AI data centers are harder to build than they look

On paper, an AI data center is a normal development deal. You secure land, get power to it, put up the building, sign a tenant to lease it and refinance once it’s stabilized. Demand looks strong. Vacancy in the main US markets was 1.4% in the first half of this year, and 80% of what’s under construction had a tenant signed before it was finished.¹

What that version leaves out is time, the chips and who pays. Power can take years to arrive. The town has to approve the project. Only a few companies are big enough to fill these buildings, and they set the terms. And the computer chips inside, which cost far more than the building, lose their value long before the building does.

This post walks through what each size of AI compute costs, why the chips set the clock for everything else, why finding someone to pay for the capacity is harder than it sounds, why the loans don’t match any of it, what goes wrong at each stage of a project, and what changes when you build your own power plant.

What each size costs

The chart puts eleven common sizes of AI computing side by side, from a desktop machine to a gigawatt campus. For each one it shows how much power it draws and what it costs to buy outright, with the chips and without them.

Two things stand out. Per megawatt, the building costs about the same at every size from 5 MW up. JLL puts the building with its power and cooling systems at $11.3M per MW for 2026, and Cushman & Wakefield puts the newest builds at $17.6M.^{2, 3} The second is that the chips are roughly three-quarters of the total at the larger sizes. Most of the money goes into the part that ages fastest.

The chips wear out first

In a normal development the building is the asset. Here the building is closer to a container for the asset. Microsoft just stretched the accounting life of its data center buildings from 15 years to 25. Its servers, the computers that hold the chips, stay at two to six years.⁴

Nobody agrees on how long the chips really last. Amazon cut part of its server fleet from six years to five in 2025 and blamed the pace of AI.⁵ Meta went the other way, to five and a half.⁶ The investor Michael Burry thinks the big cloud companies will understate this cost by $176 billion between 2026 and 2028.⁷ Microsoft’s CEO, Satya Nadella, put it simply: “I didn’t want to go get stuck for four or five years of depreciation on one generation.”⁸

Chips rarely break. They get replaced because newer ones do so much more with the same power. Nvidia’s H100 was the main AI chip of 2023 and 2024. Its newer GB200 system produces roughly 4 to 23 times as much AI output per megawatt, depending on the model it runs.⁹ When power is the scarce thing, every old chip is taking the place of a new one that would earn several times as much.

What the chips earn also swings. The price for renting out an H100 for a year fell to about $1.70 an hour in October 2025¹⁰ and was back around $2.88 by October 2026, because new capacity was sold out.¹¹ That’s good news for owners of older chips today, but it depends on a shortage lasting.

Rough math on 1 MW of computing, using the chart’s numbers. $30–35M of chips written off over five years costs $6–7M a year. The building, at $10–15M over 25 years, costs under $0.6M a year. Electricity at $60 per MWh, cooling included, is about $0.6M. Getting the chip cycle wrong costs about ten times more than getting the power price wrong.

Someone has to pay for it

There are three ways to make money from an AI data center, and each one has a different kind of payer.

• Lease the building. A tenant, usually a big cloud company, pays rent for the space and power and brings its own chips. You’re the landlord.
• Rent out the chips. You buy the chips yourself and charge customers by the hour or by the year to use them. This is what CoreWeave does.
• Sell AI output. You run AI models on your own chips and charge customers for each answer. This is what OpenAI does with its API.

In this post a tenant always means a company leasing the building, and a customer means a company paying for the chips or the output. Either way, there aren’t many of them.

Almost all of the demand traces back to a handful of companies: Microsoft, Amazon, Google, Meta, Oracle and the AI labs they back. CoreWeave rents out chips, and Microsoft, its biggest customer, was 62% of its revenue in 2024 and about two-thirds in 2025.¹² CoreWeave is also a tenant. It leases buildings from Core Scientific, and was 77% of Core Scientific’s revenue in the first half of 2026.¹³ If the company at the top changes its plans, everyone below it feels it.

