Did AI Change Where Software Is Built?

In a 2022 experiment, professional programmers were asked to build a web server in JavaScript. Among those who finished, the group without GitHub Copilot took 161 minutes on average. The group using it took 71, less than half as long. ¹

In everyday development work, researchers found a different kind of gain. Across three workplace trials involving 4,867 developers at Microsoft, Accenture and another large company, those given an AI coding assistant completed an estimated 26% more tasks. The size of the gain varied from trial to trial. ²

The tools changed too. Copilot suggested code as a programmer worked. Cursor Agent, made public in November 2024, could select relevant project context and run terminal commands. In May 2025, Codex's public preview offered to write features, fix bugs and run tests before returning a proposed change for human review. A developer could now hand over more than the next line of code. ^3 4 5

What did these years look like beyond one assignment or workplace? On GitHub, a push records an upload of changes to a public project. We follow those uploads across the same 149 economies from early 2020 to early 2026. They show when and where public activity grew, not how much of it an AI tool produced. ^7 8

In early 2020, those economies recorded 80.8 million public pushes in a quarter. By early 2026, the number was 319.2 million, nearly four times as high. A push can contain a small change or a large one; the count does not tell us how much code was written or whether it became released software. ^7 8

The total did not climb at a steady pace. By the third quarter of 2024, it had almost doubled from early 2020. In early 2025, quarterly pushes were 16% above a year earlier. In early 2026, they were 80% above (Figure 1). ⁷

India began with about one-fifth as many quarterly pushes as the US. Over the next six years it added 40.2 million pushes to its quarterly count, more than the 35.9 million the US added. The US remained ahead, but the gap between them had narrowed. ⁷

The changing balance is clearer in world shares. In early 2020, the US supplied 26.5% of pushes across the 149 economies. China followed at 6.9%, then India at 5.5%; North America and Europe together supplied nearly two-thirds. By early 2026, the US share had fallen to 17.9% while India's had risen to 14.0% (Figure 2). ⁷

The shift went further than India. Countries outside the top ten supplied 35% of pushes in early 2020 and 43% in early 2026. The map below lets those smaller country totals come into view without changing the scale as the years pass. ⁷

The peaks rise across almost the whole map. Yet the US and India remain its two tallest peaks at the end. A broader spread of activity did not make the largest contributors small.

The last year of the animation was especially broad. Between early 2025 and early 2026, pushes rose in 148 of the 149 economies. Iran was the exception. The economy in the middle of the growth ranking recorded a 93% increase, nearly doubling its quarterly uploads. ⁷

The US and India supplied a third of the added pushes that year. But set both aside and the other 147 economies still show 77% growth, close to the full panel's 80%. The two biggest countries shaped the total without creating the surge by themselves. ⁷

Was this simply a higher level of activity, or were countries growing faster too? Compare the two years ending in early 2022 with the two years ending in early 2026. Annualised push growth was higher in the recent window for 110 economies and lower for 39 (Figure 4). Slower growth need not mean fewer pushes. ⁷

We use the same 149 economies in both windows. They accounted for 99.8% of all reported pushes in early 2026, so the result is not largely a matter of GitHub adding new places to the data. ⁷

The fastest-growing places were not necessarily the ones adding the most uploads. The 30 economies with the fewest pushes in early 2025 almost tripled their combined activity over the next year. That added 1.1 million pushes to a quarter. The 30 largest grew more slowly, yet added 113.7 million (Figure 5). A small starting total can produce a striking percentage; a large one can change the world total with a modest rate. ⁷

Even the name at the top of the growth ranking depends on where we set the starting line. Without a minimum, it was Libya. Among economies that began above 100,000 quarterly pushes, it was Saudi Arabia. Above one million, it was Japan. Each answer is correct for a different-sized group of countries. ⁷

Asia supplied 56.7 million of the additional quarterly pushes between early 2025 and early 2026, or 40% of the increase. Europe and North America each supplied about 23%. Asia's position had been changing for years. Across the full 149-economy panel, its share rose from 26.6% in early 2020 to 38.3% in early 2026 (Figure 6). It was already near that level by the third quarter of 2024, the last complete quarter before Cursor Agent's November launch. The line on the chart marks a date, not an estimate of the tool's effect. ^7 4

China did not follow the regional line. Its share of pushes fell from 6.9% to 0.9% over the six years (Figure 2, Share panel). Take China out of Asia's total and the rest of the region rose from 19.8% to 37.4% of the panel's pushes. The combined Asian figure hides how much ground its other countries gained. ⁷

That is a change in GitHub activity, not a measure of all coding in China. GitHub competes there with Gitee. A separate study of public Python functions estimated that about 20% of functions in its Indian sample were AI-generated by the end of 2024, while adoption in its Chinese sample progressed more slowly. The samples suggest different paths for one kind of coding work; they do not explain either country's push totals. ⁹

