· Solveion · Perspectives · 5 min read
Cheaper to replace than to understand
A piece of electronics, almost working, one part missing. The repair cost more than a new one. That calculus reshaped how we treat physical objects, and it is now arriving for software.
One of us found a piece of electronics this week that was almost working. Everything functioned; one component was missing. The obvious instinct was to fix it, and then the arithmetic arrived: the repair would cost more than a replacement, by enough that spending an evening on it would have been an act of sentiment rather than economics. So it went in the bin, and a new one was ordered, and the whole transaction took ninety seconds and felt slightly wrong.
It is worth being precise about why it felt wrong, because the feeling is not nostalgia and the decision was not laziness. It was correct. That is exactly what makes it interesting.
How the calculus flipped
Two prices moved in opposite directions for forty years. Manufacturing and shipping a new object became extraordinarily cheap: global supply chains, containerised freight, scale that no single market could have supported alone. Skilled repair labour did not get cheaper, because an hour of a technician’s attention costs what an hour of anyone’s attention costs, and it cannot be manufactured at scale in another country.
Once the price of a new unit fell below the price of an hour of expertise, repair stopped being rational for a whole category of goods. Nobody decided this. It emerged, and then it changed behaviour, and then it changed what skills existed. There are far fewer people who can fix a television than there were thirty years ago, and not because televisions became harder to fix. It is because no one could earn a living doing it.
The consequences show up somewhere other than the receipt. The world generated 62 million tonnes of electronic waste in 2022, of which only 22 per cent was properly recycled, and the figure is heading toward 82 million tonnes by 2030. This is what an externality looks like: a decision that is individually correct, several billion times, producing a result nobody chose.
That is also why today matters. The deadline for EU member states to write the Right to Repair Directive into national law falls on 31 July 2026, requiring manufacturers to supply parts, tools and documentation on fair terms; seven US states have passed comparable laws. This is a deliberate attempt to legislate an economic calculus back into balance, which tells you how firmly it had tipped.
The same flip, arriving in software
Here is the part that concerns anyone building with AI, and it is the same shape.
Generating code has become nearly free. Understanding code has not. Reading an unfamiliar system, forming an accurate model of it, and finding the one wrong assumption three layers down still costs what an hour of expert attention has always cost. Meanwhile, “regenerate this component from scratch” has fallen to the price of a prompt.
So the same threshold is being crossed. Faced with a service that behaves strangely, the rational move is increasingly not to debug it but to rebuild it, because rebuilding is cheaper than comprehension. We have watched teams do this, and much of the time they are right, in precisely the way that throwing away a working appliance for want of one part is right.
The risk is the same too. When repair stops being economic, repair capability stops being cultivated. If a generation of engineers learns that the answer to a confusing system is a fresh generation of it, the ability to hold a large system in your head and reason about it accurately becomes as rare as television repair, and for the same unremarkable reason.
Appliances and infrastructure
The useful distinction is not between disposable and durable software. It is between the systems that are appliances to your business and the systems that are your business.
An appliance is a component with a clean boundary and a testable contract. If it misbehaves, regenerating it is fine. It was never carrying institutional knowledge, and the sooner you stop paying an expert to nurse it, the better. Most integration glue is like this. So is most tooling.
Infrastructure is different. It is where your actual business logic lives: the pricing rules with fifteen years of exceptions in them, the eligibility check that encodes a regulatory position, the reconciliation process nobody documented because everyone knew it. These systems look like code and are in fact institutional memory. Regenerating one does not restore it; it produces something that passes the tests you thought to write and quietly discards the reasoning you had forgotten was in there. The failure is not immediate, and it is not recoverable by regenerating again.
The organizations that will handle this well are not the ones that resist cheap regeneration. It is genuinely cheap, and refusing it on principle is the same sentimental error as spending an evening soldering a component into a device worth less than the solder. They are the ones that can say which of their systems are appliances and which are infrastructure, and that keep, deliberately and at real cost, the ability to understand the second category.
Globalization taught us to treat objects as disposable by making replacement cheaper than expertise. AI is making the identical offer about the things we build. It is a good offer, most of the time. The discipline is knowing which things you are not willing to stop understanding.