What aren't
we building?

A data center looks like a building and is often discussed as an abstraction. The investment is described in billions, the capacity in megawatts, the work in model training or cloud services. Yet its construction requires ordinary scarce things: a serviced site, electrical contractors, transformers, concrete, credit and a connection to a power system that already has customers.

That makes the opportunity-cost question legitimate. It does not make the answer automatic. New investment can mobilize workers and finance, raise productivity elsewhere, or help pay for infrastructure that serves other users. It can also bid resources away from homes, factories and public works. Both mechanisms can operate at once, in different places and over different horizons.

The first distinction is between a large growth rate and a large share. In July 2026, private data-center construction was running at a seasonally adjusted annual rate of $75.2 billion. That is almost one-tenth of private nonresidential construction, but 5.8 percent when the denominator includes public nonresidential work. Neither number describes the whole investment account, and neither includes the servers housed inside the buildings.

The monthly Census series measures work put in place, not announcements, project starts or the value of every component on a campus. It counts data-center buildings within office construction and excludes servers, racks and land acquisition. Calling this an AI investment series would add a classification that the source does not contain.

Three periods help organize the history without pretending to identify a treatment. Cloud expansion was already underway in 2015-2019. Pandemic digitization, supply disruptions and construction inflation reshaped 2020-2022. Generative AI accelerated infrastructure demand from 2023, but ordinary cloud computing, storage and other digital services did not stop. A shaded post-2022 period is a date marker, not an estimate of AI's causal contribution.

Crowding-out is most plausible where supply cannot respond quickly. A national pool of investment can grow while one utility has no near-term connection capacity. A country's construction workforce can expand while a particular project waits for an electrical contractor. Commercial land can be plentiful nationally and scarce in an established connectivity cluster.

The U.S. county evidence follows 694 counties with at least 25,000 private-sector employees in 2015. Eligibility is determined before the recent boom, not by selecting places with striking later outcomes. The annual geography is held consistent; Connecticut's changing county boundaries and U.S. territories are excluded. Suppressed QCEW wages and employment stay missing, even when their source cells contain zeros.

The footprint on this map is intentionally modest in its claim. NAICS 518210 covers data processing, hosting and related infrastructure services. Establishments in that industry are neither a census of data centers nor a measure of operating megawatts. Employment and establishment counts may reflect business organization as well as physical infrastructure. The map locates selected markets; it does not pretend to inventory every campus.

Electricity offers a separate test. EIA's annual state data distinguish residential, commercial and industrial retail sales and prices. The commercial category rises sharply in some data-center markets, but it also contains many other customers. Its change cannot be assigned wholesale to data centers, and industrial electricity cannot be inferred by subtracting an estimated data-center load from a different total.

Power construction and generating capacity are potential complements. Their growth can relieve a constraint, though the timing, location and firmness of capacity matter. One nameplate megawatt of generation is not automatically one megawatt of available, reliable connection capacity for a new load. The research files retain nameplate and summer capacity separately and do not attribute their expansion to data centers without a design that can support that claim.

Ireland supplies a clearer electricity numerator than most countries. Data centers consumed 7.663 TWh in 2025, about 23 percent of metered electricity, compared with roughly 5 percent in 2015. That is a large reallocation of the consumption mix. Yet other metered customers also consumed more electricity over the decade. A rising share and a growing remainder can coexist.

Housing adds counterevidence to a literal zero-sum story. The latest frozen dwelling-completion series reports 36,215 completions in 2025, revised from 36,284 in the initial release. Completions increased rather than collapsing. That observation does not tell us how many more homes might have been completed under a different allocation of construction and grid resources; it simply rules out treating every simultaneous increase in data-center use as an observed fall in all other activity.

Great Britain's new meter-matched estimates make the geography unusually concrete. External-serving data centers consumed 4.46 TWh in 2024, about 1.79 percent of metered grid electricity. In Slough, the corresponding share was 65.16 percent. Hillingdon's share was 28.13 percent. None of those local figures is described adequately by the national average.

The measurement remains incomplete. DESNZ labels the series Official Statistics in Development and excludes enterprise data centers used internally by their owners. ONS explains why investment cannot yet be isolated: the buildings, equipment and software can have different economic owners, and their reporting units can belong to different industries. A whole-economy hardware investment series cannot repair that missing classification by being renamed.

Australia makes the asset distinction visible from another direction. ABS identifies a surge in equipment investment in data services, inside the wider information industry. Its construction discussion uses commercial buildings not elsewhere classified, a broader building category. The two series describe different assets and populations; neither permits subtracting a universal data-center category from all investment.

Imported servers also matter for interpretation. They add to investment and the capital stock, while their imported value is deducted in the expenditure account for GDP. Domestic construction, imported equipment and future digital-service output therefore contribute through different channels and at different times. A large capital-expenditure announcement is not the same as an equally large addition to domestic value added.

Frankfurt's planning department states the land mechanism directly. Its June 2022 concept was adopted to guide data-center development while accounting for competition with other commercial uses. It distinguishes suitable, restricted and exclusion areas, with binding planning rules to follow. This is primary evidence that land competition is recognized in local policy, not an estimate of how many factories were displaced.

The European reporting scheme creates an additional evidence source for energy performance, but a threshold-based reporting database is not a balanced 2015-2026 national investment panel. For the Netherlands, CBS provides a dedicated-connection electricity series: 4.58 percent in provisional 2024 data. It excludes facilities embedded in institutions such as universities and hospitals. Comparisons must preserve those definitions, not erase them for a clean ranking.

Canada's clearest constraint in this source set is provincial and prospective. In June 2025, Alberta's system operator reported 29 proposed data-center projects requesting more than 16 GW, and announced an interim additional large-load connection limit of 1.2 GW through 2028. The operator tied the limit to reliability under a conservative supply-and-demand scenario.

That is a load-connection constraint, not a generation-interconnection queue or 16 GW of operating demand. The proposals would not necessarily all be built or require connections at the same time. A comparable ERCOT history could not be verified for this edition. Alberta therefore provides a documented case of restricted access, rather than a proxy for the entire North American grid.

The evidence supports a narrower and more useful conclusion than a universal construction squeeze. Data-center investment is growing quickly; its physical demands can become highly concentrated; and some grid and planning authorities have explicitly limited access or redirected development. This observational county analysis does not establish a general housing crowding-out effect, let alone a count of factories and homes not built.

Opportunity cost is still real when its precise counterfactual is not identified. The policy question is which resource is binding, who bears the marginal cost, and whether the new project expands that resource or merely secures priority over its next user. Connection conditions, infrastructure cost allocation, workforce expansion and land-use rules operate on those margins more directly than a national argument over whether the boom is good or bad.

The next decisive evidence would connect phase-level construction and energized capacity to the applications delayed, withdrawn or priced out, with consistent local outcomes before and after. Until then, the responsible answer to what we are not building is neither a confident zero nor an invented number. It is a place-specific research question with an increasingly visible physical constraint.