U.S. Data Center Industry Report 2026: Two Industries Under One Name

Forty-three gigawatts of data center capacity are under construction in the United States, and the great majority of it belongs to a dozen balance sheets. The 1,255 data centers already standing are a different business: the typical one is about 10 megawatts, six in ten were built before 2015, and resale prices run at a third of replacement cost or less. This report covers both, and it is written for the lenders, investors and developers who meet the second one.

Here is a credit memo of a kind we see more often every quarter: a 6 megawatt colocation expansion in a Midwestern metro that does not appear in any brokerage league table. The sponsor's deck opened with the number everyone opens with: the four largest cloud companies plan to spend something like $725 billion on capital expenditure this year. Then it asked for a construction loan against a building whose tenants are a regional health system, two software firms and a county government.

None of that $725 billion is coming to that building. The sponsor knew it, the lender knew it, and the deck used the number anyway, because there is no other outlook to quote. Every data center outlook published in 2026, from the brokerages to the consultancies to the rating agencies, is written about the gigawatt industry. The building in the credit memo belongs to the megawatt industry, which shares a name with the first one and almost nothing else.

This piece is written for the second industry. It uses the first as context, because the first sets the price of power, the cost of a transformer and the mood of every county board in the country. But the numbers that matter to a lender underwriting a 4, 12 or 40 megawatt facility are not the national gigawatt figures. They are the age, size, density and resale value of the buildings that already exist, and those numbers turn out to be quite different from the ones on the front page.

Two of our own datasets anchor what follows. The first is MMCG's national capacity series, which puts existing U.S. data center capacity at roughly 69 gigawatts of total load, 49 gigawatts of critical load and 43 gigawatts under construction, with hyperscale at 64 percent of existing capacity and colocation at 31 percent (1). The second is MMCG's property-level census of U.S. data centers as of June 2026: 1,778 properties after deduplication, of which 1,255 are existing, 161 are under construction and 320 are proposed (1). Power capacity is recorded for roughly a third of the census records, so we use the capacity series for megawatts and the census for what it is uniquely good at: how many buildings there are, how big they are, how old they are and where they stand. Every figure below that is described as a census figure is our own computation on those records. Where the two datasets disagree, we say so.

For the revenue side of the small industry we lean on IBISWorld's September 2025 report on data center colocation services, which sizes the U.S. industry at $17.1 billion of revenue across 2,226 enterprises, with a 7.2 percent profit margin and a forecast growth rate of 3.3 percent a year through 2030 (2). Hold those three numbers next to the $725 billion. That is the whole argument in one line.

What 1,778 buildings say that 69 gigawatts do not

Capacity statistics describe the industry the way a wealth statistic describes a country: accurately, and in a way that tells you nothing about the median resident. So start with buildings rather than watts.

The 1,255 existing U.S. data centers in the census total about 225 million square feet. The median building is 123,000 square feet. Forty-two percent of them are under 100,000 square feet, and together those 522 smaller buildings hold under 11 percent of the floor area. At the other end, 69 buildings above 500,000 square feet, 5.5 percent of the count, hold a quarter of it (1).

Power tells the same story with a harder edge. Among the 479 existing properties in the census with recorded utility capacity, 49 percent are under 10 megawatts. Those buildings hold 5 percent of the reported capacity. Twenty-eight properties above 100 megawatts, under 6 percent of the reporting count, hold 51 percent of it (1). If you are thinking of the U.S. data center market as a distribution of facilities, it is a market of small buildings. If you are thinking of it as a distribution of electricity, it is a market of a few dozen campuses. Both are true. Only one of them is a loan you will ever be asked to make.

The split by operating model is where the two industries separate on paper. Colocation is 736 of the 1,255 existing buildings, 59 percent of the count and 46 percent of the floor area. The median colocation building was completed in 2000, measures about 87,000 square feet and, where capacity is reported, carries 8.2 megawatts. Hyperscale is 334 buildings, 27 percent of the count and 43 percent of the floor area, with a median completion year of 2022, a median size of about 219,000 square feet and a median reported capacity of 49 megawatts. Enterprise facilities, the corporate and government data centers that two decades ago were most of the industry, are 157 buildings with a median vintage of 2004 and a median of 7.5 megawatts (1).

Read those medians as a sentence. The colocation building is 26 years old and 8 megawatts. The hyperscale building is 4 years old and 49 megawatts. They are not the same asset class wearing different logos; they are different generations of a technology, built to different densities, financed by different capital and sold, when they sell, into different markets.

Geography reinforces the point. Existing data centers appear in 46 states, 289 counties and 157 metropolitan markets. Of those 157 markets, 110 have five or fewer buildings. Sixty-three percent of all existing buildings, holding 55 percent of the floor area, sit outside the ten largest markets by size (1). The top five states by count are Texas with 165 buildings, California with 135, Virginia with 90, Ohio with 87 and Oregon with 69, and together they hold only 44 percent of the standing stock (1). The capacity ranking, which puts Virginia alone at over 18 gigawatts of existing hyperscale and colocation load (1), is correct and describes a different object.

We can define the second industry precisely from the census. It is the 522 existing buildings under 100,000 square feet: 398 of them colocation, 75 enterprise, 29 hyperscale. Their median completion year is 1999. Two-thirds were built before 2010. Seventy-seven percent stand outside the ten largest markets (1). That is the borrower. It has a 25 year old building, a regional tenant roster, a utility service sized for a rack density that stopped being state of the art around the time the iPhone launched, and an appraiser who will struggle to find a comparable sale.

Demand is not the question. Whose demand is.

The demand side of the gigawatt industry has been argued about so loudly since 2024 that it is worth stating the record plainly, because it has moved in one direction.

Lawrence Berkeley National Laboratory's December 2024 report put U.S. data center consumption at 176 terawatt-hours in 2023, 4.4 percent of national electricity, and projected 325 to 580 terawatt-hours by 2028 (3). Its 2025 update revised 2024 actual consumption to 192 terawatt-hours, at the low end of the earlier path, and then raised the forward trajectory anyway: a reference case of 649 terawatt-hours in 2030, nearly 12 percent of U.S. electricity, inside a scenario range of 521 to 843 (4). The single reconciliation of forecast to actual that exists in the public record therefore shows consumption tracking the conservative end of the prior estimate while the estimate itself was lifted. A separate LBNL literature review in November 2025 went further and described U.S. data center load projections as highly uncertain and likely overestimated, citing duplicate and speculative interconnection requests (5). The Energy Information Administration, for its part, stopped relying on estimates and began measuring: pilot surveys of data center operators in Texas, Washington and the Northern Virginia region in March 2026, covering 196 companies (6).

