How do demand and supply economics shape software markets?

I’ve been working in the software industry for a little over 3 years. I learned to program the traditional way: writing code line by line, getting stuck and searching for answers. When I started work in 2023, AI tools were mostly helpers that suggested the next few lines. From late 2024 they began handling whole tasks, and I’ve watched the way we build software change around them. The graph below is my picture of how demand for builders and their supply moved through 5 eras. No source counts both across that whole span, so it shows the shape of the argument, not measured numbers.

Demand for people who build software, and the capacity to build it, from the 1950s to the AI era

For most of the graph, demand sits above supply, and that gap is what kept programmers scarce. Early in the AI era the lines cross: AI multiplies how much working code people can produce. Writing code floods, while judging and owning it stays scarce.

Why does the supply and demand of software matter right now?

Both sides of the software market have moved since 2022, for 2 separate reasons. Demand for builders fell from its early-2022 peak as interest rates rose, and AI began writing a fast-growing share of new code. Anyone who builds software, studies it or is weighing a career in it is living through both changes.

The change shows in 3 figures. Software development job postings on Indeed, counted as an index where February 2020 is 100, peaked at 233.85 on 28 February 2022 and stood at 78.33 on 2 October 2026^[8]. In Python code on GitHub, one study estimates that AI-assisted code rose from about 5% of new functions in 2022 to nearly 30% at the end of 2024^[9]. And enrollment in computer and information sciences fell at every level in fall 2025, with undergraduate enrollment falling again in spring 2026^[10].

This post explains the mechanism behind those moves, era by era.

What do “supply” and “demand” mean for software?

In software, demand is how much people and businesses want programs and the people who can build them. Supply is how much of each can be provided. A copy of a program costs almost nothing, so what runs short is the capacity to build new software.

The way I think of it, the value of any skill depends on 2 things: how many people want it, which is called demand, and how many people can provide it, which is called supply.

Software is unusual on the supply side. The economist Hal Varian wrote in 1998 that “Information is costly to produce but cheap to reproduce,”^[1] a cost structure of high fixed costs and low marginal costs. Software is information in that sense, so copies are cheap, and building it never was. At a NATO conference on software engineering in 1968, a speaker reported T.J. Watson’s statement that OS/360 cost IBM over $50 million a year while it was being prepared, and at least 5,000 man-years^[11].

So software has 2 markets, and each has a supply side and a demand side:

The demand for builders is what economists call a derived demand: the demand for labor “is based on the demand for the good or service that is produced.”^[12] Applied to software, employers hire people who can build it because their customers want what the software does.

How did supply and demand change across software’s eras?

Across 5 eras, from mainframes to AI, demand for software mostly ran ahead of the people who could build it. Each boom left a wave of new graduates behind it. As I read it, the AI era is the first in which supply has moved faster than demand. Supply hasn’t grown gradually this time. It has exploded.

The eras and their boundaries are approximate. In short:

The rest of this section follows the first 4 eras, then how the golden era ended and why new supply arrives late.

Mainframes, about 1950 to 1980

In the mainframe era both sides of the market were small, and programmers were scarce from the start: by 1967, industry observers warned that the US had at least 100,000 programmers and needed at least 50,000 more, as the historian Nathan Ensmenger records^[13].

Demand was small because computers were few. Back then, they filled rooms and cost so much that only governments, militaries, large companies and universities could afford them. The counts that exist are estimates. A Cambridge University Press history reports earlier estimates of about 4,000 computers in the US in 1960 and 48,500 in 1970^[14], and a participant at the 1968 NATO conference put Europe at about 10,000 installed computers, a number growing 25% to 50% a year^[11].

