The short answer is you'll find different numbers depending on which tracker you trust: industry research firms counted somewhere between roughly 2,700 and 5,400 US data centers heading into the mid-2020s, and the gap isn't sloppiness — it's definition. One organization counts every leased colocation site and hyperscale campus; another counts only the buildings that meet a strict industrial floorplate standard. Knowing which methodology a figure rests on matters more than the figure itself, because the "true" total swings by nearly half depending on whether a server room in a suburban office park gets counted.
Why the Published US Data Center Counts Disagree By Thousands
The single biggest reason two credible sources can publish US totals that differ by more than a thousand is what they treat as a "data center." CBRE's North America Data Center Trends report, the industry benchmark cited by operators and lenders, counts leased colocation facilities and hyperscale self-build campuses that clear a meaningful capacity threshold. Uptime Institute tracks critical facilities broadly across capacity bands. Cloud scene and market trackers like those compiled by datacenterHawk and Synergy Research slice the market by megawatt capacity rather than by building.
A 100 kW server closet in a law firm's basement is, technically, a data center. So is a 300 MW hyperscale campus in Loudoun County, Virginia. No respected tracker puts those in the same bucket without saying so, but popular articles and vendor blog posts frequently do — quoting a big number from one source and a capacity figure from another as though they describe the same universe.
The capacity threshold is the deciding variable. When a firm counts only facilities above, say, 1 MW of critical IT load, the national total collapses toward the low thousands. When it counts every commercial colocation site plus enterprise-owned rooms above a smaller threshold, the count climbs past 5,000. Neither is wrong; they answer different questions. If you're sizing a market, use the capacity-threshold number. If you're mapping connectivity options, use the inclusive number.
Consider what happens when you try to reconcile them. A 2023-era CBRE count of roughly 2,700 leased and hyperscale facilities and a broader industry census topping 5,000 aren't contradicting each other — they're measuring the leased-and-hyperscale universe versus the all-comers universe. The overlap between the two lists is real, but the smaller one is essentially a subset of the larger one with enterprise-owned infrastructure stripped out. That subset logic is why you should never average the two figures to get a "true" count; you'd be counting some buildings twice and inventing a number that matches no methodology.
One practical test cuts through the confusion fast. Ask any source two questions before you quote its number: what capacity cutoff did it use, and does its count include enterprise-owned (captive) facilities or only commercial multitenant and hyperscale? If the source can't answer both, treat its total as marketing rather than measurement. The best-known trackers publish their methodology precisely because operators demand it — and that transparency is what lets you compare their figures honestly instead of treating them as competing claims about one fixed quantity.
The CBRE and Uptime Institute Baselines Compared
CBRE's semiannual North America Data Center Trends report is the most-quoted baseline on the commercial side, and it counts leased colocation inventory and hyperscale self-build capacity with enough rigor that real estate investors use it for underwriting. Its US totals have sat in the low-to-mid thousands of facilities in recent editions, with the caveat that CBRE's focus is on the inventory that trades in the leasing market. Uptime Institute, by contrast, surveys critical facilities across a wider spread of sizes and publishes findings on outages, staffing, and efficiency rather than a pure real estate census.
| Metric | CBRE-style commercial count | Broad industry census |
|---|---|---|
| Typical US facility total | ~2,600–2,800 | ~5,000–5,400 |
| What's counted | Leased colocation + hyperscale self-build | All commercial + many enterprise-owned |
| Capacity floor | Meaningful MW threshold | Smaller or none |
| Best use | Market sizing, leasing | Connectivity mapping, coverage |
| Main blind spot | Enterprise-owned rooms | Double-counting risk across categories |
The table's job is to show you the same reality through two lenses. The left column tells you how much purpose-built, commercially relevant capacity exists. The right column tells you how many places a network engineer might plausibly find racks. Buyers of colocation capacity should anchor on the left; teams planning redundant connectivity should scan the right.
Uptime Institute's value in this comparison isn't its count — it's its longitudinal survey data on what happens inside those facilities. Its annual outage analysis has repeatedly found that a majority of significant outages trace to power problems, with human error and configuration mistakes close behind. That finding reframes the count question: the number of US data centers tells you about market depth, but the failure-mode data tells you what to worry about once you're inside one.
