Housing Impacts of Data Centers

For my energy and environmental economics research, the first main blocker we ran into is availability of data on data centers. There is no public, downloadable dataset of United States data center locations that also says when each one was built or their capacity. There are only paywalled datasets from providers such as Baxtel. What followed and what turned out to be the bulk of my work with the Energy Institute at Haas, in partnership with The Opportunity Lab at Berkeley, was creating a dataset and running analysis on it related to home values.

The comparison in Part II is shaped like a difference-in-differences study: two counties that got the buildout, seven that did not, read on either side of 2021. But that is only the shape, and nothing in Part II is modeled. No regression has been fitted, so there are no coefficients, no confidence intervals, and no significance tests here, but that is the future of the research. Part II instead reads the same comparison descriptively, through permitting acceleration in the counties with data centers and what home values did around the break.

Part I: Building the dataset

There were five main sources of data center data our team identified. For our research to occur, we needed the following per data center facility: a location precise enough to place it in a county, and a date of operation or building construction precise enough to separate a before from an after. Most of the five had the first and almost none had the second. datacenters.com gives up detailed information, but only one facility-click at a time; Baxtel and Datacentermap are directories with no usable build date, and all of them put the rest behind a paywall. DC Byte is broad, covering 8,140 facilities across 138 countries, but carries no permit or construction year at all, which left three of the five unusable outright and a fourth unusable for timing. However, in an article from Business Insider I read, I noticed they had a field that no other source had: first permit issue year. This helped shape the future of my involvement in the project.

What was interesting was seeing how these sources counted facilities differently, and not all of their methodologies were public. Business Insider's map and dataset was built from public-records requests filed with air-quality regulators, so the underlying records are backup-generator air permits. It captures facilities large enough to need the permit, but misses smaller colocation sites. DC Byte is assembled bottom-up from satellite imagery, earnings releases, planning documents, and site visits.

None of the sources offered a download feature, but both DC Byte and Business Insider drew their facilities on an interactive map, and a map has to get its markers and information from somewhere. In both cases, the browser was fetching the complete underlying records to draw the maps.

So, I realized I might be able to capitalize off of this. The method I used to essentially gain their datasets was to open each map/website with browser developer tools running, watch the network panel while the map loaded and while filters were applied, find the request carrying the facility records, and take the payload. This retrieves exactly what the page already serves any visitor, all the locations and information on data centers in the map. DC Byte answers with a single 6.9 MB JSON document holding 8,147 facility objects, which when cleaned turned into 8,140 usable rows across 138 countries. Business Insider makes no separate data request, and the records are embedded in the front-end bundle, so the captured artifact is 9.8 MB of JavaScript that facilities had to be extracted from and reshaped into a 98-column table.

After obtaining the raw datasets, I then had to spend time cleaning them. Most of that work was mechanical: population and capacity figures stored as comma-separated text rather than as numbers, casing inconsistencies, a column header reading LARGE FACILITES, and single operators filed under many spellings. Equinix arrived under nine, from a bare “Equinix” to “EQUINIX LLC [formerly Verizon]”, and leaves the cleaned file as seven.

Part II: What the data shows

The treatment is a change in the pace of data center permitting, not the arrival of a county's first data center. The first design I tried compared Loudoun and Prince William against neighbouring counties, until an audit found Fairfax, sitting in the control group, hosting 45 facilities of its own. The second design I tried measured how far each ZIP code sits from the nearest data center, on the theory that homes closer to one should behave differently from homes further away. They don't, and the reason is that the facilities are not spread out: all 253 of them sit in just 11 ZIP codes, so measuring distance mostly just restates which ZIP code a home is in. The third design I tried gave each ZIP code its own start date, taken from the year its nearest facility was permitted. That fell apart too, because 23 of the 33 ZIP codes in the two counties contain no facility to date them from. What ended up working was to stop treating any individual facility as an event and read the post-2021 wave as an aggregate structural break, a design that needs no facility-level timing precision, only that the acceleration be sharp, concentrated and clearly placed in time.

So the treated group is Loudoun and Prince William, which between them hold 253 of Virginia's 329 permitted facilities, and the treatment is the change in pace rather than any dated arrival. The before period ends in December 2020 and the after period begins January 2021. That cutoff keeps the first COVID-19 shock of 2020 out of the before period while still catching the whole permitting wave, and it sits just after GPT-3 shipped in mid-2020, which is roughly when the AI demand this buildout serves began. The panel itself reaches back to January 2009, but the charts below start at January 2012, the bottom of the housing crash, because indexing to a trough is what makes two groups at very different price levels comparable at all.

From the Zillow Home Value Index dataset I had access to, I used the 74 ZIP codes across nine Virginia counties, monthly from January 2009 to January 2026, with home values in each month. Loudoun (17 ZIP codes) and Prince William (16) are the treatment counties. The other 41 ZIP codes come from seven counties that have never hosted a single data center: Arlington, Rappahannock, Orange, Clarke, Spotsylvania, Stafford, and Warren. They were chosen because they sit in the same Northern Virginia and Washington DC region. Because home prices differ so much between the two groups, each ZIP is divided by its own January 2012 value and multiplied by 100, and the collective group figure is the average of those, so every ZIP counts the same instead of the expensive ones counting for more. January 2012 is the base because it is roughly the bottom of the housing crash. This allows us to “normalize” the data somewhat.

Part III: What this does not establish

The flat gap before 2021 in Figure IV is partly an artifact of averaging. Reading Loudoun and Prince William as one group hides that the two were moving in opposite directions against the comparison counties through the 2010s: Loudoun's own gap drifts down about 0.4 index points a year before 2021, Prince William's drifts up by about the same, and the pair cancel almost exactly, which is why the combined line comes out flat. Parallel pre-trends are the assumption a before-and-after comparison like this rests on, and neither county satisfies it alone. Only the average does.

The comparison group may not have been the best choice either, which led me to question what the data would look like using every county available instead of seven handpicked counties, so the analysis was rerun statewide, across the 119 Virginia counties and 629 ZIP codes that carry a January 2012 value to index against. That version produces a larger gap, but those counties were already diverging from Loudoun and Prince William throughout the 2010s, making it difficult to draw conclusions.

Part of the reason it was difficult to draw conclusions was price level. The median January 2012 home value was $378,500 in the treated ZIP codes against roughly $159,000 across the statewide zero-data-center ZIP codes. The seven matched comparison counties sit near $270,000, much closer to the treated ZIP codes, and that closeness is what makes them the better control: when two groups start from a similar price, a similar dollar move shows up as a similar percentage, so an indexed comparison between them is measuring the housing market rather than the distance between their price tags. For example, a $40,000 gain is a much larger percentage of a cheap home than an expensive one. One thought is rural Virginia may have grown faster in percentage terms through the post-COVID boom without gaining more in dollars.

There is a good amount that the analysis I've done cannot rule out. It shows that counties with data centers diverged from comparable counties within about a year of permitting accelerating, but not that another shock hit Loudoun and Prince William over the same window. There are also only two treated counties, and no mechanism identified. Confounding variables like construction workers who needed housing, tech employment moving in, land competition impacting the housing supply, or something else unrelated to data centers aren't accounted for. The permit year is also not the opening year, so if housing markets respond when a facility opens rather than when it is permitted, the treatment date runs early by however long that gap is.

Overall, throughout this research I learned more about how to leverage tools such as Python, get creative with sourcing data, and learn more about digital infrastructure's impacts on households.