Supply was still smaller. US institutions awarded 87 bachelor’s degrees in computer and information sciences in 1964-65^[4]. The US Census Bureau counts 450,000 workers in IT occupations in 1970^[15], every IT job and not only programmers. The report of that 1968 conference records “a considerable amount of debate”^[11] on what some members called the “software crisis,” with widely differing views; one participant, Fraser, said the demands of the industry seemed to be “going ahead a good deal faster than our progress.” In an undated essay, Edsger Dijkstra put the problem in terms of ability rather than headcount: “the programming task was outgrowing our programming capacities.”^[16]

Programs became a market of their own in this era. Over 100 programs were already on sale by 1967, by a count the Cambridge history takes from the trade press of the time^[14], and the 1968 conference report listed “the highly controversial question of whether software should be priced separately from hardware.”^[11] In April 1969 Applied Data Research sued IBM, arguing that IBM’s free Flowcharter program impeded independent software makers, and in June 1969 IBM announced that from 1 January 1970 it would charge separately for its software^[14], a move the Computer History Museum describes as separating software from hardware^[17]. One member of IBM’s own task force later recalled it as “primarily to avoid a Justice Department suit.”^[18]

Personal computers, about 1980 to 1995

Personal computers took software beyond large organizations and into homes, and the US count of new computing graduates rose to a peak in 1985-86 and then fell by more than 40% over the next decade.

In 1984, 8.2% of US households had a computer^[19]. The Census Bureau’s count of US workers in IT occupations rose from 781,000 in 1980 to 1.5 million in 1990^[15]. US bachelor’s degrees in computer and information sciences reached 42,337 in 1985-86 and were down to 24,506 by 1995-96^[4], a fall of about 42%.

The internet boom and bust, about 1995 to 2005

The web set off a surge in demand for builders in the late 1990s, large enough that the US raised its cap on skilled-worker visas twice. The bust after 2000 cut that demand back before the next wave of US computing graduates peaked in 2003-04.

The web was still young. By the International Telecommunication Union’s count, 6.5% of the world’s population was online in 2000^[2], and the number of users doubled in the 5 years to 2010 and was expected to pass 2 billion by the end of that year^[20]. In the US, Congress raised the cap on H-1B visas for skilled workers from 65,000 a year to 115,000 in 1998, and to 195,000 a year for fiscal years 2001 to 2003^[21]. After the bust, new approvals at employers under the cap fell well short of it: 163,600 in fiscal 2001 and 79,100 in fiscal 2002^[22]. Cheaper money did not bring the sector back: a 2003 study found that the Federal Reserve’s response to the high-tech crisis lifted housing and spending on durable goods but had “almost no effect on the high-tech sector.”^[23] US computing degrees kept rising, to 59,488 in 2003-04^[4].

Was the shortage real? The strongest doubt came from the US government’s own audit agency. In 1998 GAO found that the Commerce Department’s 1997 shortage report had “serious analytical and methodological weaknesses that undermine the credibility of its conclusion that a shortage of IT workers exists.”^[24] That report had put the need at about 95,000 new programmers, systems analysts, computer scientists and engineers a year from 1994 to 2005^[24]. GAO also said the weak report “should not necessarily lead to a conclusion that there is no shortage.” A 2013 study for the Economic Policy Institute found that US colleges provide “an ample supply of highly qualified science, technology, engineering and math (STEM) graduates,”^[25] and that IT workers earned about what they had 14 years earlier. Read together, the evidence points to a scarcity of particular skills that booms made look larger, not a permanent shortage of everyone who could program.

The golden era, about 2006 to 2022

From about 2006, cloud computing, smartphones and app stores (and at the end, the “move online” during the pandemic) kept demand for software ahead of the people who could build it. Builders were paid far above other workers.

Amazon launched its S3 storage service on 14 March 2006^[26] and announced EC2 in beta on 24 August 2006^[27], a service for renting computing capacity in the cloud. Apple’s App Store opened in July 2008 with 500 apps^[28] and offered more than 850,000 by May 2013. The share of the world online rose from 43% in 2015^[2] to 74%, about 6 billion people, in 2025^[3]. The pandemic, in my view, pushed demand to its peak, as businesses rushed to move everything online; in the US, that peak in software postings came in February 2022.

Supply grew, and builders stayed expensive. US bachelor’s degrees in computer and information sciences more than doubled from 2011-12 to 2021-22^[29]. In May 2025, by the Bureau of Labor Statistics’ survey of employers, the median US software developer earned $135,980 a year, against $50,980 for all workers^[5], about 2.7 times.