Against that, the buyers of capacity have only raised their numbers. Amazon lifted 2026 capital expenditure guidance to about $220 billion on its July 30 call, up from $200 billion, and said it would still lack the capacity to meet demand this year and next (7). Alphabet moved its range to $195 to $205 billion on July 22 and told investors that 2027 would rise significantly again (8). Meta narrowed to $130 to $145 billion, roughly double its 2025 spending (9). Microsoft's calendar 2026 figure sits near $175 billion after an accounting change that extended data center useful lives from 15 to 25 years, and the company brought 88 data centers online in its fiscal year (10). Oracle spent $55.7 billion in the fiscal year ended May 2026, more than two and a half times the prior year, guided to about $70 billion net for fiscal 2027, and carries a remaining performance obligation of $638 billion that it is funding substantially with debt (11). Goldman Sachs puts the four largest at about $725 billion for 2026, 77 percent above 2025 (12).

The one episode that looked like digestion, TD Cowen's February and March 2025 channel checks reporting that Microsoft had cancelled a couple of hundred megawatts of U.S. leases and walked away from roughly two gigawatts of early stage projects (13), resolved within two quarters into record leasing and a chief financial officer telling analysts the company had been short of capacity for many quarters and would remain so (10). The cancellations were real. They were a reallocation of one company's portfolio as OpenAI's training workloads shifted toward Oracle, not a change in aggregate demand. The credible bear case that remains is about accounting and structure: useful-life assumptions, off-balance-sheet lease commitments, vendor financing that circles back to the vendor. It is not about whether the racks get filled.

So the gigawatt industry's demand is not the question. The question for the second industry is whose demand it can capture, because none of the capital above is addressed to a 6 megawatt building.

The demand that reaches that building is quieter and better documented than it looks. CBRE's pricing data is the clearest signal: asking rates for 3 to 10 megawatt requirements rose 12.5 percent year over year in the second half of 2025, faster than any other size band (14), and the first half of 2026 kept the mid-size tiers on top, with 3 to 10 megawatts up 8.3 percent and 500 kilowatts to 3 megawatts up 7.9 percent while the 10 megawatt-plus tier grew 6.7 percent (15). Tenants at that scale are the enterprises that either cannot put a workload in a public cloud or have decided not to: health systems with HIPAA segregation requirements, public agencies, financial firms with latency needs, and a growing cohort running inference rather than training. AlixPartners' survey of more than 400 executives found 98 percent agreeing that inference would be the key driver of future demand (16), and JLL expects inference to overtake training as the dominant workload in 2027 (17).

Then the honest caveat. Nobody counts absorption at the 1 to 10 megawatt band. CBRE's absorption series is primary-market and large-format; JLL's is the same. The most rigorous project-level pipeline dataset in circulation, Sightline Climate's, tracks 190 gigawatts across 777 projects and excludes anything under 50 megawatts by construction (18). The small industry is underwritten on tenant behavior and pricing pressure, not on counted megawatts, and a feasibility study that pretends otherwise is quoting numbers that do not describe its subject.

The pipeline has a 2028 problem

The 43 gigawatts under construction is the largest pipeline the asset class has ever carried, equal to 62 percent of everything built in the industry's history. What the headline does not show is when it lands.

Our capacity series dates roughly 19.6 gigawatts to 2026, 16.1 gigawatts to 2027 and about 5.2 gigawatts to 2028 (1). Our building count says the same thing from the other direction: of the 161 properties under construction in the census, 107 are dated for 2026, 42 for 2027 and 8 for 2028. The proposals are the mirror image. Of 320 proposed properties, 75 are dated 2027, 82 are dated 2028, 29 are dated 2029 and 14 later still (1). Deliveries in 2028 and beyond are therefore not the tail of the current construction wave. They are a separate wave that has not started, and whether it starts is a power question, not a demand question. We return to that below.

The pipeline is also a different shape from the stock. The 488 pipeline properties (under construction, proposed and in final planning) total 203.6 million square feet, 90 percent of the standing floor area, and the median pipeline building is 283,500 square feet, two and a third times the median existing building (1). Hyperscale is 58 percent of pipeline buildings and 61 percent of pipeline floor area; colocation is 22 percent of the buildings and 17 percent of the area, against 46 percent of the floor area it holds today. Among properties actually under construction the tilt is stronger: 105 of 161 are hyperscale, 61 percent are above 250,000 square feet, and only 11 percent are under 100,000. Where under-construction capacity is reported, buildings above 100 megawatts are 22 percent of the count and 62 percent of the megawatts; buildings under 10 megawatts are 10 percent of the count and four tenths of one percent of the megawatts (1).

The capacity series reaches the same conclusion by a different route: colocation is 31 percent of existing capacity and 6.7 of the 43 gigawatts underway, or 16 percent (1). Colocation's share of every new gigawatt is half its share of the old ones. That is not a market share the second industry is losing to competitors. It is a market that is being built at a scale the second industry cannot participate in, on the same maps.

Geography is moving with it. Atlanta holds 6.6 percent of existing floor area in the census and 19.7 percent of the pipeline. CBRE's first-half 2026 report has Atlanta overtaking Northern Virginia as the country's largest construction market, with 2,882 megawatts underway (15). On a megawatt basis Washington is still far ahead, at roughly 11.7 gigawatts under construction against Atlanta's 2.1 (1); the floor-area census counts proposals as well as construction, and Atlanta's pipeline is proposal-heavy, which is the point. Georgia's pipeline is 2.8 times its standing stock by floor area, Pennsylvania's 2.2 times, Virginia's 1.5 times, Arizona's 1.1 times. Texas, the largest state by building count, has a pipeline equal to 0.9 times its stock. At the other end, Oregon's pipeline is 17 percent of its stock, New Jersey's 11 percent, New York's 3 percent, and Massachusetts and Washington State have no pipeline in the census at all (1). Louisiana has three existing buildings and ten in the pipeline; Mississippi three and twelve; Maryland two and five. The megawatt ranking of construction shows the same new entrants, with Maryland, South Carolina, Wisconsin and Louisiana appearing high in under-construction capacity with almost no existing base (1).

