The Changing Global Division of Labor in Software: Emergence and Diffusion of New Programming Skills across IT Hubs

Global Software Skill Dynamics: How New Programming Expertise Emerges and Diffuses Across Cities

Software is often called "weightless." Its products can be traded globally at near-zero transport costs. This digital nature suggests that geography should not matter. Yet, the people who build software remain intensely concentrated in specific cities. This study explores the tension between the borderless nature of code and the physical reality of the developers who write it.

While software can be outsourced anywhere, the high-level knowledge required to build it might still favor dense urban hubs. Before this study, we lacked a way to track how specific technical competencies move across the globe. We did not know if software followed the same rules as traditional manufacturing or if it truly broke the link between skill and location.

The resolution gap in digital labor

Traditional economic geography often relies on coarse data. Researchers often use patent filings or broad industrial classifications to track innovation. These methods have significant blind spots. Patents are heavily skewed toward manufacturing. Industrial categories are often too broad to capture the nuance of a developer's daily work.

Furthermore, patent data involves massive time lags. There is often a long delay between an invention and its legal recognition. This makes it a poor tool for studying a field as fast-moving as software. The authors argue that we must look at specific skills to understand the shifting division of labor.

Without this granularity, we cannot see how a city moves from "general IT" to something specific like "machine learning." As seen in the rising demand for programming roles in the U.S., the scale of the sector is massive.

However, its internal composition is a moving target that traditional metrics fail to capture.

Mapping the software skill space

To overcome this, the researchers leveraged a massive dataset from Stack Overflow. They analyzed over 60 million questions and answers. Their approach to defining "skills" relies on a multi-stage statistical pipeline:

1. Tag Clustering: Instead of treating every individual tag as a separate entity, the authors use a Stochastic Block Model (SBM). This is a statistical method used to partition networks into communities of related items. Think of it like identifying "neighborhoods" of topics. If "spacy," "text-mining," and "word-embedding" always appear together, they form a single functional skill: Natural Language Processing. This process yielded a taxonomy of 237 distinct software skills .

2. User Profiling: The study links these skills to geography by geolocating three million users. By analyzing the questions a user answers, the authors build a "skill vector" for each person. This represents the proportion of their expertise dedicated to various domains.
3. Aggregating to Cities: These individual vectors are summed to create a skill profile for entire cities. To determine if a city is a specialist, the authors use Revealed Comparative Advantage (RCA). This metric compares a city's share of a specific skill to the global average.

Relatedness drives the diffusion of expertise

The study's most striking finding is that the "death of distance" in software is a myth. The diffusion of new skills is governed more by "relatedness" than by geographic proximity.

The researchers report that cities preferentially diversify into new skills that are closely related to the ones they already master .

Specifically, moving from the 10th to the 90th percentile of "relatedness density" increases the likelihood of a city entering a new skill by about 12 percentage points. This means a city's existing technical foundation is a powerful predictor of its future growth.

When looking at the speed of adoption for emerging skills, the authors used a Cox proportional hazard model. This model pits geographic distance against local capability. They found that relatedness density is the dominant predictor of how fast a city adopts a new skill. Meanwhile, the distance to the skill's origin has a negligible effect .

For example, a new skill originating in San Jose will reach a city in Bengaluru nearly as quickly as it reaches Seattle. This happens if Bengaluru already has a foundation in related technologies.

However, the "nursery" effect remains real. Brand-new skills do not pop up randomly. They emerge primarily in large, highly diversified cities like London and San Jose .

These cities act as incubators. The sheer variety of existing expertise allows for the "recombinant search" necessary to spark a breakthrough.

Limitations of the digital trace

While the dataset is vast, it is important to recognize its limits. First, Stack Overflow users are a self-selected subset of the global developer population. The authors acknowledge that usage may not perfectly mirror the entire global economy. They also note that usage of the platform has recently declined due to the rise of generative AI coding assistants.

Second, the study uses imputed salary data from a U.S.-based survey to value skills. These values represent how the U.S. market values a skill. They do not necessarily reflect what a developer in Berlin or Mumbai actually earns. Therefore, the skill rankings reflect the technical sophistication of a city's portfolio.

They do not reflect local cost of living or absolute wealth.

Finally, the analysis is centered on individual skills. It does not account for the organizational dynamics of teams. It also does not track how firms move talent between departments. These are still critical components of how industrial clusters function.

The verdict: Capability over proximity

The evidence suggests that if you want to predict where the next major software trend will land, do not look at a map. Look at a skill matrix. The study demonstrates that the software industry follows a two-stage evolution. First, there is a period of radical emergence in large, diverse urban "nurseries." This is followed by a period of widespread diffusion driven by local technical readiness.

For policymakers and educators, the takeaway is clear. Building a competitive tech sector is not just about attracting large companies. It is about cultivating a dense web of related technical capabilities. For developers, the findings suggest that career mobility is less about moving to Silicon Valley. It is more about mastering the "neighboring" skills that make your current expertise more versatile. The digital nature of the work has neutralized distance. However, it has heightened the importance of what you already know.