What Got Cheap, What Got Expensive: How AI Inverts the Cost Curve

The promise was simple: AI makes things cheaper. The reality is more interesting — it inverts the cost structure.

Things that used to be expensive because they were scarce (writing, code, design) are now nearly free because they are abundant. Things that used to be cheap because they were assumed (judgment, trust, taste, verification) are now expensive because they become the bottleneck. The substitution layer gets commoditized. The complement layer gets more valuable.

After two years of building with AI, here is what I notice.

What got cheaper

Cognitive labor at the pattern-matched layer. Writing a paragraph. Sketching a function. Drafting an email. Summarizing a meeting. Translating a passage. These used to require either skill or time. Now they require neither. The marginal cost of producing plausible content has collapsed toward zero.

Code scaffolding. A standard Next.js page. A CRUD endpoint. A CI workflow. These were real work for a junior developer. Now they are a prompt away. The 80% that fits a template is free.

Search and recall. The model of "remember the right URL or know the right query" is dying. Semantic retrieval is replacing it. You describe what you are looking for; the system finds it. Knowing-where-to-look used to be a real advantage; it is no longer scarce.

Translation. Bilingual content used to cost real money or real time. Now it costs seconds. The same is true for first-line customer support, basic data analysis, and first-draft image generation.

Tutoring at any level. Anything you can name, you can get a patient explanation of, in your language, at your level, on demand. The private tutor who used to cost $200/hour is now a sidebar.

What got more expensive

Here is the part that does not make the headlines.

Verification. When anyone can produce plausible output, the cost of verifying that output goes up. "Did the model actually understand, or did it pattern-match a confident-sounding answer?" is now a real question, and it takes real effort to answer. The person who knows the difference is suddenly valuable again.

Originality. The signal of "human-made" is becoming premium. When mass-produced content is free, it has zero value. The work of being distinctive — having a point of view, a taste, a voice, an actual experience — is what gets paid for.

Taste and curation. AI can generate a hundred logos. Knowing which one is right is still a human skill. In a world of abundance, that judgment becomes the bottleneck. The curator's job did not disappear; it became the whole job.

Trust. When deepfakes are trivial, when every email might be a phishing probe, when every paragraph could be AI-generated — the cost of believing goes up. Trust becomes a scarce resource, and the institutions that produce verified truth become more valuable, not less.

The last mile. The gap between "AI wrote me a draft" and "AI output is integrated into my real workflow, my customer's experience, my codebase" is enormous, and that gap is now where the cost lives. The 80% is free. The last 20% is most of the work.

Real-world experience. AI can write about cooking. It cannot taste your dish. As digital cognition gets cheaper, embodied skills — the things you can only learn by doing — become relatively more valuable.

Compute, energy, and GPUs. The inputs to AI are getting more expensive, not less. Per-token inference cost falls, but absolute spending on compute explodes. Someone is paying for it, and they are paying more every year.

Domain expertise that resists pattern-matching. The things AI cannot easily fake — the unwritten rules of a craft, the relationships in a community, the failures that only show up in production — these become more valuable, not less.

The asymmetry

What got cheaper is substitutable: AI does the thing a human used to do, slightly worse, much faster.

What got more expensive is complementary: AI generates the thing, but a human still has to verify, integrate, contextualize, and stand behind it.

The first category competes with humans on output.
The second category complements humans on judgment.

We are overinvesting in the first and underinvesting in the second.

The deeper pattern

Look at any technology that inverted a cost curve.

The printing press made books cheap and the rare manuscript expensive. The camera made portraits cheap and the painted portrait expensive. The container ship made global shipping cheap and the local craftsman expensive.

Every time, the substitution layer gets commoditized. The complement layer — the part that cannot be substituted — gets more valuable.

AI is doing the same thing, faster, to the cognitive layer.

A personal note

I built this blog with AI assistance. The boilerplate was free. The translations were free. The CI was free.

What was not free was deciding what to write about. What was not free was knowing which of the hundred design options actually fit the audience. What was not free was figuring out that the auto-deployment was failing because the CF Pages Dashboard does not follow repo restructures — a piece of knowledge no model had.

The pattern is consistent: AI handles the abundant, the substitutable, the pattern-matched. The expensive part is the judgment that turns output into something worth having.

What this means for builders

If you are planning a career, a company, or a project around AI, the question is not "what can AI do?" The question is:

What becomes the bottleneck when AI is in the loop?

Build for the bottleneck.