In the UK, the going rate the government sets for skilled-worker visas in programming and software development is £54,700 a year^[30]; the median for full-time employees, measured on a different basis, was £39,039 in April 2025^[31]. In India in the 1990s, researchers found software pay above other Indian industries and below US and European levels^[32], the gap behind an export industry. By the Reserve Bank of India’s survey, India’s software services exports reached $204.7 billion in 2024-25^[33], nearly 3 times the $71.4 billion of 2013-14^[34], in dollars of each year. Worldwide, exports of information and communication technology (ICT) services came to $1.33 trillion in 2025, by UNCTAD’s count^[35].

How did the golden era end?

In the US, demand for builders turned in early 2022, before ChatGPT, as interest rates began to rise. The golden era ended first on the demand side, with money. AI arrived afterward, on the supply side.

The peak in software postings, on 28 February 2022, came 16 days before the Federal Reserve’s first rate rise of 2022, on 16 March^[36]. Higher rates cut investment: a 2023 study found that a tightening of 1 percentage point cuts venture capital investment by about 25% over the following 1 to 3 years^[37]. Indeed’s economists tie the fall in US tech postings to that turn: “as the economic situation turned and the Fed began hiking interest rates, the pull-back in demand was just as sharp.”^[38] They weigh 2 candidate causes, an overhang from the earlier hiring boom as the period of zero interest rates ended, and AI. Two Google economists, writing for the Economic Innovation Group^[39], and economists at Stanford^[40] date the fall in hiring for the jobs most exposed to AI to the months after the Fed’s shift and before ChatGPT.

Why does new supply arrive years late?

Builders take years to train, so new supply reaches the market years after the demand that called for it. Across the eras, the US count of new computing graduates has risen and fallen in waves.

In the US, the Bureau of Labor Statistics names a bachelor’s degree as the typical way into software development^[5], and only 44% of 2015-16 bachelor’s graduates finished within 48 months of first enrolling^[41], so for most the usual route takes more than 4 years. In Stack Overflow’s 2017 survey, under half of professional developers had started their first coding job within 4 years of learning to program^[42]. The golden era fits these figures: supply could only grow as fast as people could learn, and demand grew faster than that.

Economists have a name for markets where producers decide on today’s price and deliver later, so that prices and supply swing back and forth: the cobweb model, which dates from 1930 and is associated with cycles in commodity markets^[43]. The National Academies applied it to engineers in a 2018 report, sketching how a jump in demand keeps wages high for about 4 years before new US graduates arrive^[44].

The US count of computing degrees runs back to 1964-65, and each era above left its mark on it:

After each peak the count fell by more than a third before it rose again: by about 42% from 1985-86 to 1995-96, and by about 36% from 2003-04 to 2008-09. Those waves are what the cobweb model predicts when supply arrives years late. Outside the US the records are short: Eurostat’s figures on people in the EU with an ICT education run from 2016 to 2025^[45].

How is AI changing demand and supply?

AI makes working code cheap and quick to produce. That floods the supply of code and of programs. Demand for software may still grow, but less of it needs a person to write the code.

How much code does AI write?

One independent measure estimates that AI wrote about 29% of new Python functions (the named blocks of code that each do one task) by US developers on GitHub by the end of 2024^[6]. That is the supply of code.

The study ran a classifier (a model that tells AI-written code from human code) over more than 30 million commits, saved changes to code, by 170,000 developers on GitHub. It was published in Science in 2026, with its figures still ending in 2024^[9], and it covers public Python code only^[46]. Its estimates differ by country:

The US leads, by what the authors call “a modest and shrinking lead over other countries.”^[6] The US figure was about 5% in 2022.

Who can build software now?

More people can build working software than ever, by every count that exists, though no count measures builders directly.

The research firm SlashData estimates just over 47 million developers worldwide in early 2025, 36.5 million of them professional, up from 21.8 million professional developers in early 2022^[47]. GitHub counts more than 180 million accounts and more than 36 million new ones in a year^[48], and says nearly 80% of its new developers use Copilot, the AI assistant it sells, within their first week. In a survey Zapier commissioned of 818 US workers who had built a work tool with AI coding tools^[49], about a third had no formal programming background. Lovable says 100,000 new projects are built on it every day^[50], and Replit says it has over 50 million users^[51].