The big tenants want hundreds of megawatts in one place, power already connected and a builder with a record of finishing on time. Without a signed tenant the money doesn’t come. In December 2025 the first prospective tenant at Fermi’s Texas campus backed out of a $150M deal to help fund construction, and the stock fell about 40%.¹⁴ Even leases that do get signed often need a giant behind them. Hut 8 leased its 245 MW campus for $7 billion to Fluidstack, which rents out chips to Anthropic, and the deal only works because Google guarantees the rent.¹⁵

The other two options have their own problems. If you rent out chips, the price moves every month, as above. If you sell AI output, you’re competing with the AI labs, where the price for the same quality of answer has been falling somewhere between 9 and 900 times a year, depending on the task.¹⁶ What you can charge in year one isn’t what you’ll get in year three.

Building leases are also getting shorter. A lease here means a tenant renting space and power in the finished building, not the land under it. These used to run 10 to 20 years, which is what let a developer borrow against the building for that long.¹⁷ Moody’s says the big cloud companies now want first terms of four to six years, because that’s how long the chips last.¹⁸ Meta’s Hyperion campus in Louisiana, one of the largest of these deals, Meta is the tenant on a first lease of just four years, and it guarantees the building’s value for 16 years in case it leaves.¹⁹ Oracle went the other way and signed building leases of 15 to 19 years as a tenant, so it carries that risk instead of its landlords.²⁰

The loans don’t match

This is the financing side. Lenders here mostly look past the building to whoever is paying you, the tenant or the customers. Houlihan Lokey says the financing “is secured by the tenant’s credit or the equipment inside, not the real estate.”¹⁷ No strong tenant or customer on a long contract means no cheap money.

That puts developers in a squeeze. Construction loans run two to five years and then have to be replaced with long-term debt, which wants exactly the long lease that’s getting harder to find. Meanwhile the profit in building and selling these projects has thinned. The gap between what a finished project yields and what investors will pay to buy it has shrunk from about 3–4 percentage points to 1.75–3 since 2024.¹⁷

Loans against the chips themselves usually cover 60–70% of what the chips are worth, so a 30–40% drop in their value wipes out the cushion.²¹ The market is already nervous about this. In October the cost of insuring CoreWeave’s debt against default went above 10% a year, and its bonds due in 2031 yielded almost 16%.²² Investors aren’t doubting the demand. They’re doubting that the cash arrives before the loan payments do.

Even the most experienced money walks when the terms look wrong. Blue Owl, one of the biggest backers of these projects, declined to fund Oracle’s $10 billion Michigan site in December 2025.²³

What goes wrong at each stage

Land and site

Picking a site now starts with two questions: can you get power, and will the town allow it. Most projects die here. Bernstein thinks only about 35% of the roughly 500 GW announced in North America will actually get built.²⁴

Power

This is the longest wait. Dominion’s quoted wait to connect a new data center in Northern Virginia went from four years in 2022 to seven in 2024.²⁵ Nadella said last year that Microsoft had chips sitting in storage it couldn’t plug in.²⁶

Utilities are also making data centers carry more of the risk. Across 55 utility rules for very large power users, the typical minimum contract went from 5 years to 12, and you pay for at least 80% of the power you reserved whether you use it or not.²⁷ That’s a real commitment made years before you have a tenant. When AEP Ohio brought in rules like this, data center requests in its area fell from about 30 GW to 13 GW.²⁸

The wait lists are full of projects that won’t happen. Wood Mackenzie expects only about 28% of roughly 1,066 GW of data center requests to get a firm commitment from grid operators.²⁹ The other route is to make your own power, which brings its own problems. More on that below.

Permits and neighbors

Data Center Watch counted 75 projects worth about $130 billion blocked or delayed in the first quarter of 2026 alone.³⁰ Its counts mix delays with cancellations, but the big losses are real. The 2,100-acre Digital Gateway in Prince William County, Virginia, was voided in court, and the ruling held on appeal this spring,³¹ and Google pulled its 468-acre rezoning in Indianapolis.³²

Design and equipment

The big electrical gear has to be ordered before you know what you’ll be housing. The largest transformers were taking more than three years to arrive in early 2026.³³ Over the same stretch, the power drawn by one rack of servers went from about 40 kW with H100s to about 130 kW with a GB200. Nvidia’s next design, due in 2027, is around 600 kW,³⁴ and it is already planning for 1 MW racks.³⁵

A building designed today may not fit the chips that are current when it opens. Microsoft paused the second phase of its campus in Mount Pleasant, Wisconsin, in January 2025 to rethink the design around “recent changes in technology.”³⁶

Construction

Sightline expects 30–50% of the large projects due in 2026 to slip.³⁷ The delay doesn’t stay with the builder, either. CoreWeave cut its 2025 revenue forecast because one outside developer was late.³⁸ The builders who hit their dates tend to be the ones who started on sites that already had power, often old crypto-mining sites.

