US Data Center Construction Spending After ChatGPT?

Census files data centers under private office construction as buildings that contain the hardware for storing, processing and transmitting digital information. The value covers the shell, electrical work, air-conditioning, emergency backup power, design fees, and interest and taxes paid during construction. The definition excludes racks and servers.

Epoch AI Model · Servers 55.9%, Facility 30.2%

Epoch AI's May 2026 cost model puts a stylised US hyperscaler 1 GW AI site at $37.9bn up front: servers $21.2bn (55.9%), facility $11.4bn (30.2%) and network equipment $4.9bn (13.0%). The facility line is built from a data center construction cost index, close to what Census counts. Epoch says it is not an estimate for any specific facility.

Microsoft describes the same split. Of $41bn of capital expenditure in April to June 2026, including finance leases, roughly two thirds went on short-lived assets, mainly CPUs and GPUs, CFO Amy Hood said. The rest was long-lived assets, which covers more than buildings, so a third is a ceiling for the building share.

Headlines need the same care. The IEA says capex at five large technology companies passed $400bn in 2025, while Census puts all US private data center construction at $49.7bn that year in current dollars. Company capex includes servers and sites abroad, so the two are an order-of-magnitude check, not a ratio.

The Before window is not a flat baseline. In real terms, US data center construction rose about 4.4 times from $2.5bn in 2014 to $11.1bn in 2019, so the window opens after a long climb. Inside it, monthly spending still grew 1.45x, from $0.83bn in January 2019 to $1.21bn in October 2022.

Most of that growth came late. Calendar-year spending in real terms went from $11.1bn in 2019 to $14.3bn in 2022, only 1.29x, and monthly spending averaged $1.18bn across January to October 2022 against $0.92bn in 2019. The upturn began in early 2022, months before ChatGPT arrived in November.

Back-Cast History · Published Only Since July 2024

None of this history was watched as it happened. The AGC reported that Census released its first data center estimates in July 2024, with monthly values back to January 2014; before that, the buildings sat inside the broader office category. The 2014 to 2022 figures are a reconstruction, with no independent series to check them against.

There is no step at the pivot. October 2022 stood at $1.21bn and November at $1.25bn; the line simply keeps rising and gets steeper. Monthly spending first reached $2bn in September 2023, $3bn in June 2024, $4bn in April 2025 and $5bn in April 2026, then passed $6bn in June 2026.

Calendar years show the same staircase in real terms: $21.9bn in 2023, $37.0bn in 2024 and $51.6bn in 2025, against $14.3bn in 2022. January to August 2026 already totals $45.6bn, 37% above the same eight months of 2025.

The steepest stretch is the latest. Spending rose from $4.84bn in March 2026 to a preliminary $7.54bn in August, up 56%, and the final ten months alone, November 2025 to August 2026, add up to $54.6bn, more than the whole $47.1bn Before window. That is also where revision risk is highest.

November 2022 is where the comparison splits, not a switch the industry flipped. Census does not separate AI data centers from cloud, colocation and enterprise ones, so "Before AI" and "After AI" are labels for dates. Microsoft's Hood said customers increasingly build solutions that use both AI and non-AI infrastructure.

Build times rule out a quick effect. Epoch AI found that gigawatt-scale sites which broke ground in the past three years took 1 to 3.6 years to reach 1 GW, and Census records value put in place, meaning work done in the month, not contracts signed. Spending in early 2023 mostly reflects projects approved earlier.

The defensible reading is that a buildout already accelerating in 2022 kept going, and AI demand is now the reason companies give for it, as Alphabet's June 2026 equity raise shows. The later months, at several times the pre-2022 pace, are where an AI link is plausible. The early ones are not evidence of it.

Census publishes no state breakdown for this category that we found, so geography comes from brokers' trackers, which count megawatts rather than dollars. JLL's midyear 2026 report says 77% of North American capacity under construction is in frontier markets, naming West Texas, Ohio, Louisiana, Indiana and the Carolinas. CBRE's first-half 2026 report calls Atlanta the most active construction market in North America for the first time, with nearly 2,900 MW under way, overtaking Northern Virginia.

Gigawatt figures depend on who is counting. CBRE counts 7,481 MW under construction in primary markets, up 24.8% on a year earlier. JLL counts more than 66 GW under construction across North America. Sightline's estimate, reported by Bloomberg on 1 April 2026, had only about 5 GW of roughly 12 GW due online in the US in 2026 under construction, though SemiAnalysis answered that its denominator was flawed and that the two largest hyperscalers alone exceed 5 GW.