My thought is that the change is larger than these counts show. With AI, someone with only a basic understanding can describe what they want and see a working prototype within minutes. It won’t be perfect, and it may not be secure, but it works. The number of people who can produce working software has multiplied in a way that no earlier era came close to. Founders are building their own first products, and domain experts, people who know a field deeply, are building tools for their own work. Every one of those good-enough tools is a project that, a few years ago, would have gone to a developer.

Does AI really lower the barrier to working software?

The strongest case against the flood is that AI-written code is often wrong or insecure, so the barrier to software that works safely has fallen much less than the barrier to software that runs.

Developers say so themselves. In Stack Overflow’s 2025 survey, 46% of respondents distrusted the accuracy of AI tools, against 33% who trusted it^[52], up from 31% distrusting a year earlier^[53]. In the same survey, 66% named AI answers that are “almost right, but not quite” as their biggest frustration, and 45% found debugging AI-generated code more time-consuming. Early peer-reviewed studies, on models from 2021 and 2022, found about 40% of the programs GitHub Copilot wrote in security-sensitive scenarios to be vulnerable^[54], and found that people with an AI assistant wrote less secure code while being more likely to believe it was secure^[55]. In a security vendor’s tests in 2025, AI-generated code introduced risky security flaws in 45% of tests, and newer models did no better^[56]. A security firm’s scan of over 5,600 public apps built on vibe-coding platforms^[57], where people build software by describing it to an AI, found more than 2,000 vulnerabilities and over 400 exposed secrets. Google’s DORA research team describes AI’s main role in software development as “that of an amplifier”^[58], magnifying the strengths of high-performing organizations and the dysfunctions of struggling ones.

This doesn’t undo the flood of code. It shows where the scarce work went: into judging whether the code is right and safe, which the section “Where does the scarcity go now?” takes up.

Is demand for coding work falling while demand for software grows?

Demand for software is still growing. Official outlooks describe routine coding moving to AI, so typing code is a shrinking share of what software buyers pay for.

On the side of software, the analyst firm Gartner forecasts worldwide software spending of about $1.47 trillion in 2026, up 15.5%^[59]. IDC expects software to take more than 33% of all ICT spending in 2026^[60]. The job site Indeed finds software development’s share of all job postings rising in most large developed economies it tracks, Germany and France excepted^[61]. The US Bureau of Labor Statistics builds the same reasoning into its projections: “increased productivity from the use of AI may lower prices and increase demand for software products.”^[62]

On the side of coding work, Canada’s official job outlook for New Brunswick says AI is likely to automate “routine, junior-level tasks such as code writing.”^[63] Skills England, the UK’s official skills body, describes “a shift away from routine coding and testing tasks toward roles focused on oversight, verification, and communication,”^[64] while still projecting more demand for programmers and software development professionals than for any other occupation in its digital sector, and calling the evidence for changes in employment “currently weak.”

One more change may bypass some coding work altogether. In November 2024, Anthropic open-sourced the Model Context Protocol (MCP), “a new standard for connecting AI assistants to the systems where data lives.”^[65] Microsoft^[66], OpenAI^[67] and Google^[68] added support in 2025, and in December 2025 Anthropic gave MCP to the Agentic AI Foundation, under the Linux Foundation^[69]. I expect assistants that act through MCP to take some work away from building screens and forms, in booking and ordering, for example: a customer tells an assistant what to order, and the assistant places the order with the business’s system without the customer ever opening a web page. That would take away much of the work of frontend (the screens people use) and design teams.

Is demand really lagging behind supply?

Where it’s measured, supply now runs ahead of demand for new programs and for advertised jobs. Demand for builders as a whole isn’t simply lagging, though: postings for senior and AI roles are rising again.