Where the two methodologies bite hardest is at the edges. A colocation provider that opened a small edge site in a Tier 2 city might appear in the broad census and miss the commercial tracker's capacity floor entirely. A hyperscale campus built for a single tenant's own workloads may be captured differently again. So when a headline says "the US has X data centers," the honest follow-up is usually "under which definition, and as of when?" The count is a moving target because construction keeps adding facilities faster than any tracker can re-survey them.
Where American Data Centers Actually Cluster
Geography is where the count stops being abstract. Northern Virginia's "Data Center Alley" around Loudoun County holds the densest concentration of hyperscale and colocation capacity on earth — several million square feet of powered shell in a few contiguous jurisdictions. Dallas–Fort Worth, Silicon Valley, Chicago, Phoenix, and Atlanta round out the top US markets by capacity, and each has a distinct mix of hyperscale self-build and multitenant leasing. Counting facilities by state produces a lopsided map: a handful of states hold a large share of national capacity even as facilities are technically present in all fifty.
That lopsidedness has a physical signature you can feel driving through Loudoun County — the low hum of cooling yards, the windowless two-story shells, the substations packed shoulder to shoulder with office parks. It's not scenery; it's the reason power interconnection queues in that region stretch for years and why operators have been fanning out to secondary markets like Columbus, Ohio; Richmond, Virginia; and the Salt Lake City corridor.
The clustering also distorts any nationwide average you might calculate. If you divide total US capacity by total facility count, you get a number that describes almost no actual facility, because the distribution is so skewed toward a few enormous campuses and a long tail of smaller colocation sites. A more useful per-market figure is capacity per facility within a given metro, and in Northern Virginia that ratio runs far higher than in, say, a midwestern market dominated by smaller enterprise colocation.
Power is the binding constraint on where the next clusters form, and that's reshaping the count's geography in real time. Utilities in the densest markets have imposed multi-year interconnection timelines, prompting operators to sign power agreements years before a building exists. The practical consequence for anyone reading facility counts: recent totals understate the pipeline badly, because a facility with a signed power contract and a poured foundation may not appear in a commercial tracker until capacity is actually deliverable.
For network planning, the clustering cuts both ways. Density in a single metro like Northern Virginia gives you abundant carrier options and low-latency cross-connects within the region. But it also concentrates risk: a wide-area power event or a fiber cut in Data Center Alley can take out redundancy you thought was physically diverse. That's the edge scenario national counts never capture — two facilities "in different data centers" a few miles apart may share a substation, a fiber conduit, and a single point of failure.
Hyperscale Campuses Versus Colocation: Two Different Counts
Hyperscale facilities — the massive self-built campuses operated by Amazon Web Services, Microsoft Azure, Google Cloud, Meta, and a handful of others — are the fastest-growing slice of the US count and the one most often miscounted. A hyperscale "campus" can span multiple buildings, each with its own utility feed, and different trackers count it as one facility, five facilities, or a capacity figure in megawatts. AWS, Microsoft, and Google each operate dozens of US campuses, and their build-out pace means any published hyperscale count ages within quarters.
Colocation, by contrast, is the multitenant business — Equinix, Digital Realty, CyrusOne, QTS, and regional operators leasing cage and cabinet space to enterprises. Colocation counts are more stable because the facilities are commercial and trackable, but they include everything from massive interconnection hubs in Ashburn to single-building regional sites. The two categories overlap in purpose (both host third-party IT) but almost never in counting method, which is why mixing a hyperscale campus count with a colocation site count produces a meaningless sum.
| Category | Typical operator scale | How trackers count it | Main counting pitfall |
|---|---|---|---|
| Hyperscale self-build | Dozens of US campuses; each campus may be multiple buildings | Often by campus or by MW | Building vs. campus ambiguity |
| Multitenant colocation | 50–250+ US sites per major provider | By leased facility | Small edge sites missed or included |
| Enterprise-owned | Thousands of small rooms | Rarely counted consistently | Thresholds vary wildly |
The pitfall column is the practical takeaway: no matter which category you're reading about, the unit of count (building, campus, or megawatt) changes the total. A hyperscale operator announcing "20 new data centers" may mean 20 buildings across four campuses, which a tracker might log as four facilities. That's a 5x discrepancy baked into a single press release, and it's why vendor announcements should never be converted directly into census additions.
