The headline reads: Big Tech is laying off tens of thousands of developers, AI is taking over programming, and the profession is dying out. The headline has got the situation completely wrong. What is happening right now is not the end of software development, but the beginning of its greatest expansion. And the reason for this is a 160-year-old economic observation that, of all people, the CEO of Microsoft has brought back into the spotlight.
Anyone who draws the wrong lesson from the redundancy figures is missing the very window of opportunity in which custom software is becoming affordable for small and medium-sized enterprises for the first time. To understand why, we need to take a brief look back at the year 1865.
What the Jevons Paradox has to do with your software
In 1865, William Stanley Jevons made an observation in *The Coal Question* that defies intuition. James Watt’s steam engine consumed significantly less coal per unit of work than its predecessors. One would have expected Britain to consume less coal as a result. The opposite happened. Because the engine became more economical, it suddenly became viable for use in factories, mines and trains, where it had previously been too expensive. Coal consumption rose dramatically.
This is the crux of the paradox: as the use of a resource becomes more efficient, its overall consumption rises. The low price opens up new applications, and these new applications consume more than the efficiency gains have ever saved.
It was precisely this pattern that Satya Nadella applied to AI in January 2025, when the rise of DeepSeek was causing unease in the markets. “Jevons’ paradox strikes again,” he wrote. “As AI becomes more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.” In other words: the more efficient and accessible AI becomes, the more its use increases, until it becomes a commodity of which one can never get enough.
Nadella was referring to computing power. The same mechanism applies to software development, just one level higher. In this context, what AI produces – code – becomes a commodity.
The waves of redundancies tell a different story to what you might think
The figures are real; there’s no question of that. In the US alone, over 120,000 tech workers lost their jobs in 2025; worldwide, the figure stood at over 200,000, with year-end estimates pointing towards 235,000. The job cuts continued in 2026. And the narrative surrounding this is shifting noticeably towards AI: in May 2026, the outplacement firm Challenger, Gray & Christmas attributed around 40 per cent of job cuts to AI; in January of the same year, the figure was just 7 per cent.
However, this narrative does not hold up as well as it sounds. A large proportion of these redundancies simply corrects the excesses of the pandemic years, when the same corporations were hiring in bulk at zero-interest rates. Added to this are rising interest rates and a reallocation of capital: savings made in one area flow into the next AIdata centre investment. Strikingly often, the same companies report record turnover and job cuts in the very same quarter. This is portfolio restructuring whilst coffers are full, not a contraction in demand for software.
In this complex situation, ‘AI’ is also the most convenient label. It sounds better on the stock market than ‘we got our calculations wrong in 2021’. Anyone who takes the headline at face value is confusing the reorganisation of a handful of corporations with the state of an entire professional sector.
A look at the official forecasts supports this. The US Bureau of Labour Statistics expects employment for software developers to rise by 15 to 16 per cent between 2024 and 2034, “much faster than the average for all occupations”, and around 129,000 job vacancies per year in the broader developer category.
The bottleneck has always been the price
Ask any medium-sized business about their wish list for software, and you’ll get it straight away: the ERP that finally integrates seamlessly with the online shop. The quotation calculation, which currently resides in a sprawling Excel spreadsheet. The customer portal that the sales team has been asking for for years. The backlogs were never empty. They were frozen.
Frozen by the price. With traditional custom software, the development budget is the smaller item. It is the total costs over the system’s lifetime that tip the balance. Studies on software economics regularly put the maintenance share of life-cycle costs at 60 to 80 per cent. A project with a development budget of 200,000 euros can therefore quickly run to between half a million and a full million over its lifetime. Faced with these figures, small and medium-sized enterprises draw a firm line: they only develop what tangibly sets their business apart. Everything else is left undone, purchased from third parties or done manually.
This limit is pure mathematics. And mathematics changes when any one of the factors changes. This is precisely where AI comes in: the cost of implementation, the only factor that has ever shifted this limit.
Custom software is becoming affordable, and that is precisely the boom
When AI reduces the cost of producing code, the threshold at which a project becomes economically viable shifts downwards. Take the customer portal from the wish list mentioned earlier. With traditional development costs, it ended up costing well over a quarter of a million over its lifetime and remained there, year after year, in every budget round. As development costs fall, this very portal slips below the threshold at which senior management gives the go-ahead. This makes custom software viable for projects that were previously never cost-effective.
This is the Jevons paradox in its business context, and it hits small and medium-sized enterprises harder than any Silicon Valley corporation, because that is where pent-up demand is greatest.