The clearest case is the market for programs. Apple says its reviewers handled more than 9.1 million app submissions, updates included, in 2025^[70], against over 7.7 million in 2024^[71], and attributes the surge to “powerful AI development tools.” Users did not keep pace. By Sensor Tower’s estimate, downloads across the iOS and Google Play stores grew 0.8% in 2025, to nearly 150 billion^[72], and were flat in the second quarter of 2026^[73]. Sensor Tower figures reported by the New York Times show new App Store apps doubling to about 560,000 in the first half of 2026 while downloads grew 2%^[74], and Appfigures data was reported as showing worldwide app releases up 60% in the first quarter of 2026^[75]. In mobile games, an analysis of Sensor Tower data found that a 77% jump in game releases brought only 14% more games earning over $20,000^[76].

The market for builders shows the same at the door. In the UK, graduate vacancies in every sector drew 140 applications each in 2025^[77], the same as the year before and up from 86 in 2022-23^[78], among the Institute of Student Employers’ members; its digital and IT vacancies drew 205 each in its 2024 survey^[79]. In data from the recruiting software Ashby, the average recruiter processed 291 applications per hire by early 2026, against roughly 100 in early 2021^[80]. On the forum Hacker News, the monthly “Who wants to be hired?” thread draws about twice the top-level comments of “Who is hiring?”^[81], where in 2022 it drew between about a third and a half as many.

Demand for builders is recovering, though, in one part of the market. US software development postings on Indeed rose almost 15% from late February 2025 while all US postings fell 7%^[61], and in June 2026 they were still about 27.5% below February 2020. Of the rise from May 2025 to May 2026, 71% came from senior roles and 37% from jobs with AI in the title; the 2 groups overlap. The Bureau of Labor Statistics projects 10% growth for software developers, quality assurance analysts and testers over 2025-35^[5]. So the lag shows in programs, in code-writing and in entry-level work, and the recovery is in senior and AI roles.

Where does the scarcity go now?

As per my understanding, the market for builders is splitting in 2: writing code from clear instructions, where the old scarcity is turning into a flood, and designing, judging and owning what the code does, where scarcity remains. The official projections draw a similar line.

The Bureau of Labor Statistics projects US employment of software developers to grow 10.2% from 2025 to 2035, and of computer programmers, the narrower job of writing code, to fall 7.3%^[7]. The 2 jobs differ in scale: about 111,000 programmers against about 1.7 million developers in 2025. The Bureau gives as one reason that companies are expected to use AI to automate repetitive programming tasks^[82]. Skills England describes the same move from routine coding toward oversight.

Here’s one small example from my team’s own work. A few months ago, we used Claude Code AI to build a proof of concept (POC) for connecting to external APIs. By hand, we would normally have spent about 7 to 8 days on something like this. With AI, we had a working version in 2 and a half days, covering the frontend, backend, and data integration for 4 authentication types: no auth, API key, bearer token, and basic auth. We left OAuth (client credentials and authorization code) out on purpose. It was a lot more complex, and we wanted to get the POC in front of people quickly.

A few days later, we realized our product needed to connect to HubSpot, Salesforce, Meta, and other APIs that rely on OAuth. So we went back and extended the POC. AI helped us move fast again, and the OAuth work took about 1 more day, where we would have guessed 3 to 4 days before.

Still, AI wasn’t the one making the calls. We decided which auth types to build first and which to leave for later. We noticed when our product’s needs made OAuth unavoidable, and we worked out how to shape the connector design so it could grow. We’re not claiming we got every call right, but those were the parts that needed us. The code arrived quickly. Deciding what to build, what to cut, and where to take the project next was still our job.

Will cheaper software create more or less programming work?

It comes down to a race. If demand for software grows faster than AI cuts the work each program needs, there will be more programming work. If not, less. No one has measured which is happening yet.