For enterprises choosing between the two models, the counting ambiguity is a distraction from the real comparison. Hyperscale gives you elastic capacity and managed services but locks you into one provider's ecosystem and pricing. Colocation gives you carrier neutrality and physical control but makes you responsible for your own hardware lifecycle. The count of US facilities in each category tells you about market maturity; it tells you nothing about which model fits a given workload.
Reading the Numbers: How to Use a Facility Count Without Being Misled
Use a facility count the way you'd use a census — as a denominator, never as a verdict. If a vendor claims "access to 5,000 US data centers," ask whether that number includes facilities they can actually deliver service into, or whether it's a partner-map total padded with sites that have no interconnection agreement. The gap between a directory count and a serviceable count is often large, and it's the number that affects your latency and redundancy, not the headline total.
Three derived ratios make a count useful. Capacity per facility tells you market density. Facilities per provider tells you concentration risk. Facilities per metro tells you where competition keeps colocation pricing honest. None of these requires a perfect national total — they work fine on a consistent methodology, which is why you should pick one tracker and stay with it across years rather than cherry-picking the biggest number each time you need a citation.
That consistency principle has a cost you should plan for. Re-surveying a national facility census takes time, so published totals lag reality by months, and construction in the fastest-growing markets outpaces any annual update. If you need a current figure for a decision, the defensible path is to combine a published total with the named tracker's stated as-of date and treat the number as a floor, not a ceiling. Writing "roughly 2,700 leased and hyperscale facilities as of the tracker's latest edition" is honest; writing "there are 2,700 data centers in America" is not.
Treat vendor-published counts with particular suspicion when they appear in marketing materials rather than methodology documents. A provider's "global footprint" page will list every announced site, including ones under construction or merely contracted, because the number is a sales asset. A research firm's count will exclude unbuilt capacity because the number is an analytical asset. Same word, opposite incentives — and the incentive tells you which number to trust for which purpose.
What the Count Misses: Edge Sites, Micro Data Centers, and Enterprise Rooms
The published counts systematically miss the long tail. Enterprise server rooms, edge micro data centers tucked into telecom huts, and on-premises closets in hospitals and factories number in the tens of thousands, and almost no tracker counts them because they aren't commercial inventory. That omission matters because a large share of enterprise IT still runs in exactly those spaces — and when they fail, they fail without the redundancy a colocation facility provides.
Edge computing is compounding the omission. As latency-sensitive workloads (content delivery, industrial control, retail point-of-sale) push compute closer to users, operators are deploying small facilities in places no census was designed to capture: basements, rooftops, telecom central offices. A micro data center of a few racks can serve thousands of users, yet it satisfies no reasonable capacity threshold and appears in no commercial facility count.
The failure case here is instructive. An enterprise that believes its workloads are "in a data center" because a partner's directory lists 5,000 facilities nationwide may discover its actual redundancy depends on two small edge sites sharing a landlord's single power feed. National counts can't see that; only a site-level review of power paths, fiber entry points, and generator coverage can. The count tells you the market is deep. The site review tells you whether your specific pair of facilities is diverse.
This is where the "how many" question gives way to a better one: how many facilities can you actually reach with the redundancy and latency your workload requires? That number is smaller than any national total and far more useful. It's built from carrier availability, power topology, and proximity — not from a census.
How the US Count Compares With the Rest of the World
The US leads the world in data center capacity by a wide margin, and most global counts reflect that. Europe's combined total trails the US even when you sum its major markets, and Asia-Pacific's total is concentrated in a few countries — China, Japan, Singapore, India — with China's hyperscale build-out the fastest-growing anywhere. The US advantage isn't just facility count; it's the depth of interconnection ecosystems in markets like Northern Virginia and Dallas that no other region matches at the same scale.
The comparison has a caveat that global headlines usually skip. Different countries define and count data centers differently, so cross-border totals are even less comparable than domestic ones. A European count may exclude enterprise facilities entirely, while an Asia-Pacific census may include telecom central offices. Quoting "the US has X and Europe has Y" without matching methodology is the same error as mixing hyperscale campuses with colocation sites, just at a larger scale.
What the global picture does establish reliably is direction. US capacity has grown steadily as hyperscale demand accelerated, with the fastest growth concentrated in a handful of markets constrained by power. Europe and Asia-Pacific are growing too, but from smaller bases and with their own power constraints — grid congestion in parts of Europe, land and cooling limits in Singapore. The US count's trajectory is upward; the question is how fast power interconnection can keep up.
