The first thing I would put next to any data center cost-per-square-foot number is not another number. It is a question: what kind of square foot are we pricing?
That sounds fussy until you look at the current U.S. data. Gordian's RSMeans model puts the Q1 2026 national average for a one-story computer data center at $504.76 per square foot. ConstructConnect's August 2026 report, looking at tracked data center construction starts rather than a standardized building model, puts the year-to-date median at $570 per square foot and the average at $859.
Those figures are far enough apart that simply quoting “the 2026 average” would be misleading. The useful part is understanding why they differ and deciding which benchmark is closest to the project you are actually trying to budget.
Three current benchmarks, three different views of the market. They should not be averaged together.
A useful starting range is roughly $400–$700 per square foot — before project mix pushes it higher
For early planning of a conventional one-story facility, the most defensible public U.S. benchmark I can find is Gordian's current RSMeans dataset. Its Q1 2026 figures range from just under $400 per square foot in the Austin/Dallas model to nearly $690 in Palo Alto. The national model lands a little above $500.
Gordian is very explicit about the limitation: these square-foot models are for planning and budgeting, not detailed estimating. That matters. A standardized model is useful precisely because it holds many assumptions steady; a live project portfolio does the opposite and captures whatever developers are actually starting, including unusually expensive facilities.
| Market | Q1 2026 | Position vs. national model |
|---|---|---|
| Austin / Dallas, TX | $397.80/sf | About 21% below |
| Charlottesville, VA | $423.39/sf | About 16% below |
| Nashville, TN | $431.32/sf | About 15% below |
| Atlanta, GA | $433.34/sf | About 14% below |
| Tempe, AZ | $445.18/sf | About 12% below |
| Denver, CO | $454.34/sf | About 10% below |
| National average | $504.76/sf | Baseline |
| Boston, MA | $592.92/sf | About 17% above |
| Jersey City, NJ | $600.15/sf | About 19% above |
| Chicago, IL | $613.38/sf | About 22% above |
| Palo Alto, CA | $688.26/sf | About 36% above |
The source here is Gordian's July 2026 RSMeans update. I would use that table to localize a conceptual estimate, not to conclude that a real Dallas project will necessarily cost $397.80 for every square foot it builds.
Why another 2026 dataset is already showing $859 per square foot
ConstructConnect is measuring something different. Its August 2026 Data Center Report looks at projects in its construction-starts data, removes projects under 2,000 square feet and trims the highest and lowest 10% by square-foot cost. Even after that trimming, the year-to-date median is $570 and the average is $859.
The average sits roughly 51% above the median. That gap is a useful warning that a smaller number of very expensive projects are pulling the arithmetic mean upward.
I find the median more useful than the average when I want a quick feel for what a “middle” tracked project looks like. The average is still informative, but it answers a different question: how much spending per square foot is showing up across the project mix, including the large high-cost tail.
The change from 2025 is also striking. ConstructConnect reports a 2025 median of $415 per square foot and an average of $514. Its 2026 year-to-date figures are therefore up about 38% and 67%, respectively. I would not automatically interpret all of that as pure inflation. A shift toward larger AI-oriented, high-density or otherwise technically intensive projects can change the mix as well as the underlying prices.
The same 100,000 square feet can produce very different budgets
Square-foot pricing becomes tangible when you put it against a building size. The arithmetic is simple; the interpretation is not.
Lower-cost conceptual market benchmark.
Close to the current RSMeans national model.
ConstructConnect's 2026 YTD median.
ConstructConnect's 2026 YTD average.
The mistake would be to choose one of those four totals without asking what power capacity, redundancy, cooling architecture and cost scope sit behind it. Floor area is only the denominator. In a data center, much of the money is being spent to deliver reliable power and reject heat.
This is why cost per MW often tells me more than cost per square foot
A conventional commercial building is usually discussed in square feet because the space itself is a major part of what is being delivered. Data centers are different. Two buildings with identical floor areas can support very different IT loads.