Two things should temper the pipeline figure. First, Sightline's estimate that 30 to 50 percent of the 2026 pipeline may not materialize on schedule (18) is consistent with what the queues show. Texas's large-load interconnection queue stood near 233 gigawatts at the start of 2026 and above 438 gigawatts by mid-year, while only about 9 gigawatts had approval to energize and observed peak large-load consumption was under 4 gigawatts (19). Second, the census records 21 abandoned proposals totaling 9.7 million square feet, among them a cluster of eight buildings in Glen Allen, Virginia, a 1.66 million square foot proposal in Douglasville, Georgia, a 1.5 million square foot technology park in Mooresville, North Carolina and a 1.08 million square foot campus in Plain City, Ohio, plus two deferred projects including a 2.15 million square foot campus outside Augusta (1). Proposals die. Buildings under construction with 80 percent preleasing (15) do not.

Power is the clock

Every data center outlook in 2026 says power is the constraint. Fewer say what that means in contract terms for a borrower, so here is the template the market has converged on, because a feasibility study now has to model it.

There is still no federal rule. FERC opened docket RM26-4-000 after the Energy Secretary's October 2025 directive on large-load interconnection, and on June 18, 2026 issued show-cause orders to all six regional grid operators, each with 60 days to justify or reform its tariffs on application and study processes, cost allocation, co-located generation and flexible-load services (20). That is a process, not a standard, and a study should assume region-by-region divergence for at least the next two years. PJM, which had already received a co-location order in December 2025, is furthest along.

PJM's own auction is the clearest price signal in the country and it is a capped one. The 2027/2028 base residual auction cleared at $333.44 per megawatt-day, the ceiling negotiated with the governors of Pennsylvania and Maryland, and it cleared 6,623 megawatts short of the reliability requirement, the first time the region has failed its one-in-ten-year standard. PJM's uncapped simulation would have cleared at about $530 per megawatt-day for the region and $543 in the Dominion zone (21). A pro forma in the mid-Atlantic that uses the capped number is using a political ceiling that expires after the 2027/2028 delivery year.

The utilities have written their side of the bargain into tariffs, and the terms rhyme. Dominion's GS-5 rate class, approved by the Virginia commission on November 25, 2025 and effective January 1, 2027, applies to loads of 25 megawatts and above: a 14 year contract, a ramp of up to four years, a minimum charge on 85 percent of contracted transmission and distribution demand and 60 percent of generation demand, and collateral of $1.5 million per megawatt where the customer fails credit tests (22). AEP Ohio's data center tariff, approved July 9, 2025, runs 12 years with an 85 percent monthly take-or-pay minimum and an exit fee equal to three years of minimum charges (23). Georgia Power's rules for new loads of 100 megawatts and above allow contracts up to 15 years, with early termination requiring only two years of minimum bills, a materially softer exit than Ohio or Virginia (24). In Texas, the curtailment rule under Senate Bill 6 was adopted on March 26, 2026 and is in force: a co-located large load must be able to curtail fully within 30 minutes of an ERCOT emergency instruction, without compensation, and is barred from paid demand-response programs. The financial gate, still a draft as of this writing with a final rule expected by year end, sets security at $50,000 per megawatt for loads of 75 megawatts and above (25).

Notice the thresholds. Twenty-five megawatts in Virginia, 25 in Ohio, 100 in Georgia, 75 in Texas. The second industry sits below every one of them. That is a real advantage for a 6 or 12 megawatt facility, which can still take standard commercial service without a 14 year take-or-pay contract. It is also a warning: the small operator connects to a grid whose capacity price and reserve margin are now set by customers a hundred times its size, and it has no seat at that table.

The equipment clock is the part most feasibility studies still get wrong. GE Vernova's combined gas turbine backlog and slot reservations reached 116 gigawatts in the second quarter of 2026, and management confirmed on its July 22 call that a heavy-duty turbine ordered now would not be delivered until 2031 (26). Siemens Energy reports lead times beyond 40 months and a firm backlog booked out to fiscal 2028 (27). Wood Mackenzie expects turbine prices to reach $600 per kilowatt by the end of 2027, a 195 percent increase since 2019, with global orders of 110 gigawatts against manufacturing capacity of 60 to 70 gigawatts a year (28). Large power transformers run 128 weeks in Wood Mackenzie's survey, U.S. tier-one manufacturers are quoting 48 to 60 months, and a Department of Energy laboratory report in May 2026 found transformer bushings alone running up to 130 weeks, longer than the transformer (29). Against a data center build cycle of 12 to 24 months, grid interconnection of three to seven years, and a turbine that arrives in 2031, the premium for on-site power is no longer a premium. Bloom Energy delivered an operational fuel cell system to an Oracle site in 55 days (30), and diesel generator capacity at U.S. data centers nearly tripled from 20 gigawatts in 2018 to 55 gigawatts in 2024 (31).

Our census carries the physical trace of this shift. Among existing buildings with reported capacity, utility power per square foot of building area has a median of 72 watts for pre-2000 construction, 84 watts for 2000s construction, 87 watts for the 2010s and 163 watts for buildings completed since 2020 (1). The siting data adds the consequence: the share of new data centers built within a mile of a substation fell from 86 percent in 2016 to 60 percent in 2026 to date, and the average distance more than doubled to about 1.1 miles (1). The industry is walking away from the substations because the substations are full.

What a megawatt earns, costs and sells for

The tightness is real. CBRE's first-half 2026 survey of the primary North American markets recorded 1.4 percent vacancy, supply up 33.7 percent year over year to 10,903 megawatts, and 80.4 percent of under-construction capacity preleased, leaving under 1,500 megawatts available, about six months of supply at the current absorption pace. Northern Virginia vacancy was 0.2 percent. Asking rates for 10 megawatt-plus requirements rose 19 percent in the New York tri-state area and 14.5 percent in Atlanta over the half, while Northern Virginia rose 1.5 percent and Silicon Valley and Hillsboro were flat (15). JLL's midyear report has vacancy at 1 percent for the third consecutive year against 66 gigawatts under construction, 95 percent pre-committed, and puts 77 percent of the capacity under development in frontier markets (17).

