When Software Can Be Used Everywhere
This series examines a specific subject: how U.S. firms adopt custom software, how that adoption relates to the number of establishments they operate, and why the relationship matters for concentration. The central question is not whether software is “productive” in the abstract. It is whether one investment can be reused across many parts of the same firm, changing the economics of where the firm ends and how far it expands. By the end, the reader should be able to distinguish what the Census evidence observes, what the model infers, and which real-world conditions limit the result.
The question is not how much software a firm buys
Software investment and firm concentration have grown in the same broad historical direction. The Census study reports that firm-level investment in custom software rose from roughly 2% of U.S. nonresidential fixed investment in 1980 to more than 8% in 2021. Over approximately the same period, the shares of sales and establishments associated with the largest 1% of firms rose by about ten percentage points. That co-movement is important, but it is not yet an explanation. Large firms may buy more software because they are already large; software may help them grow; or both outcomes may reflect other changes in markets and organizations.
The more revealing question is therefore about the firm boundary. Imagine a retailer developing an internal system for inventory, pricing, and replenishment. A forklift has to be purchased and positioned for each place where it is used. The software system, by contrast, may be developed once and then used by stores in several locations at the same time. Adding a store still requires workers, buildings, equipment, and local management. But the software investment need not be recreated in full for every store. The value of the original investment can spread across the firm's establishments.
That difference creates a reason to study establishments rather than software spending alone. A firm with one establishment and a firm with one hundred may both report a software investment, but the economic role of that investment is not identical. In the larger firm, the cost can be shared across more operating units. The software may become part of the infrastructure that makes a multi-establishment organization viable, not merely another item in its capital budget.
This is the tension at the center of the research. If software can be reused, larger scope may make adoption more attractive. Yet among firms that have already adopted, software can account for a smaller share of total capital expenditure as scope rises, even while the absolute level of software investment is higher. Adoption and intensity are different margins. Treating them as one statistic makes the evidence look contradictory when it is actually pointing toward a change in the firm's cost structure.
It also helps to separate scope from scale. Scale usually refers to how much a firm produces or how many resources it controls. Scope refers here to the number of establishments under one firm. A company can grow by making one site larger, by adding sites, or by doing both. Non-rival software is especially relevant to the second route because the investment may support several operating units without being replicated in the same proportion. The question is not whether large firms are efficient in every respect. It is whether a particular input changes the relative appeal of one organizational arrangement over another.
That distinction gives the concentration statistic a specific meaning. If the largest firms own a larger share of establishments, the result says that operating units are increasingly grouped inside a small number of corporate boundaries. It does not, by itself, tell us whether those firms charge higher prices, earn more profit, or eliminate competitors. The Census paper uses concentration as an outcome connected to firm scope, not as a synonym for market power. That restraint matters because a technology can make multi-establishment organization cheaper while still producing uncertain effects on consumers, workers, and potential entrants.
The opening puzzle therefore has three levels. At the firm level, does custom software make adoption more attractive as scope grows? Within adopters, how does the input change the mix of software, labor, and physical capital? Across firms, do those choices shift the distribution of establishments and sales? A convincing explanation has to keep all three levels connected without assuming that a pattern at one level automatically determines the next.
The property that changes the firm boundary
The relevant term is non-rival. An input is rival when one unit's use prevents that same unit from being used elsewhere at the same time. A cash register or a forklift is rival in this practical sense. Custom software is non-rival within the firm when the same investment can operate simultaneously across several establishments. The distinction is about reuse, not about whether something is digital, intangible, or technologically sophisticated.
That qualification matters. The Census paper's baseline treats custom software as non-rival and non-excludable within a firm, while remaining excludable across firms. A company can use its own investment across its sites, but a competitor cannot simply use it without permission. This is different from an idea that spills freely across the entire economy. The mechanism is increasing returns within an organizational boundary, not an assumption of economy-wide free access.
Nor does every software purchase have this property to the same degree. Off-the-shelf software, vendor licensing, and establishment-specific customization can change how much of an investment is reusable. A license priced per site may turn some of the apparent sharing benefit into a recurring cost. A system built for one operation may require substantial adaptation before another operation can use it. The paper later treats partial excludability and specificity as important qualifications. For the opening model, however, the key question is simple: does adding another establishment require a proportionate new software investment, or can the existing capability travel with the firm?