Consider a simplified example using the same $11 million-per-MW construction benchmark for two 20 MW facilities:
$220M ÷ 200,000 sf = $1,100/sf
$220M ÷ 400,000 sf = $550/sf
Nothing happened to the cost per MW in that example. Only the amount of floor area associated with the same IT capacity changed, yet the calculated cost per square foot was cut in half.
That is why I would pair this article with our data center cost per MW guide rather than treating the two metrics as substitutes. Square feet are useful for real-estate planning and conceptual building budgets; megawatts are usually better at expressing what the technical facility is being built to deliver.
Use $/sf when...
- the building footprint is known before the final IT load;
- you are comparing geographic construction cost models;
- you need a conceptual real-estate or shell budget;
- you are sanity-checking a building-level estimate.
Use $/MW when...
- power capacity is the key project constraint;
- you are comparing facilities with different densities;
- cooling and electrical infrastructure dominate the decision;
- you want a closer link between capital cost and IT capacity.
Density is now making the square-foot metric even harder to read
High-density AI infrastructure exaggerates the problem. A traditional cloud facility and a liquid-cooled AI facility may have similar footprints but very different electrical and mechanical requirements.
Turner & Townsend's current U.S. analysis finds that similarly sized liquid-cooled data centers are typically 7% to 10% more expensive to construct than equivalent air-cooled facilities. Its 2025–2026 Data Centre Construction Cost Index also notes that the market is splitting between more stable conventional cloud builds and less standardized AI projects.
JLL approaches the issue from another angle. Its 2026 outlook forecasts an average shell-and-core construction cost of $11.3 million per MW globally and notes that tenant technology fit-out for AI infrastructure can reach as much as $25 million per MW. The important point for square-foot analysis is that a substantial share of the economic intensity can sit in systems that are only loosely related to the number of square feet being constructed.
What actually moves the number
Market and labor
RSMeans' 2026 city table moves from roughly $398/sf in Austin/Dallas to $688/sf in Palo Alto before we change the basic building model.
Power density
Higher IT load per unit of space can require more expensive distribution and cooling without increasing the footprint proportionally.
Cooling architecture
Air-cooled and liquid-cooled facilities should not be treated as cost-equivalent simply because their buildings have the same area.
Redundancy
Electrical and mechanical resilience adds equipment and capacity that a simple floor-area metric does not describe.
Scope
Shell, fit-out, M&E equipment, land, utility work and active IT can be included or excluded differently across published figures.
Turner & Townsend's methodology is useful here because it spells out the construction categories captured in its own benchmark: shell and core, architectural fit-out and finishes, mechanical and electrical fit-out, general contractor preliminaries, margin, contingency and M&E equipment. The methodology is worth reading before comparing its figures with another source.
A practical 2026 budgeting approach
If I were building an early spreadsheet today and only knew the intended size and location, I would start with the local RSMeans square-foot model rather than a generic $1,000/sf headline. Then I would create a separate adjustment layer for the things the model cannot know about my project.
Start local. Use a city or regional baseline instead of a national number when one is available.
Define the area. Be clear whether you are using gross building area, white space or another measure.
Add the technical brief. Capacity, density, cooling and redundancy can move the estimate more than the shell.
Separate exclusions. Land, off-site power infrastructure and IT equipment deserve their own lines rather than being quietly mixed into $/sf.
Cross-check against $/MW. If the square-foot budget and the capacity budget tell very different stories, investigate before trusting either one.
For a conventional one-story U.S. data center, roughly $400 to $700 per square foot is a reasonable conceptual band in the current RSMeans market models. For the broader project market, the current ConstructConnect data shows why budgets can land much higher: its 2026 median is already $570 and its average is $859, with many tracked projects above $1,000 per square foot even after outliers are trimmed.
I would therefore be suspicious of both extremes: a single $500/sf number presented as universal, and a blanket claim that every new data center now costs $1,000 or $2,000 per square foot. The market is expensive, but the spread is real. A useful estimate explains the project behind the number.