Rents by business model, as compiled in CREFC's January 2026 primer from KBRA and CBRE data on 2025-vintage leases, run roughly $100 to $150 per kilowatt-month for single-tenant hyperscale deployments of 40 megawatts and up on 10 to 15 year terms, $150 to $250 for wholesale colocation of 5 to 40 megawatts on 5 to 10 year terms, and $200 to $400 for retail colocation of 1 to 10 megawatts on one to three year terms (32). Chicago quoted $200 to $230 per kilowatt-month for 250 to 500 kilowatt requirements in the first quarter of 2026, Northern Virginia $190 to $235 (15). The lease structure itself is shifting toward triple-net terms, with take-or-pay power floors of 60 to 85 percent of allocated capacity becoming common and ready-for-service fit-out dates stretching from weeks to months, and past six months where an air-cooled hall is being converted to liquid (15).

Cost has decoupled from all of that. Cushman & Wakefield's 2026 development cost guide puts greenfield all-in cost at $8.9 million to $23.3 million per megawatt across the United States and Canada, excluding the chips, with an average of $17.6 million for the most modern facilities, up 21 percent per megawatt since the fourth quarter of 2024 against roughly 5 percent general inflation. Power infrastructure is the largest line at 21 percent of the stack; switchgear prices are up 60 percent since late 2021, copper 46 percent, generators 33 percent, cooling equipment 32 percent, transformers 29 percent. The fully built cost per square foot is approaching $1,000 (33). JLL puts shell and core alone at about $11.3 million per megawatt for 2026 (17). A warehouse conversion, where an operator buys an existing industrial box with a large service and retrofits it, runs about $7 million to $12 million per megawatt, 10 to 15 percent below greenfield and six to twelve months faster, according to the operator and brokerage commentary gathered for this piece; treat the range as directional.

Now reconcile the small industry's revenue statistic against the stock, because the two do not match and the mismatch is instructive. IBISWorld's $17.1 billion of colocation services revenue (2), divided by the 15 gigawatts of existing colocation critical load in our capacity series (1), gives about $1,140 per critical kilowatt per year, or roughly $95 per kilowatt-month. That sits below the bottom of the hyperscale rent band and at less than half the retail band. Either the industry statistic counts a narrower set of firms than the physical stock (IBISWorld's definition is the NAICS 518210 colocation provider, which excludes hyperscaler-owned capacity and much wholesale REIT revenue), or a large share of colocation-classified megawatts is leased at wholesale prices under long contracts, or both. Our read is both, and the practical lesson is that industry-level revenue and margin figures are the wrong instrument for a single asset.

The right instrument is a per-megawatt operating statement. IBISWorld's 7.2 percent profit margin (2) and Digital Realty's stabilized net operating income margin in the low to mid 50 percent range for 2025 (34) describe the same economics at different altitudes: the gap is almost entirely depreciation, about 31 percent of Digital Realty's revenue, plus corporate overhead near 9 percent (34). A lender sizing debt to an asset should be looking at property-level NOI per megawatt after power pass-through, staffing and maintenance, not at an industry EBIT margin that has been through the accountant's depreciation schedule twice.

The capital that funds the first industry is scaling faster than the criteria that rate it. Data center asset-backed and commercial mortgage-backed issuance reached $27 billion in 2025 according to KBRA, up from an average of about $4 billion a year in 2018 to 2022 and around $10 billion in each of 2023 and 2024 (35). J.P. Morgan projects $30 billion to $40 billion a year in both 2026 and 2027 (36), and the Structured Finance Association, citing Barclays, sized the data center securitization market at $61 billion as of July 2026, up from about $4 billion in 2020 and now about 12 percent of the esoteric ABS market, with a path to $180 billion by the end of 2028 (37). Construction debt on the named 2026 financings we reviewed priced around 250 to 400 basis points over SOFR, at 65 to 70 percent of cost for hyperscale-leased projects and 50 to 60 percent with recourse for speculative colocation. Going-in capitalization rates in the 2026 brokerage research cluster near 5 to 5.5 percent for hyperscale assets leased to investment-grade tenants, 6 to 7 percent for wholesale colocation, and 7 to 8.5 percent and wider for powered shell (15)(17).

None of that capital stack reaches a 6 megawatt building in a market with four other data centers. The securitization market's minimum is a portfolio; the private credit market's minimum is a campus. The second industry is financed by regional banks, SBA lenders and the sponsor's own equity, and it is priced off a comparable set that, as the next section shows, barely exists.

The collateral gap

Here is the number a credit committee should carry out of this piece. A modern data center costs about $17.6 million per megawatt to build (33). The average data center that traded over the last three years sold for $6.3 million per megawatt (1). The median one in our census sold for about $3.2 million.

The census sales record is thin but it is ours, and it is consistent with the national distribution. We found just over 100 data center trades with a disclosed price between January 2023 and June 2026. The median price per square foot was $239 in 2023, $129 in 2024, $214 in 2025 and $261 in the first months of 2026. Colocation buildings, the bulk of the sample, traded at a median of $240 per square foot with an interquartile range of $126 to $392. Fifty-five percent of the buildings sold were built before 2000 (1). For the roughly 50 trades where capacity is also recorded, the median price was $3.2 million per megawatt of reported utility capacity, the 20th percentile was $0.8 million and the 80th percentile $6.9 million; a handful of trades where the price bought far more than the reported capacity, enterprise and interconnection assets in Minneapolis, Dallas, Kansas City, Silicon Valley and metro Atlanta, reached $17 million to $23 million per reported megawatt (1). The three-year national distribution has the same shape: a top quintile averaging $19.6 million per megawatt, a bottom quintile under $1 million, and an average of $6.3 million (1). The modern end of the market prices near cost: Blue Owl's $860 million purchase of a 72 megawatt Gainesville, Virginia facility in the first half of 2026 works out to about $11.9 million per megawatt (15). It is the old end that does not.

So the median existing asset trades at under a fifth of modern replacement cost, and the median 2023 to 2026 buyer bought a building from the last century. There are four reasons, and a lender needs all four.

The first is vintage. Median power usage effectiveness in our census runs 1.50 for buildings completed in the 1980s and 1990s, 1.35 for the 2000s, 1.30 for the 2010s and 1.20 for buildings completed since 2020 (1). The national colocation series shows the same curve, from about 1.58 for pre-2000 stock to 1.3 for post-2020 vintages (1). A building at 1.50 spends a quarter more electricity per unit of computing than a building at 1.20, in a market where electricity is the tenant's largest cost after the hardware.