In April to June 2026, Alphabet, Amazon and Meta spent $129.2bn on property and equipment, up 82% from $71.2bn a year earlier. Census-measured data center construction rose 36% over the same quarter in current dollars, from $12.5bn to $17.1bn. The scopes differ, so the gap fits money shifting to chips without proving it.

The One Peak Forecast · $86bn in 2026, $63bn by 2030

If more of each dollar goes to hardware, construction can plateau while total capex rises. The only explicit forecast of a peak we found, from Moca Systems via Bisnow in November 2025, has US data center spending at $86bn in 2026 and $63bn by 2030. Census's seasonally adjusted rate for preliminary August 2026 is $85.0bn a year, level with that peak, and the line has not turned. The forecast predates the current data vintage and does not say whether it is nominal or real.

Lease accounting adds a wrinkle. Microsoft is extending data center useful lives from 15 to 25 years, moving more leases from finance to operating and cutting its calendar 2026 capex guidance to about $175bn with underlying investment unchanged. A building a developer constructs and Microsoft leases appears in Census as construction but may not appear in Microsoft's capex.

The strain shows in cash flow. Alphabet's second-quarter free cash flow was -$5.9bn: $39.1bn of operating cash flow against $44.9bn of purchases of property and equipment. It raised $49.6bn of net equity in June 2026, with proceeds earmarked, among other uses, for capital expenditures to scale AI infrastructure. Amazon's trailing-twelve-month free cash flow is -$7.6bn, down from +$18.2bn a year earlier.

Regulators are naming the risk. On 30 September 2026 the Bank of England's Financial Policy Committee said the rapid increase in AI-related debt issuance had raised capital markets' exposure to AI. Morgan Stanley put global AI-related issuance at about $450bn by early September, double 2025's, according to reports. After July's sharp falls in AI and semiconductor stocks, the committee warned, a more significant shock could trigger a sharper repricing.

The revenue side is a projection. Bain & Company's 2025 technology report says $2tn of annual revenue would be needed by 2030 to fund the computing power AI demand implies, and that even with AI-related savings the world is $800bn short. None of this points to a stop in the Census series; it points to who carries the cost if revenue arrives late.

Brokers see demand, not a bubble. JLL's midyear 2026 report says concern about a potential bubble "should be viewed in the context of 99% occupancy, particularly when the largest data center tenants rank among the world's most profitable and highest-rated companies." It reports 1% vacancy for the third consecutive year and says most tenants securing space today are contracting for 2028 deliveries.

CBRE's first-half 2026 data agree. Vacancy held at 1.4% across primary markets, and more than 80% of capacity under construction is already preleased, up from 74.3% a year earlier. The two brokers measure different universes, which explains 1% against 1.4%. Tenants signing before buildings finish is not what speculative overbuilding looks like today; the open question is whether those tenants can keep paying.

Census measures value put in place, the work done in a month. It is a voluntary survey of owners, nonrespondents are imputed, and private nonresidential estimates are increased by 25% to cover projects missing from the sample frame. The series used here is not seasonally adjusted, so it shows actual monthly spend.

Inflation Basis · CPI in August 2026 Dollars

Every dollar figure is deflated with the consumer price index (FRED CPIAUCSL) into August 2026 dollars, so the comparison does not partly measure the 2021 to 2023 inflation spike. BLS never published October 2025 CPI because of a lapse in appropriations, so that month is interpolated as the geometric mean of September and November 2025 and flagged in the data.

A construction-cost index, an equal-weighted blend of BLS producer prices for new industrial and new warehouse building, would shrink the Before window's growth to 1.15x instead of 1.45x, because building costs rose about 48% over it against about 18% for CPI. The headline holds either way: 5.41x against 5.01x from November 2022 to June 2026, matching Our World in Data's five-fold finding.

August 2026 is preliminary and June and July are revised, so the last three points can still move. Census revises this series, and the current vintage sits well above the figures quoted in earlier reporting, so numbers from older articles should not be compared with it. We could not find a Census note dating that revision, so we do not attribute it to a particular release.

What the data settles: US data center buildings have gone up far faster in the latest 46 months than in the 46 before, 3.37x in total, and much of that is a recent surge, with the final ten months exceeding the whole earlier window. The buildout was already growing, and AI demand is the explanation owners now give for the later acceleration.

What it does not settle: how much of the building is AI, how much of AI spending the buildings represent, or whether the climb continues. The buildings are about a third of a modelled AI site's bill, Census cannot split AI from cloud, and the financing evidence puts the risk in who pays rather than who rents. The next Census releases, which revise August's preliminary figure, are the cleanest test.