line. Switch signed a $1.9 billion supply deal with Schneider for power modules and cooling equipment. For the contractor, preconstruction becomes procurement. When the owner buys the gear, you don’t carry the price risk, but you don’t control when it shows up either, so receiving, storage, warranty start dates and late-delivery responsibility need to be spelled out in the contract.

Chips: The Design Can Change Mid-Build

The chips change faster than the buildings, so the design can change while you’re building it. Nvidia now releases a new generation about every year, and each one needs more power and cooling. A server rack that drew under 10 kW a few years ago can now draw about 120 kW, enough to require liquid cooling. By 2027, Nvidia expects 600 kW. At least one major operator switched from air to liquid cooling partway through construction. That shows up as redesigns, change orders and resequenced work in the mechanical and electrical scope, where most of the money is. A fixed price on a design that’s still moving is a real risk, which is why EMCOR says more of these jobs are moving to GMP contracts.

Labor: There Aren’t Enough Skilled Workers

The industry doesn’t have enough skilled workers, especially electricians. ABC estimated the industry needed to attract about 349,000 new workers in 2026. Data center construction work pays roughly 25% to 30% more than comparable jobs elsewhere (Skillit and The Wall Street Journal), and in this year’s AGC and NCCER survey, 58% of firms that did data center work in the past year said those projects are increasing competition for skilled workers. The shortage is sharpest at the end of the job, in commissioning, where specialists test every system before the owner accepts it.

Permitting: Approval Is No Longer a Given

Local approval used to be close to a formality, but that’s changing fast. New York put a moratorium on large data centers this summer, a Texas audit ordered in August has slowed new grid connections, and in October Charlotte extended its pause on new data centers for another 11 months. In a March poll, Gallup found 71% of Americans would oppose an AI data center in their area. A project can now stall even with financing and power lined up.

Debt: The Risk That Shows Up Last

Financing tends to be the last risk to surface, often after the work has started. Project-level leverage is high, the debt is often tied to a single tenant, and it runs for decades against chips that may be outdated in a few years (Brookings). If the tenant leaves or the money dries up, the project can stop mid-build.

Contractors are already seeing that risk in their backlog, which Part 3 covers.

Where It All Leads: The Work Will Continue Through 2027, No Matter What

Add it all up and the industry is so far behind that even if the AI boom cooled tomorrow, data center builders would still be busy through 2027. A J.P. Morgan analysis reported by The Wall Street Journal found more than 60% of the capacity planned for 2027 wasn’t under construction this spring, and another 7% was already late. SynMax satellite monitoring found that of the 37 GW IIR Energy expects in 2027, roughly 20 hadn’t started land clearing as of Aug. 4.

It takes about 18 months to build a data center, so anything not already under construction won’t finish in 2027. Some of those projects will die. Most of the demand will resurface somewhere else, so the work slides into 2028 and stacks up where the power is.

On a data center, the GC is no longer the main driver of the finish date. Power, chips, gear lead times, labor and permits set the schedule, and the contractor controls none of them. That changes the GC’s job. The value is in seeing those outside risks early, having the work ready when each one clears, and protecting the firm when one slips.

Part 3How This Affects GCs

Whether or not a firm builds data centers, the boom is changing its business. The place to start is how much of the money actually reaches construction.

Where the Money Goes: Most of It Buys Chips, Not Buildings

Construction gets a smaller share of the boom than the headline spending numbers suggest. Amazon, Microsoft, Google and Meta spent about $57 billion a month worldwide in Q2, while all private U.S. data center construction ran about $7 billion a month in August (an $85 billion annual pace, per the Census Bureau). The two figures don’t measure exactly the same thing, but the gap is clear: most of the money buys chips, not buildings. Bernstein estimates the GPUs alone are close to 40% of an AI data center’s total cost, and adding the servers and networking around them brings the equipment to just over half.

Data center construction is mostly MEP: about 50% electrical and 22–33% mechanical. The shell is only 9–14% of construction cost (Turner & Townsend). These are electrical jobs with buildings around them, which is why so many of the risks in Part 2 land in the electrical and mechanical scope.

How a Data Center Job Is Different

Because of where the money goes and the constraints in Part 2, a data center job runs very differently from a typical commercial job:

Who’s Building: The Biggest Firms Are Getting Most of the Work

So far, the work is concentrated among the biggest contractors. ABC found 42% of contractors with more than $100 million in revenue have data center work, compared with 7% of smaller firms. The leaders show how fast their businesses have changed:

• Turner: about 37% of backlog in data centers; 2025 revenue up 40%.
• Mortenson: revenue up about 60%, from $6.7 billion to $10.8 billion, putting it in ENR’s top 10.
• EMCOR: mechanical revenue from network and communications work, which includes data centers, more than doubled in Q2.
• Comfort Systems: record $17.8 billion backlog at the end of September; technology work, mostly data centers, was 58% of first-half revenue, up from 40% a year earlier.
• Sterling Infrastructure: known for highways and heavy civil work; its E-Infrastructure business, which does site, electrical and mechanical work for data centers and factories, made up 78% of second-quarter revenue.