The second is density, and it is the one that has moved. AFCOM's 2026 State of the Data Center survey puts average rack density at 27 kilowatts, up from 16 a year earlier and 7 in 2021, a near quadrupling in three years that its authors describe as a step change (38). Uptime Institute's broader survey of the installed base is more conservative, with a modal density of 11 kilowatts and an average near 8 once the AI outliers are excluded, which is the more relevant figure for a lender because it describes the buildings that exist rather than the ones being planned (39). Uptime has also been widely cited for the finding that about 68 percent of enterprise data centers built before 2015 lack the power density and cooling to host modern AI workloads; we have not been able to trace that figure past secondary reporting and treat it as a single-source estimate (39). Current AI training racks run 100 to 200 kilowatts. A building designed at 5 to 10 kilowatts per rack, which describes most of the 610 pre-2010 buildings in our census, 463 of them colocation (1), cannot be re-tenanted at 50 or 100 kilowatts per rack by installing liquid cooling. STL Partners estimates a liquid cooling retrofit at about $2 million per megawatt against $11 million and up for a greenfield liquid-cooled build (40), but the retrofit only helps where the utility service, switchgear and distribution have headroom, and a building that draws 72 watts per square foot was not built with it. The constraint is the substation, not the chiller.

The third is that the comparable set for distress does not exist. The two material data center bankruptcies of the period both resolved as going-concern sales, not vacant buildings. Brookfield agreed in November 2023 to buy substantially all of Cyxtera's assets for $775 million, about 245 megawatts across some 60 mostly leased sites, roughly $3.2 million per megawatt on the platform figure, then bought the underlying real estate at seven sites for a combined outlay near $1.3 billion; Cyxtera's failure was interest expense that more than doubled into a 2024 maturity wall, not empty halls (41). Sungard Availability Services sold eight data centers and about 400 customers to 365 Data Centers for $52.5 million in 2022 (42). No data center asset-backed note has defaulted or been downgraded, which is a fact about the age of the asset class (the first deal priced in 2018) more than about its safety (35). When primary-market vacancy is 1.4 percent, a failing operator's space is re-leased before the building goes dark. That is good for today's collateral and bad for tomorrow's appraisal, because it leaves the appraiser with theory where comps should be.

The fourth is the appraisal itself. A data center is limited-market, special-purpose property in the sense the Interagency Appraisal and Evaluation Guidelines have used since 2010, and those guidelines require that market value exclude value in use and going-concern value, which for a data center is most of what the lender is really financing (43). Where the comp set is thin, appraisal practice defaults to the cost approach, replacement cost less depreciation, and the cost approach is the one method structurally unable to capture the functional obsolescence described above. A cost indication that sits far above the income and sales indications is not evidence of equity; it is a measurement of the obsolescence the appraiser could not quantify. The Small Business Administration's own standard, which defines special-purpose property as a limited-market property whose design or layout restricts it to the use for which it was built and requires a certified general appraiser with going-concern experience to allocate value among land, building, equipment and intangibles (44), is a useful discipline on conventional deals too.

Our census records the exits that have actually happened. Six demolitions, all of buildings completed between 1975 and 1989, four of them in Silicon Valley. Two conversions. No modern facility gone dark (1). The historical record of decommissioning, compiled by Columbia Law's Sabin Center this August, is a record of enterprise facilities from the 1990s sitting vacant for years before public brownfield money paid for abatement (45). What a lender can salvage from an obsolete data center is the land and the power: powered shells trade in a range that trade coverage puts around $105 to $275 per square foot against $625 to $1,135 for a fully fitted new build, and a reported pair of Loudoun County shells built for about $259 per square foot is the cleanest comp we have seen. The fit-out is roughly four fifths of the cost and nearly all of the obsolescence.

Two more things now sit inside the collateral question that did not five years ago. The first is tenant credit. A growing share of new leases underwrite to AI-native operators whose standalone credit is speculative grade: CoreWeave carries Ba3, BB- and B+ ratings from the three agencies, and its two largest customers were 77 percent of 2024 revenue (46). Fitch's September 2025 criteria for data center securitizations view facilities used for AI training less favorably than those used for cloud, citing obsolescence, lease breaks and limited alternative use, and S&P's methodology assigns its most conservative revenue-volatility assumptions to assets with weaker tenants in secondary markets (47). No data center securitization has been rated above single-A (35). The second is the residual. When Meta financed its Hyperion campus in October 2025 through a $27.3 billion private credit structure, it provided the joint venture a residual value guarantee for the first 16 years of operations, capped, payable if a lease is not renewed (48). The most creditworthy tenant in the industry had to guarantee the building's residual value to get the deal done. That tells you what the market thinks a data center is worth without its tenant.

2026 was the year policy turned

For a decade the public sector's position on data centers was an incentive. In 2026 it became a study committee, and in several places a tax.

Ohio's governor directed the state tax credit authority on May 27, 2026 to pause new applications for the data center sales and use tax exemption while a joint legislative committee reviews the program. The trigger was arithmetic: the exemption cost more than $1.5 billion in 2025 against a forecast of $136 million (49). A House effort to override the governor's 2025 veto of a repeal stalled short of the 60 votes it needed, and companion repeal bills sat in committee through the summer. Arizona's standalone repeal bill, HB 2631, died in committee in January; the budget bill signed June 13, 2026 did the work instead, barring the Commerce Authority from accepting new data center exemption applications from July 1, 2026 through June 30, 2029 (50). Georgia's repeal bill, SB 410, passed the Senate 32 to 21 in March and died in the House; the exemption stands to 2032 while the state projects about $2.5 billion of foregone revenue in fiscal 2026 (51). Virginia kept its exemption, extended to 2050, and added something new: a data center electricity consumption tax of $0.011 per kilowatt-hour effective July 1, 2026, sunsetting in mid-2028, with collections above $600 million refunded pro rata, alongside a new permit process for projects of 100 megawatts and above that mandates sound assessments (52). The National Conference of State Legislatures counts 38 states with data center incentives, 28 in which lawmakers introduced bills to curb them in 2026, and nine that considered repeal (53).