This also shifts the old fundamental decision. Whether to rent standard SaaS or build your own software was long a question of budget: building was the expensive premium option, whilst buying was the pragmatic default. As the price of building falls, the balance tips in favour of in-house solutions in more cases. Anyone wishing to clearly map out the strategic framework behind this will find it in our decision matrix for SaaS versus bespoke software; the technical flip side – that is, when building your own code really pays off – is analysed in Build vs. Buy from an engineering perspective.
A glance at the industry leaders shows that this is no theoretical scenario. At Microsoft, according to Satya Nadella, AI writes up to 30 per cent of the code; Google CEO Sundar Pichai stated that by the end of 2024, AI would already account for over a quarter of new code. Nevertheless, neither company has done away with its development organisations. They are building more ambitiously, not more economically. What is happening at the top offers a preview of what is coming to medium-sized businesses: greater capacity to produce code leads to more software and more development.
The catch: cheaper code is not automatically better code
For the curve to point upwards, one condition must be met, one that is often overlooked amidst all the optimism about efficiency: the software must be reliable. This is precisely where the surge in demand becomes challenging.
The productivity gains are, in fact, less clear-cut than the 30 per cent figures suggest. In the summer of 2025, a controlled study by the METR research institute had sixteen experienced open-source developers work on 246 real-world tasks, half using AI tools and half without. The result was uncomfortable: with AI, the developers took 19 per cent longer on average. Even more revealing is the perception gap. Afterwards, the same developers estimated that AI had made them 20 per cent faster. They were slower but thought they were faster.
It’s important to put this into context: the tools tested were from spring 2025, and METR itself describes the result as a snapshot and subsequently revised the study design. Nevertheless, the point remains: there is a disconnect between perceived speed and actual speed when using AI tools.
This is a warning against the naïve approach to AI development. Code adopted without scrutiny looks like finished work, but in reality is often technical debt brought forward. Where this line is drawn, why the lowest conceivable threshold is what “works”, and how quickly unread code becomes a security risk is described in detail in Vibe Coding vs. Good Code in detail.
The actual effect of AI is therefore a shift in value. Typing code becomes cheap. Deciding what should be built in the first place, and verifying whether what has been built is sound, are becoming more expensive and more important. An AI that builds the wrong feature three times as fast hasn’t helped anyone. On the other hand, those who select the right projects and ensure their quality are in greater demand than ever.
What this means for the future of AI software development
For small and medium-sized enterprises, this is the real story behind the headlines about redundancies. It’s not that ‘software is becoming redundant’, but rather that ‘software is becoming affordable’. Projects that were previously on the wish list but considered too expensive are now within reach. And the window of opportunity in which early access to them provides a real competitive edge is now open.
What is changing is the role played by a service provider or an in-house team. Value now lies in the ability to select the right projects from a growing backlog, to deploy AI in a disciplined manner during implementation, and to ensure quality so that the software will still be viable in five years’ time. Efficient production without this judgement merely produces legacy systems more quickly.
For SMEs, now is the time to reassess their own frozen backlog: which projects will the reduced production costs make profitable first? This question is both an architectural and an economic one, and it is precisely at this intersection that we come in with enterprise and bespoke software.
Jevons was right about coal. He’s right about code, too.
Sources (verified, to be checked again before publication):
- Jevons, W. S. (1865): The Coal Question, on the efficiency–consumption paradox.
- Nadella, S. (27 January 2025): public post on the Jevons paradox in the context of DeepSeek; covered by Fortune, amongst others (https://fortune.com/2025/01/27/microsoft-ceo-satya-nadella-deepseek-optimism-jevons-paradox/).
- Nadella (Microsoft, up to 30% AI code, LlamaCon 29 April 2025) and Pichai (Google, over a quarter of new code, Alphabet Q3 2024 earnings call): including Tom’s Hardware (https://www.tomshardware.com/tech-industry/artificial-intelligence/microsofts-ceo-reveals-that-ai-writes-up-to-30-percent-of-its-code-some-projects-may-have-all-of-its-code-written-by-ai).
- Tech redundancies 2025/2026, US vs. global, and AI attribution (Challenger, Gray & Christmas, May 2026 report): Layoffs Tracker (https://layoffs.fyi/).
- US Bureau of Labour Statistics, Occupational Outlook Handbook, Software Developers (Growth 2024–2034): https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm.
- METR (10 July 2025): Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/.
- Maintenance accounts for 60–80 per cent of lifecycle costs: a common figure in software economics, see e.g. R. L. Glass, Facts and Fallacies of Software Engineering (2002).
- Security/quality status of AI-generated code: see evidence in the article Vibe Coding vs. Good Code (Veracode 2025, amongst others).