When Expertise Gets Cheap

Will AI end capitalism? Almost the wrong question. Capitalism does not require a priesthood of specialists. It requires property, contracts, incentives, and something scarce enough to price. The better question is narrower and meaner: what happens when the cost of acquiring economically useful knowledge approaches zero?

For most of the industrial and professional centuries, that cost was high. A person who spent twenty years becoming competent in five disciplines was rare and ruinously expensive. Society optimized around the cheaper unit: the specialist. Economists studied economics. Lawyers studied law. Engineers studied engineering. Corporations assembled these people into teams because nobody could hold the whole problem in one skull. Expertise was the scarce resource. The org chart was a cost function wearing a necktie.

The production function flipped

Ask a model to explain modern monetary theory. Then the semiconductor supply chain. Then Taiwan’s position in both. Then the second-order effects on rates, inventories, and freight. The incremental cost of the next domain collapses toward the price of tokens and a few minutes of attention. You have not become a central banker or a foundry engineer. You have rented a slice of their reasoning.

That is a different relationship with knowledge than the one universities sold. They taught content, yes. They also certified who was allowed to claim competence. The degree, the bar card, the MBA, the lab affiliation — these were trust technologies. They told a hiring market that a person had survived a slow filter. AI attacks the first function immediately. It leaves a live question about the second: what happens when demonstrated capability becomes cheaper than credentialed capability?

Capitalism never required the credential. It required a way to allocate opportunity under uncertainty. Specialization and diplomas were one allocation machine. If models make the underlying capability cheap to instantiate, that machine loses pricing power even if the letterhead remains.

The inversion is easy to miss because it does not feel like intelligence arriving in your skull. It feels like access. You do not need to own a lawyer’s knowledge to obtain legal reasoning. You do not need a decade of statistics before you can run a regression. You temporarily instantiate expertise. The specialist used to possess the scarce stock. Now the specialist’s stock can be leased by the query.

Knowledge, execution, orchestration

Think in layers, not vibes.

Layer 1 is knowledge — what Bayesian inference is. This is nearly free.

Layer 2 is execution — use it on this dataset; write the function; draft the memo. This is getting cheap in the domains with tight feedback. Cursor-class coding agents already compress months of software delivery into days, which is why enterprises are yanking production back in-house. Programming got there first because the compiler and the test suite answer back, a verification loop other kinds of knowledge work still lack.

Layer 3 is orchestration — take inference, market microstructure, behavioral bias, chip geopolitics, and the policy reaction function, and build an account of why this tape is doing that. The human here is not necessarily the one who knows more than the model. The human chooses which domains matter, which links are real, which questions deserve oxygen, and what the resulting object is for. The model is the workforce underneath.

That layer is why the old résumé starts to invert. The prior economy asked what are you? Doctor. Engineer. Counsel. The emerging one asks what can you connect? A person who is merely pretty good at economics, machine learning, history, markets, and code used to look unfocused — a departmental orphan. If depth can be rented on demand, the bizarre combination is the asset. AI can make narrow specialization less scarce while making unusual human joints more valuable. Not because the person mastered everything. Because they can direct many things into an object that did not exist.

You can watch firms act this out without admitting the theory. Microsoft’s 4,800-job cut was not a parable about robots at desks. It was opex harvested so capex could keep compounding — labor treated as the flexible input once models and racks became the claimed engine. Headcount is what you liquidate when you believe execution is about to get cheaper.

Scarcity does not vanish. It moves.

If that were the whole story, the piece would be a TED talk. It is not. AI does not abolish scarcity. It changes which scarcities bind.

Compute still costs. Energy still costs. Packaging and silicon still bottleneck; TSMC’s capex is a reminder that tokens sit on wafers, not on wishes. Distribution, reputation, trusted brands, legal liability, and the right to transact remain expensive. Attention remains finite. Relationships remain slow. Ownership still matters. Capitalism mutates around whatever is still hard to copy.

So the sequence is directional, not complete: from physical production, to machines, to information, to distribution, to judgment. Not all the way, not evenly, not this quarter. Directionally, the competitive battlefield slides from who knows toward who synthesizes — and toward who owns the remaining bottlenecks underneath the synthesis.

The historical inversion is almost tender. For most of modern professional life, becoming capable meant narrowing yourself: pick a field, survive the filter, become useful inside a box. The emerging production function allows a different bet. Learn broadly. Stay curious. Find the joints. Rent depth from machines. Build the thing nobody’s department was chartered to notice.

The strange person — the one whose interests never fit a single door plaque — is no longer, automatically, poorly specialized. They may be the scarce unit the new cost curve actually wants: not a walking encyclopedia, but the one who can say, wait. These four things are the same problem.