Local government moved faster. The most rigorous tracker of local moratoria, maintained by the ALEA Institute, counted 533 instruments across 42 states as of August 19, 2026, up from 222 in April; its data center-specific series runs 7 moratoria in 2023, 6 in 2024, 59 in 2025 and 294 in the first seven months of 2026 (54). Good Jobs First's narrower tally of data center actions counted 63, of which 54 passed (55). Monterey Park, California became the first city to ban data centers permanently by ballot; Charlotte adopted a 150 day pause; New York's legislature passed a one-year statewide moratorium on projects of 20 megawatts and above that awaits the governor's signature (54).

Water and noise are where the ordinances bite the site plan. California's AB 93, which would have required data centers to disclose projected water use at licensing, was vetoed on October 11, 2025; the companion SB 57, directing the utilities commission to study data center cost shifting to other ratepayers by January 2027, was signed the same day (56). Loudoun County eliminated by-right data center development on March 18, 2025, requiring a special exception with public hearings, and is drafting use-specific standards on generators, noise, setbacks and on-site power for adoption late this year or early next (57). Prince William County removed the commercial HVAC exemption from its nighttime noise limits on October 28, 2025, with a one-year sunset (58). Virginia's Department of Environmental Quality concluded on July 28, 2026 that a data center with evaporative cooling was unlikely to find sufficient groundwater in the eastern coastal plain under current conditions (59).

Read the thresholds again. Most moratorium and permit triggers cluster at 10, 20 or 100 megawatts. The 6 megawatt facility usually sits below the tripwire, which is real relief on entitlement timing. But the incentive it counted on in Ohio or Arizona may now be closed to new applicants, and its feasibility study has to say which side of the line the project sits on in each jurisdiction, and what the exemption was worth in the pro forma if it is gone. Our Ohio and Arizona state pages carry the working assumptions we currently apply: MMCG does not assume the Ohio exemption for new projects while the pause stands , and every Arizona study must address the exemption program's status and the interconnection queue explicitly.

The borrower a bank actually sees

Strip away the gigawatts and the borrower in the credit memo is an operating business, which is the single most important thing about it for a lender, because it determines which programs will touch it.

A colocation provider is classified under NAICS 518210, computing infrastructure providers, data processing, web hosting and related services, with a small business size standard of $40 million in average annual receipts (60). The Census definition places the rental of server and networking space in data centers in that code and not under the real estate lessor codes, and the distinction carries the whole SBA analysis. SBA cannot finance passive real estate. It can finance an operating company that sells metered power, cooling, interconnection, physical security and remote hands to customers who bring their own hardware, and that occupies the building it is financing. Under SOP 50 10 8, in effect since June 1, 2025, business models that lease space are ineligible unless revenue is for services rather than rent, customers have no assigned space of their own and the business supplies the necessary equipment; a retail colocation model on metered contracts passes that test far more comfortably than a wholesale lease of a dedicated hall to a single tenant, which looks like a landlord to the reviewer because it is one (44). The occupancy rules then apply as written: 51 percent owner occupancy of an existing building, 60 percent of new construction with up to 20 percent leased permanently and 20 percent temporarily, and an eligible passive company holding the real estate and leasing it in full to the operating company where the sponsor wants that structure (44). Loan numbers issued from October 1, 2026 fall under SOP 50 10 8.1 (61), and the passive-business and occupancy language should be re-checked against that text before any application is dated.

SBA money will not build the campus. A 7(a) loan tops out at $5 million and a 504 debenture at $5 million, or $5.5 million where a public policy goal such as energy efficiency applies, which at 2026 costs buys well under a megawatt of modern capacity (about a third of one at the modern average, a bit over half at the low end of the greenfield range), a shell around a leased power block, or one phase of a facility that a bank is financing in whole. That is a real use. It is not the $45 million a 5 megawatt warehouse conversion costs, and a sponsor who arrives with an SBA structure for a project ten times that size has misunderstood the program.

USDA is the program that sponsors in rural counties ask about and the one that has never once been used. The Business and Industry guarantee caps at $25 million, with FY2026 guarantee percentages of 85 percent under $5 million and 80 percent up to the cap under the OneRD rule (62). A $25 million guarantee covers the electrical room of one building at $9 to $15 million per megawatt. Our own review this June found no data center financed through B&I, REAP or the rural electric programs anywhere on the public record, and the reason is structural rather than a matter of eligibility: in Virginia each permanent data center job has represented roughly $54 million of investment, about 168 times the average job, and a program scored on rural employment cannot underwrite an asset that employs almost nobody. We wrote that up in The Fifty-Four-Million-Dollar Job, and nothing in the summer's data has changed the conclusion. Roughly two thirds of planned capacity is heading to rural and frontier markets, financed by hyperscaler equity, private credit and local tax abatement, not by Rural Development.

What the second industry has that the first does not is a rent roll. A 6 megawatt building with a health system, two software firms and a county on multi-year contracts is a multi-tenant credit, and multi-tenant credit is what regional bank underwriting is built to evaluate. The first industry's credit is a single lease, and increasingly a single lease to a company rated below investment grade whose revenue depends on one or two customers (46). Tenant diversity at 1 to 10 megawatts is not a consolation prize for the small operator; it is the collateral, and a feasibility study should present it as such, with the concentration, remaining term and credit of each material tenant laid out the way a rating agency would lay out a securitized pool.

So a lender-grade data center feasibility study in 2026 has five jobs, and only the first of them existed in 2021. It counts the market in the borrower's class of building and megawatt, not the national gigawatt, and it names the buildings: the census we ran for this piece is the kind of count we mean, and it is the difference between a 6 megawatt facility in a market with four competitors and the same facility in a market with forty. It treats power as a schedule and a contract, with the utility's large-load tariff, the interconnection study status, the equipment lead times and the capacity price all dated in the pro forma rather than assumed. It prices the collateral at powered-shell value, with a density and service audit that says what the building can host at 30 kilowatts per rack across its floor plate and what it would cost to get there. It stresses tenant credit, not just the asset. And it maps the policy tripwires in the specific jurisdiction: the moratorium threshold, the status of the state exemption and its value in the pro forma, the noise and water rules that now have dollar figures attached. The cost build-up discipline we described in the construction loan feasibility study applies with one amendment: the contingency on the electrical package should be carried separately, because that is where the 60 percent switchgear inflation lives. The scope we run for these engagements is described on our data center feasibility study page.

What we expect

We write outlooks so that they can be checked, so here are ours, dated September 2026, with what would change them.