Asia's larger world share does not tell us whether more Asian countries shared in its growth. Between early 2023 and early 2026, pushes within Asia became more concentrated in fewer countries. Africa moved the other way: activity there spread more evenly across its countries (Figure 7). ⁷

Africa's public pushes rose 37.5% from early 2023 to early 2026. The world grew faster, reducing Africa's share from 7.6% to 4.7%. Both things happened while activity was spreading more evenly among African countries. The region was growing, reaching more of its own countries, and still losing ground in the world total. ⁷

The country detail makes that clearer. In early 2026, 29 of the 33 African economies in our comparison exceeded their previous quarterly push highs. Ghana, Kenya, Nigeria and Sudan had grown over the latest year but were still below earlier peaks. Small starting totals produced the fastest percentage gains. Nigeria, Egypt, South Africa and Kenya together supplied just over half of Africa's added pushes. A falling world share did not mean that activity was falling across the continent. ⁷

Europe offers another version of this story. Every one of its 39 economies in the comparison recorded more pushes between early 2025 and early 2026. Only seven gained world share, with Poland making the largest gain. Europe's combined share fell from 29.1% to 26.2% because growth elsewhere was faster. ⁷

In South America, the other nine countries in the comparison grew faster than Brazil over that year. Its share of regional pushes slipped from 61.5% to 58.3%. Brazil nevertheless supplied 54.7% of South America's additional pushes, more than the other nine countries combined. The fastest rate and the largest addition belonged to different places. ⁷

At the country level, Japan made the largest gain in world push share between early 2025 and early 2026. The US and India followed (Figure 8). ⁷

For Japan, the gain was a recovery. Throughout 2022, 2023 and 2024, it recorded more quarterly pushes than in early 2020 yet held a smaller slice of the growing world total. Figure 9 follows that share through the six years. ⁷

Japan's quarterly pushes then rose from 4.5 million in early 2025 to 10.4 million a year later. That was enough to lift its world share just above where it had stood in early 2020. ⁷

The US's share gain was briefer. From early 2021 through late 2025, its share had been lower than a year earlier in every quarter; early 2026 broke that run. Across the four quarters ending then, however, the US supplied 16.8% of pushes, almost unchanged from the preceding four quarters. The Philippines followed a steadier climb, gaining share in all 21 year-on-year comparisons from early 2021 to early 2026 and rising from 41st to 19th by quarterly pushes. A one-quarter reversal and a six-year ascent look alike on a map of latest-year gains, but they tell different stories. ⁷

India ranked second by total pushes in the final quarter of 2025. Yet it ranked 81st by pushes per million residents among 155 reporting economies with at least one million people. The second measure puts uploads in relation to population. It does not tell us what fraction of residents code. ^7 10

India was still far behind the US in pushes per resident, but that measure grew much faster in India between early 2020 and early 2026 (Figure 10). The chart follows selected countries from the main 149-economy panel, using a 2020 population estimate and a 2026 projection. It shows how their per-person activity changed, not a ranking of every country. ^7 11

More uploads could come from more accounts, more uploads from existing accounts, or both. Over the full six years, recorded account totals grew faster than pushes. In the final year the order reversed. Pushes rose 80% while accounts grew 31% (Figure 11). An agent might help one person prepare more changes, but these totals cannot identify tool use or show how productive any individual became. They include inactive accounts too. ^7 8

Pushes outpaced account growth in 144 of the 149 economies. Cabo Verde, Iran, Kyrgyzstan, Rwanda and Ukraine were the five exceptions. The pattern was widespread, but an account total is not a headcount of people actively writing code. ^7 8

A country can have many public projects without many recent uploads. In early 2026, China held 4.31% of public repositories in the panel but only 0.91% of its pushes. A repository can stay on GitHub without receiving a new upload. Canada, meanwhile, gained world share of GitHub organisations over six years while losing share of accounts and repositories. Organisations include informal teams as well as companies, and GitHub assigns locations differently across these measures. Neither a project count nor an organisation count is a count of current coding work. ^7 8

The opening experiment showed a large time saving on one assigned task. Another experiment found the opposite result for a different kind of work. In early 2025, 16 experienced open-source developers worked on issues in repositories they already knew. With AI tools allowed, they took 19% longer to finish. ¹² A follow-up with newer tools found possible savings, but problems with who joined the study and how time was measured prevented a reliable estimate. ¹³ The participants, tasks and tools differed from the Copilot test. Neither experiment measures what caused a country's push count to rise.