Signing a tenant

About a fifth of what’s under construction still has no tenant.¹ Tenants also change their minds. Microsoft dropped about 200 MW of building leases in early 2025 when its forecasts changed.³⁹

Replacing the chips

This is the stage most models skip. The chips age out in a few years, whether they’re the tenant’s or yours, the lease comes up, and the next generation needs more power and different cooling than the building was designed for. Upgrades are rare and expensive. There’s also no real example yet of an AI data center being leased to a new tenant after the first one failed, so nobody knows what that looks like.

Making your own power

Because the grid is so slow, more developers are building their own power plant on or next to the site. Usually that means gas turbines, sometimes gas engines or fuel cells, often with batteries. Some run on it only until the grid connection arrives. Others plan to run on it for good.

The biggest builders are already doing it. xAI got its Memphis-area sites running on temporary gas turbines and batteries before the grid caught up.⁴⁰ The Stargate campus in Abilene, Texas, has about 1.2 GW of gas generation ordered for the site.⁴¹ Oracle has signed for up to 2.8 GW of fuel cells.⁴² Meta went through its utility instead, and Entergy now plans ten gas plants, about 7.5 GW, for Meta’s Louisiana campus.⁴³

The appeal is speed, and the math backs it up. On the rough numbers above, a megawatt of chips costs $6–7M a year whether it’s running or not, while its power costs well under $1M. Paying more for power to get the chips running a year or two sooner usually pays off.

But now you have a second development project, a power plant, with its own problems.

The equipment is scarce. Gas turbine prices are up about 195% since 2019, and Wood Mackenzie puts the wait for large ones at six years.⁴⁴ GE Vernova, which makes them, says it’s closer to three.⁴⁵ A new gas plant cost about $2,157 per kW of capacity in 2025, up from under $1,500 two years earlier.⁴⁶ At that price, the power plant for a 100 MW data center costs more than $200 million before any building goes up.

The power costs more. Lazard puts electricity from a large, efficient gas plant at about $90 per MWh in 2026, a 17-year high.⁴⁷ The smaller turbines and engines that fit on a data center site usually burn more gas for the same power. You also need a gas pipeline nearby with room to spare.

There’s no backup. On the grid, if one power plant trips, others cover for it. On your own plant, a turbine going down takes chips offline, so you have to buy spare units that mostly sit idle.

The permits are a fight. xAI ran turbines under a rule that lets temporary generators skip an air permit, and the NAACP sued in April 2026 over the turbines powering its Mississippi site.⁴⁸ The state had approved a permit for 41 turbines there a month earlier.⁴⁹

Regulators still have a say. Even connecting a data center straight to an existing power plant can need federal approval. When Amazon tried to take more power directly from Talen’s nuclear plant in Pennsylvania, regulators turned it down, and the deal was rebuilt to run through the grid.⁵⁰

So making your own power doesn’t remove the risk. It swaps the wait for the grid for equipment, fuel, permits and a higher power bill. It works best as a bridge until the grid arrives, or on a site that already has a gas line and a county that wants the project.

Small sites have their own problems

Going small avoids some of this. Most utility rules for very large power users start around 20 to 25 MW,²⁷ so a 5 MW site can often skip the long grid wait and the minimum payments. But the big tenants want hundreds of megawatts in one place, so a small site usually ends up with smaller tenants with weaker credit, or renting out chips by the hour. It also carries much of the same fixed cost on far less revenue. People who invest in these deals are careful. “Anything under one MW per site, and I start to get a bit worried,” one told Latitude Media.⁵¹

That’s why the highlighted rows in the chart sit where they do. They’re big enough to carry the fixed costs and mostly small enough to skip the longest waits for power.

What it adds up to

The demand is real, and in the big markets the buildings are filling up. But cost per megawatt is the easy part of the model. Projects get into trouble on the chip cycle, the wait for power, the approvals and finding a tenant or customer with good credit who will commit for long enough. Each of those needs its own line in the numbers.

The building lasts 25 years. The chips that pay for it last maybe five, the tenant may sign for four, and the power can take seven to show up. Getting those to line up is most of the work.