A small thought experiment makes the distinction concrete. Suppose two firms each build an internal scheduling system. One firm operates a single site; the other operates twenty. The system may improve both firms' coordination, but the second firm can spread the development cost across twenty operating units. If the system also makes it easier to coordinate those units, the software has two effects at once: it reduces an input cost per establishment and makes a broader organizational form more workable. That is why the paper treats software as a variable production input that is shared across establishments, rather than simply as a fixed cost attached to the firm.
The mechanism does not imply that a firm should expand without limit. More establishments create management burdens. They may compete for the same customers, reducing the value of adding another site. The firm must also decide whether the software fits its operations well enough to justify adoption in the first place. Non-rivality creates a possible advantage to scope; it does not erase the costs of governing scope.
There is a useful difference between calling software a fixed cost and calling it non-rival. Both ideas can suggest that an investment is shared across output, but they answer different questions. A fixed cost lowers average cost because it is paid once for a production operation. A non-rival input changes the way the cost is distributed across a firm's establishments while still entering production as an active input. That distinction lets the model ask how software intensity and factor shares change as the firm adds sites, rather than assuming that software simply disappears into an overhead category.
The boundary of reuse is consequently an economic boundary. Within a firm, a common data model or workflow may be valuable precisely because it coordinates otherwise separate establishments. Across firms, the same capability can remain protected by ownership, secrecy, licensing, or the cost of rebuilding complementary routines. This is why the mechanism can produce increasing returns inside a firm without implying that the entire economy has increasing returns from free access to the same software. The firm may scale because it can share an asset that its rivals cannot simply appropriate.
But reuse is never just a property of code. It depends on compatible processes, data, contracts, training, and local decisions. A system may be technically deployable at a second site yet require so much adaptation that most of the expected saving disappears. The more those frictions grow, the less the investment behaves like a common input and the more it resembles a collection of site-specific projects. The economic question is therefore about effective reuse, not about the label attached to the product.
How to read the evidence
The research uses two complementary lenses. First, the authors link the Annual Capital Expenditures Survey to the Longitudinal Business Database to observe firm-level investment and establishment structure. Their focal measure of custom software is vendor-customized and own-account software intended for internal use; it excludes prepackaged software. They primarily measure firm scope by the number of establishments. This is a deliberate, narrow measurement choice: it captures a concrete organizational boundary, but it is not a complete measure of products, industries, expertise, or brands that a firm may control.
Second, the authors build a model in which firms choose whether to adopt custom software, how much to invest, and how many establishments to operate. The model is not a replacement for the data. It is a way to ask a question the data alone cannot answer: if the rental cost of software changed, how much of the historical movement in firm scope or concentration could follow from that change under a specified set of assumptions?
The distinction between those lenses should remain visible. The data can show that adoption rises with scope, that adopters' software share falls with scope, and that software-intensive industries tend to be more concentrated. Those are observations or conditional associations. The model can show how those patterns might arise from cost sharing and what a counterfactual price change would do inside the modeled economy. It cannot turn a co-movement into proof that software alone caused concentration.
This also explains why the subject is larger than a technology-adoption rate. A rate answers whether firms crossed the adoption boundary. Intensity asks how much software matters within firms that crossed it. Scope asks how many establishments a firm operates. Concentration asks how activity is distributed across firms. The series follows the connections among these questions because the economic significance of non-rival software appears in their interaction.
The measurements themselves impose discipline. The Census study focuses on capitalized internal-use custom software, not every subscription, app, or digital practice that a business might use. That choice makes the input more comparable to an investment, but it also means that the findings should not be casually extended to all forms of digitization. A firm may use a widely available cloud service without owning a reusable capability of its own; conversely, a modest-looking software expenditure may embody a workflow that becomes valuable only when deployed across a network of establishments.
The broader adoption question is also part of the subject, even though it cannot be answered by the U.S. firm-scope data alone. Evidence from the OECD's study of UK SMEs shows why a potentially scalable technology may remain concentrated among firms with the money, skills, confidence, and local support to implement it. Cost, relevance, vendor trust, peer validation, and regional capability affect whether a firm reaches the adoption threshold at all. Those conditions do not replace the Census mechanism; they help explain why its benefits may arrive unevenly.
The next step is to inspect the Census measures closely and see the two-margin pattern in the data. Once that pattern is clear, the apparent paradox becomes a useful clue: software can be more attractive to a larger firm while occupying a smaller share of that firm's investment. The explanation lies in what can be reused, what still has to be repeated, and how the firm chooses its boundary.