Deliveries in 2026 and 2027 will land close to the dated figures, because the buildings are up, preleased at 80 percent and largely equipped. The first year in which national deliveries could fall by half is 2028, and whether it does depends on how many of the 82 proposals dated for that year clear an interconnection process that is being redesigned in six regions at once. Our base case is that 2028 deliveries come in well below 2027 and that the shortfall is concentrated in PJM, where the reliability gap is measured, and in ERCOT, where the batch study is not due to return capacity offers until April 2027 (25). A FERC final rule or a PJM auction for 2028/2029 that clears materially above the cap would make us more constructive on 2028 volumes and less constructive on power cost; a data center securitization downgrade or a public capex cut at one of the four largest buyers would make us less constructive on both.

Capacity prices in PJM stay at the cap through the 2027/2028 delivery year and the uncapped simulation is the right planning number after it. Rent growth continues fastest in the 500 kilowatt to 10 megawatt tiers, because that is where available inventory still exists and where the competition for contiguous space is most intense; the retail band holds above $200 per kilowatt-month in the primary markets and drifts up in the secondary ones.

Transaction prices for pre-2015 air-cooled stock will keep falling relative to replacement cost. Our census has the median trade at 18 percent of modern build cost already, and the gap widens as the buyers of second-generation space finish pricing the retrofit against the service capacity. We expect the first properly distressed comps, buildings sold vacant and below cost rather than as going concerns, to appear in 2027 and 2028, and we expect them to be small, old and outside the top ten markets, which is to say in the second industry. That is not a prediction of losses. It is a prediction that appraisers will finally have something to look at.

Incentive rollbacks continue. The count of local moratoria has roughly doubled every four months in 2026 and there is no mechanism in view that stops it. Projects under the 10 to 20 megawatt thresholds will keep clearing while campuses stall, which is, for once, an advantage the second industry did not have to pay for.

And the number in the credit memo stays wrong. There will be a $725 billion figure again next year, larger, and it will describe the first industry, and the building it is being used to finance will still be 25 years old and 8 megawatts. The lender's job is to underwrite the building.

Frequently asked questions

Is the U.S. data center market in a bubble?

The demand record does not support that reading in 2026: measured consumption has tracked the conservative end of federal estimates while the four largest buyers have raised spending three years running, primary-market vacancy is 1.4 percent and 80 percent of construction is preleased. The risk that exists is financial and structural: speculative-grade tenants, residual values that the most creditworthy tenant in the industry had to guarantee, an appraisal method that cannot see obsolescence, and a 2028 delivery cliff that depends on power the grid has not yet agreed to supply. Those are underwriting problems, not a bubble, and they fall hardest on assets that are old, small and single-tenant.

How much does it cost to build a data center in 2026?

Cushman & Wakefield's 2026 guide puts greenfield cost at $8.9 million to $23.3 million per megawatt excluding IT hardware, averaging $17.6 million for the most modern facilities, up 21 percent per megawatt since late 2024; shell and core alone is about $11.3 million per megawatt in JLL's estimate. Power infrastructure is the largest line at about a fifth of the budget. A warehouse conversion with an adequate existing service runs roughly $7 million to $12 million per megawatt.

What is an existing data center worth per megawatt?

Far less than it costs to build one. The national three-year average is $6.3 million per megawatt with a top quintile near $19.6 million and a bottom quintile under $1 million; the median disclosed-price trade in our census of 2023 to 2026 sales was about $3.2 million per megawatt of reported capacity and $240 per square foot for colocation buildings. Value concentrates in secured power and interconnection, then the shell; an air-cooled fit-out designed for 5 to 10 kilowatt racks carries most of the obsolescence.

Can an SBA 7(a) or 504 loan finance a data center?

Yes, within limits. A colocation operator under NAICS 518210 is an operating business, not a passive landlord, provided revenue is for metered services rather than rent and the operator supplies the infrastructure; it must then meet the 51 percent (existing) or 60 percent (new construction) occupancy tests and treat the property as special-purpose in the appraisal. The program maximums of $5 million (7(a)) and $5 million to $5.5 million (504 debenture) size it to a fraction of a megawatt of modern capacity or one phase of a facility, not a campus.

Will USDA Business and Industry finance a rural data center?

There is no closed example on the public record, and we do not expect one. The $25 million guarantee cap is a fraction of one building's electrical package, and a program scored on rural jobs cannot rate an asset that creates almost none. Rural data centers are being financed by hyperscaler equity, private credit and state and local abatement.

What should a data center feasibility study include in 2026?

A market count in the borrower's class of building and megawatt rather than the national figure; a dated power schedule with the applicable large-load tariff, interconnection status, equipment lead times and capacity price; a collateral analysis at powered-shell value with a density and service audit; a tenant credit stress; and a jurisdiction-specific map of moratorium thresholds, exemption status and noise and water rules. Each is described in the sections above, and the scope we run is on our data center feasibility study page.

September 10, 2026 by Michal Mohelsky, principal of MMCG Invest, LLC, a national SBA and USDA feasibility study consultancy

