Every developer who calls me about a data center budget starts with the same question: what does it cost per megawatt? And every honest answer starts with the same follow up: included in what?

That is not evasion. In this market, the scope is the number. One source will quote you a shell. Another will quote you a finished building with power and cooling. A third will quote you an all in development budget with land, and a fourth will quietly add the tenant's AI fit out on top. All four can call their figure cost per megawatt, and all four can be right about a different thing.

Here is how I sort it when I am underwriting a project, and what the current figures actually cover.

The three numbers you will hear

First, construction alone. Turner and Townsend puts construction, meaning the building plus mechanical and electrical work and excluding land and IT, at about $9.8 million to $13.3 million per megawatt in major U.S. markets. Phoenix and Columbus sit near the low end at about $9.8 million per megawatt, Atlanta at about $9.9 million, New Jersey at about $12.9 million, and Silicon Valley at the top at about $13.3 million. Geography alone moves the same basic building by roughly a third.

Second, the global shell and core benchmark. JLL puts average construction cost in the largest global markets at $10.7 million per megawatt in 2025, rising to $11.3 million per megawatt in 2026. That sits inside the Turner and Townsend U.S. range, which is reassuring. Two different houses, two different methods, roughly the same neighborhood for the building itself.

Third, all in greenfield development. Cushman and Wakefield puts all in greenfield development costs in the U.S. and Canada at an average of $17.6 million per megawatt for the most modern facilities, excluding chips and GPUs. That is the figure I use for a total project check, because it is the one that tries to capture land, site work, the building, power infrastructure, and the soft costs and contingencies that get a site to operating. It is not a construction quote. Comparing it to a construction quote is the most common budgeting error I see.

What is inside the all in figure, and what is not

Power infrastructure is the single largest category in the Cushman and Wakefield all in figure, at 21% of total greenfield cost. That surprises people who picture a data center as a warehouse with servers. The building shell is not the expensive part. Bringing power to the building, distributing it, backing it up, and cooling what it feeds is where the money goes.

Contingencies rank second. Cushman and Wakefield ties that directly to long material lead times, labor shortages, and a shifting regulatory environment. I read that as an appraiser: the market is pricing uncertainty as a line item, not as a footnote. Skilled mechanical, electrical, and plumbing wages rose an average of 5% per year from 2021 to 2024, with even steeper growth in secondary and tertiary markets, and those trades are the bottleneck on almost every schedule I review.

What the $17.6 million figure does not include matters just as much. It excludes chips and GPUs. And that exclusion is enormous. AI optimized fit outs can add up to $25 million per megawatt on top of shell and core costs. In plain terms, the tenant's technology can cost more than the entire building that houses it. If a budget adds that fit out into a real estate cost per megawatt, it will overstate the real estate. If it leaves the fit out out of a total capital plan, it will understate the project by more than half. Keep the two buckets separate, always.

The AI premium shows up inside construction too. Liquid cooled, AI ready facilities cost 7% to 10% more to build than comparable air cooled facilities, before you buy a single GPU. Density is driving that. The cooling plant, the power distribution, and the floor loading for dense racks are not the conventional build with a new label.

Why the number keeps moving

Cushman and Wakefield found construction costs have risen an average of 21% per megawatt since the fourth quarter of 2024. That is not ordinary construction inflation. The firm attributes it to supply chain constraints, rising skilled labor wages, evolving land fundamentals, and escalating costs for critical components.

For valuation, that pace creates a specific trap. A cost comp from 2024 is not a comp for a 2026 budget without a time adjustment, and a time adjustment built on a general construction index will miss what switchgear, generators, and cooling equipment have actually done. When I use the cost approach on a newer facility, I want the source, the date, and the scope of every cost figure in the workfile, because a two year old number with the wrong scope is worse than no number at all.

There is also a retrofit reality check. Upgrading an older facility runs $50 million to $100 million per 10 megawatt facility, or $5 million to $10 million per megawatt in round terms. That is close to new construction money for a building that will still have an older building's layout, ceiling heights, and power path. It explains why so much capital chooses greenfield, and why older data center stock does not automatically become AI capacity just because demand is strong.

How I use these figures in an appraisal

I start by fixing the scope in writing: construction only, shell and core, all in development excluding IT, or total capital including fit out. Then I pick the benchmark that matches that scope, and only that scope. The land check runs separately, using the method I laid out in my piece on how to value data center land, where powered land and total project cost per megawatt set the context. The underwriting check runs through the same discipline I use in my appraiser's checklist for underwriting a data center, because a cost figure that does not survive the power, tenant, and residual questions is just a number.

Then I sanity check the total. A 50 megawatt project at $17.6 million per megawatt is $880 million before chips and GPUs. At the JLL shell and core figure of $11.3 million per megawatt, the building alone is $565 million. The gap between those two totals, about $315 million on that example, is land, power infrastructure beyond the building, site work, soft costs, and contingency. If a developer's budget shows almost no gap, something is missing. If it shows a much larger gap, I want to know which line item grew and why.

The honest gaps

Two cautions I give every client. First, there is no single audited cost database for data centers the way there is for some other property types. These figures come from brokerage and cost consultancy guides built on their own project samples. They are the best public benchmarks available, they agree with each other more than they disagree, and they are still samples, not censuses. A specific project's actual cost will move with its redundancy level, its density, its market, and its power path.

Second, per megawatt only works if the megawatt means the same thing. Some figures are per megawatt of IT load, others per megawatt of facility power, and a source does not always say which in the headline. Before I compare any two figures, I confirm the denominator. A 20% apparent cost difference can be nothing but a denominator difference.

The bottom line

In 2026, budget about $9.8 million to $13.3 million per megawatt for U.S. construction depending on market, about $11.3 million per megawatt as a global shell and core average, and about $17.6 million per megawatt all in for a modern greenfield project in the U.S. and Canada, excluding chips and GPUs. Price the AI fit out separately, at up to $25 million per megawatt, and never let it blur into the real estate.

The question is never just what does it cost per megawatt. It is what is included, in which market, at what density, and as of what date. Get those four answers in writing, and the cost figure becomes usable. Skip them, and you are comparing headlines, not projects.

Sources: Cushman and Wakefield 2026 Data Center Development Cost Guide (via Business Wire, September 2026); Turner and Townsend Data Centre Construction Cost Index 2025 to 2026; JLL 2026 Global Data Center Outlook; Data Center E-Primer, February 2026.