Agents were producing visible work on GitHub by then. The AIDev study catalogued 932,791 pull requests attributed to five coding agents as of August 1, 2025. Each pull request is a proposed change; the count does not say how many were accepted or how many national pushes came from agents. ⁶

What happened to developers' work after adopting agents? An observational study followed more than 500,000 GitHub developers and compared adopters of interactive agents with similarly active developers observed a year earlier. It measured how each group's activity changed from its own earlier level, rather than comparing their raw output. That design gives a view of changes associated with adoption, not a randomized test. ¹⁴

For the interactive agents in the study, the estimates for weeks 21 to 30 after adoption were 153% for weekly commits, 47% for the number of repositories contributed to and about 20% for weekly release records, compared with matched developers' changes. A commit is a saved change; a release record marks a published project version. The release need not contain the adopter's own code, and neither measure counts successful products. ¹⁴

The researchers also observed more rewriting and more branches for trying alternatives. Their interpretation was that agents made it easier to test possible changes, while deciding what to keep and publish remained a separate step. This does not explain the national push totals. It does explain why we should resist treating every added upload as finished software. ¹⁴

GitHub also shows which kinds of projects attract contributors. TypeScript builds on JavaScript, adding checks that can catch certain errors before a program runs. In early 2020, just three of 93 comparable economies had more developers pushing to repositories containing TypeScript than to those containing Java. By early 2026, all 93 did (Figure 12). This is a change in the projects contributors touched, not a measure of the code they wrote in either language. ^7 15

One developer can appear in both language counts, and a repository can contain several languages. Project topic labels are looser still. Among the six economies reporting all three selected AI-related topics in early 2026, India led participation in repositories tagged "generative-ai", while the US led "llm" and "ai-agents". The tags are optional and overlapping. They identify projects, not which contributors used an AI tool. ^7 8

Knowing where someone uploads code is not the same as knowing whose project receives it. A contributor in India, for example, can work on a repository assigned to the US. To see that direction, we leave the country push totals and follow the link from contributor to repository. ⁸

GitHub reports cross-border contribution weight for that link. It combines public pushes and opened pull requests to someone else's repository when the contributor and repository are assigned to different economies. An India-to-US link means work from an India-assigned contributor went to a US-assigned repository. It does not mean the change was merged, and a repository's assigned location need not describe every member of its team. ⁸

Kenya shows why the distinction matters. Its public pushes rose 63% between early 2025 and early 2026, while contribution weight on its published foreign routes present in both quarters fell 7.6%. Ukraine's pushes rose 21% as weight on its matched routes fell 4.7%. These are changes on the routes we can compare at both dates, not a count of either country's entire foreign collaboration. More uploads need not mean more recorded work on those existing links. ⁷

The largest routes in early 2026 converged on the US. Nine of the ten biggest published country-to-country links ended at US-assigned repositories. The tenth ran from the US to India. Looking beyond those top routes, US destinations received 47.0% of all published cross-border contribution weight between individual economies. Their share had been 47.5% in early 2020. Figure 13 begins with the six largest sending economies so the directions are legible; these six are not the denominator for the 47% figure. ⁷

Five of those six senders, Germany, the UK, Canada, India and France, sent more than half their recorded cross-border weight to US-assigned repositories. The sixth sender was the US. The pattern also survives a stricter comparison using only links reported in every quarter of the six years. US destinations received about 50% of weight on those routes at the start and 51% at the end. That keeps the routes fixed rather than letting new or disappearing links change the mix. ⁷

India came closest to matching the US in uploads. In early 2020, its quarterly push count was 21% of the US count. By the third quarter of 2025, it had reached 98%, almost level (Figure 14). But a near-match in uploads did not make India an equally large destination for work from abroad. ⁷

The near-parity did not last. By early 2026, India's pushes were 78% of the US count. India had not stopped growing: its quarterly pushes were 32% higher than in the third quarter of 2025. The US count rose 66% over the same stretch, reopening the gap. ⁷

Other comparisons show how substantial India's presence had become. From mid-2022 onward, it had more developers pushing to repositories containing Java than the US. An uninterrupted lead in JavaScript began in the third quarter of 2023; Python and TypeScript followed in the second quarter of 2025. The US remained ahead in C, C++, Go and Rust. These overlapping counts concern contributors to repositories containing each language, not how much code they wrote in it. ⁷

The Python-function study offers a second, narrower view. It found India catching up from 2023 in the estimated share of sampled functions made with substantial AI assistance, though the US share remained higher at the end of 2024. The study stops before India's later lead in Python-project contributors and its near-match in pushes. It cannot tell us whether AI caused either change. ⁹

The link between the two countries makes the destination question concrete. In early 2026, the India-to-US route was the fourth-largest published cross-border link. It carried about 2.5 times the contribution weight of the US-to-India route. India was approaching the US as a source of public uploads, while US-assigned repositories continued to draw much more recorded work from India than the reverse. ⁷

The geography of uploads has widened. Countries outside the top ten now supply a larger share of them, and India briefly came within reach of the US total. The geography of where international work goes has moved much less. US-assigned repositories still receive nearly half the published cross-border contribution weight. Agents may make proposed changes easier to produce, but these records cannot tell us how many came from agents or which survived review. They can show the difference between more code being uploaded around the world and the projects to which that work flows. ⁷