Request a Feasibility Study https://calendar.app.google/EJzWEz3GCqLY2jU86

Michal Mohelsky, J.D. | Principal | mmcginvest.com 

Contact: michal@mmcginvest.com

Phone: (628) 225-1125

Sources

(1) MMCG Invest, U.S. Data Center Census (June 2026) and national capacity series (Q3 2026), built on licensed CoStar Group property, sale and market data; census figures are MMCG computations on 1,778 deduplicated U.S. property records, capacity recorded for 630 (2) IBISWorld, Data Center Colocation Services in the US, Industry Report OD6203, September 2025 (3) Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, LBNL-2001637, December 2024 (4) Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update (5) Lawrence Berkeley National Laboratory, Large Load Literature Review, November 2025 update (6) U.S. Energy Information Administration, Short-Term Energy Outlook, September 2026, and EIA data center energy survey pilot announcement, March 2026 (7) Amazon.com, Inc., second quarter 2026 results and earnings call, July 30, 2026 (8) Alphabet Inc., second quarter 2026 results and earnings call, July 22, 2026 (9) Meta Platforms, Inc., second quarter 2026 results and earnings call, July 29, 2026 (10) Microsoft Corporation, fiscal 2026 fourth quarter results and earnings call, July 2026, and fiscal 2026 first quarter earnings call, October 29, 2025 (11) Oracle Corporation, fiscal 2026 fourth quarter and full year results, June 2026 (12) Goldman Sachs Global Investment Research, hyperscaler capital expenditure estimates, 2026 (Amanda Lynam), as reported (13) TD Cowen, data center channel checks, February 21 to 24 and March 26, 2025 (Michael Elias, Cooper Belanger, Gregory Williams) (14) CBRE, North America Data Center Trends H2 2025, February 2026 (15) CBRE, North America Data Center Trends H1 2026, August 27, 2026 (Gordon Dolven), and CBRE Global Data Center Trends 2026, June 17, 2026 (16) AlixPartners, 2026 Global Data Center Market Outlook, survey of more than 400 executives (17) JLL, North America Data Center Report, Midyear 2026, and JLL 2026 Global Data Center Outlook (18) Sightline Climate, Data Center Outlook, first quarter 2026 (19) Electric Reliability Council of Texas, large load interconnection queue reporting and 2026 Long-Term Load Forecast, and Public Utility Commission of Texas Project No. 58481 (20) Federal Energy Regulatory Commission, Docket No. RM26-4-000 and the June 18, 2026 show-cause orders in Docket Nos. EL26-67 through EL26-72 (21) PJM Interconnection, 2027/2028 Base Residual Auction results, December 17, 2025 (22) Virginia State Corporation Commission, final order in Case No. PUR-2025-00058 (Dominion Energy Virginia biennial review and GS-5 rate class), November 25, 2025 (23) Public Utilities Commission of Ohio, opinion and order approving the AEP Ohio Data Center Tariff (Schedule DCT), July 9, 2025 (24) Georgia Public Service Commission, large-load rules for new customers of 100 megawatts and above, January 23, 2025 (25) Public Utility Commission of Texas, 16 TAC section 25.205 (adopted March 26, 2026) and proposed 16 TAC section 25.194 (Project No. 58481), and ERCOT Batch Zero study procedures (26) GE Vernova, second quarter 2026 results and earnings call, July 22, 2026 (27) Siemens Energy, third quarter fiscal 2026 results, August 2026 (28) Wood Mackenzie, gas turbine market commentary, April 1, 2026, and power transformer lead time and price survey, 2025 to 2026 (29) U.S. Department of Energy national laboratory report TP-5700-96742 on transformer supply chain and bushing lead times, May 2026 (30) Bloom Energy, press release on fuel cell deployment for Oracle, April 14, 2026 (31) Better Data Center Project, U.S. data center diesel generator capacity analysis, March 2026 (32) CRE Finance Council, Data Center E-Primer, January 2026 (33) Cushman & Wakefield, 2026 Data Center Development Cost Guide (John McWilliams) (34) Digital Realty Trust, Inc., Annual Report on Form 10-K for fiscal year 2025 (35) KBRA, data center ABS and CMBS issuance research, March 10, 2026, and Data Center ABS Global Rating Methodology, January 9, 2026 (36) J.P. Morgan, securitized products outlook for data center ABS and CMBS, 2026 (37) Structured Finance Association, data center securitization market note citing Barclays Research, July 23, 2026 (38) AFCOM, State of the Data Center report, 2026 (tenth annual edition) (39) Uptime Institute, Global Data Center Survey 2026, and Uptime Institute analysis of pre-2015 enterprise data center AI readiness as reported in secondary coverage (40) STL Partners, The Retrofitting Roadmap, May 2026 (41) Cyxtera Technologies, asset purchase agreement with Brookfield Infrastructure, November 1, 2023, and U.S. Bankruptcy Court for the District of New Jersey plan confirmation, January 2024 (42) Sungard Availability Services, Chapter 11 sale of eight data centers to 365 Data Centers, U.S. Bankruptcy Court for the Southern District of Texas, August 2022 (43) Interagency Appraisal and Evaluation Guidelines, 75 Fed. Reg. 77450, December 10, 2010 (44) U.S. Small Business Administration, SOP 50 10 8, Lender and Development Company Loan Programs, effective June 1, 2025 (45) Columbia Law School Sabin Center for Climate Change Law, Decommissioning Data Centers, August 2026 (46) Moody's Ratings, Fitch Ratings and S&P Global Ratings, rating actions on CoreWeave, Inc., 2025 to 2026, and CoreWeave Form S-1, March 2025 (47) Fitch Ratings, Data Center ABS and CMBS rating criteria, September 16, 2025, and S&P Global Ratings, Data Center Securitizations Methodology, June 13, 2024 (48) Meta Platforms, Inc., press release on the Hyperion data center financing with Blue Owl Capital, October 21, 2025 (49) Office of the Governor of Ohio and Ohio Tax Credit Authority, pause on data center sales and use tax exemption applications, May 27, 2026, and Good Jobs First cost analysis of the Ohio exemption, 2026 (50) Arizona HB 4168, Laws 2026, Chapter 140, signed June 13, 2026, and Arizona HB 2631 (2026 session) (51) Georgia SB 410 (2026 session) and Georgia Department of Audits fiscal estimate of the data center sales tax exemption, fiscal 2026 (52) Commonwealth of Virginia, 2026 to 2028 biennial budget as signed June 30, 2026, including the Data Center Electricity Consumption Tax, and Virginia HB 153 (2026) (53) National Conference of State Legislatures, 2026 state data center legislation summary (54) ALEA Institute, Moratorium Nation local moratorium tracker, August 19, 2026 release (55) Good Jobs First, local data center moratoria and restrictions tally, 2026 (56) State of California, veto message on AB 93 (Papan), October 11, 2025, and SB 57 (Padilla) as signed (57) Loudoun County Board of Supervisors, Zoning Ordinance Amendment Phase 1 on data center uses, March 18, 2025, and Phase 2 schedule (58) Prince William County Board of County Supervisors, noise ordinance amendment, October 28, 2025 (59) Virginia Department of Environmental Quality, eastern coastal plain groundwater study for data center cooling, July 28, 2026 (60) 13 CFR 121.201, small business size standards, and U.S. Census Bureau, NAICS 518210 definition (61) U.S. Small Business Administration, Information Notice 5000-880695, SOP 50 10 8.1, effective for loan numbers issued on or after October 1, 2026 (62) 7 CFR Part 5001 and U.S. Department of Agriculture, fiscal year 2026 Business and Industry guarantee terms, 91 FR 11272