The past points one way. In the US, real business investment in software grew about 208 times from the first quarter of 1980 to the second quarter of 2026^[83], while its official price index fell about 58%^[84]. In the eras above, each step made software cheaper to build, so far more of it was wanted, and demand for developers kept growing. Economists know the pattern from coal. In The Coal Question, first published in 1865^[85], Jevons wrote: “It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth.”^[85]

Whether the pattern holds this time turns on one condition, which Federal Reserve Board economists Crane and Soto set out. Demand is elastic when a lower price brings a lot more buying, and inelastic when it brings little. In their words: “If the demand for coding services is inelastic this could lead to a fall in employment, as fewer coders are required to satiate demand.”^[86] If demand is elastic, cheaper coding could raise coder employment instead, which they judge “more likely in the long run than in the short run.” Box’s chief executive, Aaron Levie, makes the case for more work, as Tim O’Reilly relays it: “if agents make an engineer two to ten times more productive, all the software projects that were never economically viable suddenly become viable.”^[87]

Good-enough software also creates some work later on, when quickly built systems get hacked, lose data or can’t grow. But as AI improves, even that clean-up work may shrink, so it isn’t a safe foundation for a career.

Conclusion

For most of software’s history, the scarce thing was the ability to write it. AI has made that cheap, so the scarcity is moving to the work around the code: deciding what to build, judging whether it’s right and safe, and owning what it does once it runs. Demand is recovering there first, in senior and AI roles, and it’s the part of my team’s judgement that AI didn’t do.

Whether cheaper software means more programming work overall is still open. Which half of the job is getting cheaper is not.

Frequently asked questions

Why are software developers paid more than most workers?

Demand for people who can build software has grown faster than they could be trained, so employers have paid for a scarce skill. In May 2025 the median software developer in the US earned $135,980 a year, against $50,980 for all workers^[5].

Why is building software expensive when copying it is nearly free?

A copy of a program costs almost nothing, but writing it does not. Hal Varian put it as information being “costly to produce but cheap to reproduce,”^[1] so the scarce side of software has been the people who can build it.

Were there ever too many programmers before AI?

The evidence shows a shortage smaller than the booms made it look, rather than a glut. After the dot-com bust, new H-1B approvals at capped employers fell to 79,100 in fiscal 2002, well under the cap of 195,000^[22], and a 2013 study for the Economic Policy Institute found that US colleges provide “an ample supply of highly qualified science, technology, engineering and math (STEM) graduates.”^[25]

Are fewer people studying computer science?

Yes. In the US, enrollment in computer and information sciences fell at every level in fall 2025, and undergraduate enrollment fell again in spring 2026^[10]. Only 44% of 2015-16 bachelor’s graduates finished within 48 months of first enrolling^[41], so a fall in enrollment reaches the job market years later.

If you’re in this software field and want to share something, write to me.