The gap shows up in backlog, the months of signed but unfinished work a contractor has lined up. Contractors with data center work averaged 11.3 months of backlog in September, about three months more than everyone else (ABC). Smaller firms mostly get in as subs, and some owners now line up the electrical contractor first and pair it with a GC.

How Data Center Builders Are Changing: Better Methods, More Strain on People

Data center work is changing the firms that do it, and not all of that is bad. The work pushes firms toward prefabrication, tighter coordination and better use of technology, and those habits carry over to the next hospital or lab. A Stratus survey of MEP contractors found 88% of data center work now involves some prefab.

The downside is the strain on crews and staff. In the AGC and NCCER survey, among firms that did data center work, 49% reported more wage pressure and 30% more turnover. Large contractors are pulling crews away from smaller ones, and burnout was on the agenda when MEP contractors and data center owners met this year. Inside some firms, it creates an A team and a B team. The best PMs and superintendents get locked onto data center jobs, and other clients get the next tier down.

Their backlog carries more risk than it appears. Many data center buildings are financed with borrowed money, underwritten on the assumption that tech spending continues. Rising interest rates haven’t slowed that borrowing yet, but The New York Times reports some investors are rethinking speculative projects without firm tenants. Comfort Systems warns in its annual report that its backlog is subject to unexpected cancellations, and one customer made up about 13% of its 2025 revenue. EMCOR only counts the funded phase of a campus as backlog. Bernstein analysts expect cancellations to rise into 2027.

If a big job pauses, payments stop while crew and equipment costs don’t. A contractor that let its steady clients go to chase data center work has nothing to fall back on.

Contractors Outside the Boom Pay More, Wait Longer and Compete for Less Work

Contractors who don’t build data centers still feel the boom. If you’re building a clinic or a school in a market with data center activity, you’re competing with Amazon, Google, Microsoft and Meta for the same gear and the same people.

Gear and materials cost more and take longer. Switchgear averaged about 44 weeks in 2025 (Wood Mackenzie), which is hard to absorb on a 12-month schedule, and construction input prices were up 11.2% year over year in September (ABC), with metals like copper among the fastest risers. Labor costs more too, and JLL says markets with heavy data center development are seeing the biggest construction cost jumps.

The work itself is shrinking. Private nonresidential construction outside data centers was down 7.9% year over year in June (ABC). Subs with data center work are getting more selective about what else they bid. And in September the Fed raised rates for the first time in three years, and the meeting minutes showed officials pointing to AI investment as one source of inflation pressure. Tech companies barely notice, but owners financing offices, apartments and medical buildings do.

So contractors doing ordinary commercial work get squeezed from both sides: less work and more firms bidding on it, while costs rise and owners’ budgets don’t.

Part 4Looking Forward

What to Look For

October brought mixed signals. Meta kept spending, narrowing its 2026 capital spending guidance to $135–145 billion, while CoreWeave’s debt scare showed how quickly lenders can turn. Here’s what to watch from here.

• Big tech’s quarterly earnings: another spending increase keeps 2027 busy; a pullback tightens financing fast.
• Credit markets: private credit and bond buyers fund the buildings, so nervous lenders would slow new starts first. Neocloud bonds like CoreWeave’s are a good early warning.
• Who mobilizes this winter: raw ground in spring means a 2028 job.
• State and local approvals: especially the Texas audit and local moratoriums like Charlotte’s.
• Electrician wages: one of the best early reads on labor across all commercial work.

Recommendations for Data Center Builders

• Know who you’re working for. Confirm the owner entity, its financing, who stands behind the lease and which phases are funded before you mobilize.
• Put power and gear delays in the contract. Spell out receiving, storage, warranty start dates, late delivery of owner-furnished gear, temporary power and delayed commissioning.
• Don’t fix the price on a moving design. Where the design isn’t settled, push for GMP or cost-plus terms instead of a fixed number.
• Watch backlog concentration. Don’t let one client become a large share of your revenue.
• Staff commissioning early and protect your best people. The end of the job is where specialists are scarcest and burnout is highest.

Recommendations for Builders Not in the Boom

• Go after neglected owners. Public construction spending is still growing slightly (about 1% this year through August, per the Census Bureau), and healthcare, education, municipal and tenant improvement owners who feel ignored by their usual contractor are winnable now.
• Chase the work around the campuses. Power infrastructure construction grew from about $9 billion in 2021 to $37 billion in 2025, and substations, utility and site work follow the data centers. Mid-size firms can often get in as subs.
• Protect your schedule and price. Buy long-lead gear in preconstruction when the owner allows, put escalation language in fixed-price bids, approve alternate manufacturers early and check lead times on a set schedule.
• Interview your MEP subs. Confirm crews, experience and shop capacity before you carry their numbers.
• Use steady work to recruit. Data center jobs end when the building does, and the schedules are grueling.