For anyone benchmarking, the practical rule is to compare like with like. If you're using a commercial-tracker US total, compare it against a commercial-tracker European total, not a broad census. If you can't find a matched methodology, describe the comparison qualitatively and skip the ratio. A mismatched ratio is worse than no ratio because it looks authoritative.
What's Driving the Count Upward Right Now
Two forces are adding US facilities faster than trackers can count them: AI compute demand and the power-and-land chase it triggered. Training and inference workloads need dense, power-hungry capacity, and operators are racing to secure megawatts in markets where utilities can deliver them. That's pushing new construction into secondary markets and reviving sites that sat dormant, which means any facility total published a year ago is already a floor.
The AI build-out also changes what gets counted. A single large AI campus can draw hundreds of megawatts, dwarfing the capacity of dozens of traditional colocation sites combined. When those campuses come online, they shift national capacity totals dramatically while barely moving the facility count — one new building, enormous capacity. That's the clearest illustration of why facility counts and capacity figures must be read together: the count measures breadth, capacity measures depth, and AI is widening the gap between them.
Power interconnection is the binding constraint on how fast the count can rise. Utilities in the densest markets have acknowledged multi-year queues for new large-load service, and operators have responded by signing agreements for capacity years before construction begins. The practical result is a growing inventory of "announced" and "under construction" facilities that count differently depending on the tracker — some include them, some don't, and almost none label them clearly.
The human side of the build-out deserves a mention because it limits growth in ways power doesn't. Skilled trades for electrical and mechanical fit-out, plus operations staff trained for high-density cooling, are in short supply, and staffing a 300 MW campus takes a different skill set than running a 5 MW colocation site. Uptime Institute's staffing research has flagged this gap for years. The count can grow only as fast as the people who keep the facilities running.
FAQ
Does the US really have more than 5,000 data centers?
Under a broad census that includes enterprise-owned rooms and smaller commercial sites, yes — some counts top 5,000. Under the narrower commercial methodology used by CBRE and similar trackers, the leased colocation plus hyperscale total sits closer to the high 2,000s. The two numbers describe different populations, so both can be true at once.
Why do some sources say 2,700 and others say 5,000+?
Because they apply different capacity thresholds and inclusion rules. A capacity floor above roughly 1 MW of critical load strips out thousands of small enterprise and edge rooms. Sources rarely state their threshold in the headline, which is why the figure alone is unreliable — the methodology behind it is the real information.
Is a hyperscale campus one data center or several?
Often several. A single campus can contain multiple buildings, each with its own power feed and cooling plant, and trackers variously count it as one facility, one per building, or a megawatt figure. That ambiguity is why a hyperscale operator's "new data centers" announcement can't be added directly to a census.
Do enterprise server rooms count as data centers?
Technically yes, and that's the definitional fault line. Enterprise-owned rooms number in the tens of thousands and are excluded from most commercial counts because they aren't leased inventory. Their exclusion is the main reason broad and narrow US totals diverge by thousands.
Where are most US data centers located?
Northern Virginia's Loudoun County holds the densest concentration, followed by Dallas–Fort Worth, Silicon Valley, Chicago, Phoenix, and Atlanta. These markets dominate capacity even though facilities exist in every state. Power availability, not geography alone, now determines where new clusters form.
How does the US compare with Europe on data center count?
The US leads on capacity by a wide margin, but cross-border comparisons are unreliable because European and US trackers use different inclusion rules. A commercial-tracker US total should be compared only against a commercial-tracker European total; mixing methodologies produces a ratio that looks authoritative but isn't.
What's the biggest mistake people make with these numbers?
Treating a facility count as a measure of redundancy or serviceability. A national total says nothing about whether two facilities you depend on share a power feed or fiber conduit. Site-level review of power paths and carrier diversity answers the question the count can't.
Will the US count keep rising?
Published totals are a floor, not a ceiling, because construction and power-contracted projects outpace survey updates. AI compute demand is pulling new capacity into secondary markets and reviving dormant sites. The limiting factor is power interconnection timelines and skilled staffing, not land or capital.
The figures in this article are general market context drawn from named industry trackers and survey organizations and will shift as new capacity comes online; for a decision that depends on current totals, confirm the methodology and as-of date on the tracker's own published report.