Sources

87 references

1. Markets for Information Goods, University of California, Berkeley, 16 October 1998.
2. ITU releases 2015 ICT figures, ITU, 26 May 2015.
3. Facts and Figures 2025: Internet use, ITU, 2025.
4. Table 325.35. Degrees in computer and information sciences conferred by postsecondary institutions, by level of degree and sex of student: Academic years 1964-65 through 2021-22, National Center for Education Statistics, 2023.
5. Software Developers, Quality Assurance Analysts, and Testers: Occupational Outlook Handbook, U.S. Bureau of Labor Statistics, 27 August 2026.
6. Who is using AI to code? Global diffusion and impact of generative AI, arXiv, 20 November 2025.
7. Occupational projections and worker characteristics, U.S. Bureau of Labor Statistics, 27 August 2026.
8. Software Development Job Postings on Indeed in the United States (IHLIDXUSTPSOFTDEVE), FRED, Federal Reserve Bank of St. Louis, 7 October 2026.
9. AI Is Already Writing Almost One-third Of New Software Code, Complexity Science Hub, 22 January 2026.
10. Computer Science Enrollment Is Cooling, National Student Clearinghouse, 11 August 2026.
11. Software Engineering: Report on a conference sponsored by the NATO Science Committee, Garmisch, Germany, 7th to 11th October 1968, NATO Science Committee, 1969.
12. 4.1 Demand and Supply at Work in Labor Markets, Principles of Economics 3e, OpenStax, 14 December 2022.
13. Letting the "Computer Boys" Take Over: Technology and the Politics of Organizational Transformation, Cambridge University Press (International Review of Social History), 2003.
14. Intangible Machines, Cambridge University Press.
15. Occupations in Information Technology (ACS-35), U.S. Census Bureau, 2016.
16. Programming: from craft to scientific discipline (EWD 566), E.W. Dijkstra Archive, 25 August 2008.
17. Software Becomes a Product, Computer History Museum.
18. Oral History of Hugh Williams and Burton Grad, Computer History Museum, 9 December 1985.
19. Computer and Internet Use in the United States (P20-569), U.S. Census Bureau, 2013.
20. ITU estimates two billion people online by end 2010, ITU, 19 October 2010.
21. Education and Training Funded by the H-1B Visa Fee and the Demand for Information Technology and Other Professional Specialty Workers (RL31973), Congressional Research Service, 28 April 2005.
22. Education and Training Funded by the H-1B Visa Fee and the Demand for Information Technology and Other Professional Specialty Workers (RL31973), Congressional Research Service, 1 July 2003.
23. Monetary Policy and Sectoral Shocks: Did the FED react properly to the High-Tech Crisis?, NBER, 2003.
24. Information Technology: Assessment of the Department of Commerce's Report on Workforce Demand and Supply (GAO/HEHS-98-106), GAO, 20 March 1998.
25. New analysis finds no shortage of STEM workers in the United States, Edward J. Bloustein School of Planning & Public Policy, 24 April 2013.
26. Eight Years (And Counting) of Cloud Computing, AWS News Blog, Amazon Web Services, 14 March 2014.
27. Announcing Amazon Elastic Compute Cloud (Amazon EC2) - beta, Amazon Web Services, 24 August 2006.
28. Apple's App Store Marks Historic 50 Billionth Download, Apple Newsroom (Australia), 16 May 2013.
29. Undergraduate Degree Fields, National Center for Education Statistics, 2024.
30. Immigration Rules Appendix Skilled Occupations, GOV.UK (Home Office), 8 October 2026.
31. Employee earnings in the UK: 2025, Office for National Statistics, 23 October 2025.
32. The Software Industry and India's Economic Development, UNU-WIDER, 2001.
33. Survey on Computer Software and Information Technology Enabled Services Exports: 2024-25, Reserve Bank of India, 4 November 2025.
34. Survey on Computer Software & Information Technology Enabled Services Exports: 2013-14, Reserve Bank of India, 10 March 2015.
35. Africa and Latin America and the Caribbean account for less than 3% of global ICT services exports, UNCTAD, 29 September 2026.
36. Federal Reserve issues FOMC statement, Board of Governors of the Federal Reserve System, 16 March 2022.
37. Monetary Policy and Innovation, NBER, 2023.
38. The US Tech Hiring Freeze Continues, Indeed Hiring Lab, 30 July 2025.
39. Looking for the Ladder: Is AI Impacting Entry-Level Jobs?, Economic Innovation Group, 2026.
40. What is really happening to jobs? Separating AI hype from reality, Stanford Institute for Economic Policy Research (SIEPR), 2026.
41. Fast Facts: Time to degree, National Center for Education Statistics.
42. Stack Overflow Developer Survey 2017, Stack Overflow, 2017.
