Software Is Becoming Cheap. What Remains Scarce?

The claim that AI makes software cheap is now common enough to be uninteresting. What follows from it is not. The same arithmetic reaches engineering, planning, analysis and administration — the general form of the argument is that technology is becoming easier to build and competitive advantage is not — but software is where it became visible first, so it is the clearest place to work through it.

If the cost of producing a given piece of working software falls sharply, and falls for everyone at roughly the same time, then whatever advantage rested on that cost falls with it. Not all advantage — only the part that was doing the work of being expensive. The interesting question is which part that was, and what is still standing afterwards.

Six candidates come up repeatedly. What matters is not the list, which is unremarkable, but the test attached to each one. Every company believes it has distribution, data and stickiness. Very few can pass a question designed to disprove it.

1. Distribution

When products become cheap to build, the constraint moves to attention. A buyer’s willingness to evaluate anything at all is finite and does not scale with the number of available products. Owning access to that buyer becomes more valuable than owning the code they end up using.

This is why the same product can be a good business inside one company and a bad one inside another. The asset is not the software. It is the standing permission to sell to someone.

2. Proprietary data

Model weights are rentable. Data produced by operating a business is not always. The distinction that matters is not volume, and it is certainly not the word “proprietary” in a management presentation. It is reconstructability.

Data that describes what happened is usually purchasable somewhere. Data that describes what was corrected, what failed, what the exception was and what the outcome turned out to be tends not to exist outside the company that lived through it.

3. Workflow ownership

There is a difference between software a company uses and software a company runs through. The first is a tool sitting beside the work. The second is the shape of the work itself.

Replacing a tool costs a purchase decision. Replacing a workflow costs retraining, migration, a period of degraded output, and the political capital of whoever proposed it. That second bill is why unremarkable incumbents survive competitors with better products, and why “we could rebuild that” so rarely becomes “they will switch”.

4. Trust and accountability

Where a decision carries regulatory, financial or physical consequence, buyers are not purchasing capability alone. They are purchasing someone who is answerable when it goes wrong. Automation lowers the cost of producing the analysis. It does not lower the demand for a name underneath the conclusion.

This is a real and durable asset, and it is also the one most often claimed without basis. Being trusted by customers is not the same as carrying liability for them.

5. Physical integration

Software that controls, certifies or ships inside physical equipment inherits that equipment’s constraints: approval cycles, service networks, spare parts, liability, installed base, and the simple fact that someone has to drive there.

These constraints are frequently described as legacy burdens. In a world of cheap software they are also the reason a new entrant with a superior product cannot arrive next quarter. Atoms remain harder than bits, and difficulty is what a moat is made of.

6. Judgment

The least discussed consequence of cheap production is that it raises the cost of building the wrong thing, because far more wrong things get built. When execution was expensive, the plan was scrutinised. When execution is cheap, the plan is often skipped.

The scarce input becomes the decision itself: what to own, what to rent, which programme to stop, which capability will still matter in five years. This is not a technology skill, and unlike engineering capacity it is not getting cheaper.

Where this leads

None of the six is a technology moat. That is the finding, not an oversight. As the technology component of an advantage depreciates faster, the non-technology components carry more of the weight — which means the parts of a business that are hardest to examine are becoming the parts that matter most.

For an operator, that changes what deserves capital. For an investor, it changes what a technology diligence has to cover. The reproducibility question — how much of the customer-visible product could be rebuilt, and how quickly — is worked through separately in the 80% rebuild test.