43. Cobweb Theory, Market Stability, And Price Expectations, Cambridge University Press (Journal of the History of Economic Thought), 2023.
44. Appendix D: Cobweb Model of the Engineering Labor Market, Understanding the Educational and Career Pathways of Engineers, The National Academies Press, 2018.
45. ICT education - a statistical overview, Eurostat, 2026.
46. Who is using AI to code? Global diffusion and impact of generative AI, arXiv, 20 November 2025.
47. There are 47.2 million developers in the world - Global developer population trends 2025, SlashData, 2 May 2025.
48. Octoverse: A new developer joins GitHub every second as AI leads TypeScript to #1, The GitHub Blog, 28 October 2025.
49. 34% of people shipping software using AI tools have no formal programming background, Stacker, for Zapier, 11 August 2026.
50. One year of Lovable: Welcome to the age of the builder, Lovable, 18 November 2025.
51. The Future is Actually Very Human, Replit, 11 March 2026.
52. AI, 2025 Stack Overflow Developer Survey, Stack Overflow, 2025.
53. Stack Overflow's 2025 Developer Survey Reveals Trust in AI at an All Time Low, Stack Overflow, 29 July 2025.
54. Asleep at the Keyboard? Assessing the Security of GitHub Copilot's Code Contributions, arXiv, 16 December 2021.
55. Do Users Write More Insecure Code with AI Assistants?, arXiv, 18 December 2023.
56. GenAI and Code Security: What You Need to Know, Veracode, 2025.
57. Methodology: How we discovered over 2k high-impact vulnerabilities in apps built with vibe coding platforms, Escape, 29 October 2025.
58. DORA 2025 State of AI-assisted Software Development Report, Google Research, 2025.
59. Gartner Forecasts Worldwide IT Spending to Grow 14.2% in 2026, Totaling $6.37 Trillion, Gartner, 27 July 2026.
60. Global ICT Spend to Reach $4 Trillion in 2026, Driven by Accelerated AI Platform Investments in Aerospace and Defense, IDC, 24 February 2026.
61. AI and Job Postings: From Destruction to Creation?, Indeed Hiring Lab, 8 July 2026.
62. Incorporating AI impacts in BLS employment projections: occupational case studies, U.S. Bureau of Labor Statistics, Monthly Labor Review, 10 February 2025.
63. Software Design Engineer in New Brunswick: Job prospects, Job Bank, Government of Canada, 10 December 2025.
64. Sector Skills Needs Assessment – Digital and technologies, GOV.UK (Skills England), 6 October 2026.
65. Introducing the Model Context Protocol, Anthropic, 25 November 2024.
66. Extend agents with Model Context Protocol (MCP) in Copilot Studio, Microsoft Copilot Blog, 19 March 2025.
67. Release v0.0.7 · openai/openai-agents-python, GitHub, 26 March 2025.
68. Google I/O 2025: Updates to Gemini 2.5 from Google DeepMind, Google, 20 May 2025.
69. Donating MCP to the Agentic AI Foundation, Anthropic, 9 December 2025.
70. The App Store stopped over $2.2 billion in fraudulent transactions in 2025, Apple Newsroom, 20 May 2026.
71. The App Store prevented more than $9 billion in fraudulent transactions, Apple Newsroom, 27 May 2025.
72. 2026 State of Mobile: AI Moves Mobile into Its Next Phase, Sensor Tower, 2026.
73. Q2 2026 Digital Market Index: Monetization Depth Outpaces Download Growth, Sensor Tower, 2026.
74. App Store added nearly as many new apps in H1 2026 as in all of 2025, 9to5Mac, 20 July 2026.
75. The App Store is booming again, and AI may be why, TechCrunch, 18 April 2026.
76. Is The Era of Mobile "AI Slop" Games Here?, Naavik, 19 April 2026.
77. The application explosion: key insights for employers, ISE, 28 April 2026.
78. ISE top 10 stats of 2025 you need to know, ISE, 8 December 2025.
79. Record graduate job applications, ISE, 17 October 2024.
80. Recruiter Productivity, 2026 Talent Trends Report, Ashby, 28 April 2026.
81. Ask HN: Who is hiring? (July 2026), Hacker News, through the Algolia search API.
82. Computer Programmers: Occupational Outlook Handbook, U.S. Bureau of Labor Statistics, 27 August 2026.
83. Real private fixed investment, chained quantity index: Nonresidential: Intellectual property products: Software (B985RA3Q086SBEA), FRED, Federal Reserve Bank of St. Louis, 30 September 2026.
84. Private fixed investment, chained price index: Nonresidential: Intellectual property products: Software (B985RG3Q086SBEA), FRED, Federal Reserve Bank of St. Louis, 30 September 2026.
85. The Coal Question, Econlib, 5 February 2018.
86. AI and Coder Employment: Compiling the Evidence, Federal Reserve Board, 20 March 2026.
87. The World Needs More Software Engineers, O'Reilly Media, 7 April 2026.