The Economics of Intelligence

Efficiency died as a competitive advantage the moment it became an off-the-shelf item. You can buy world-class logistics, payment infrastructure that processes billions, support that answers in seconds, models that write, translate, and reason — all of it by monthly subscription, with a credit card, no meeting required. What for fifty years was the battlefield of companies — doing the same thing faster, cheaper, with fewer errors — collapsed into commodity. And a commodity doesn't defend margin. It only transfers value to whoever sits a layer above.

The question that separates the companies that will matter from the ones that will evaporate over the next ten years isn't "how do I do this better." It's: what layer do I live in, and what do I accumulate that no one can rip away from me. This is the economics of intelligence. Not in the sense of "using AI" — almost everyone will use it, the way almost everyone uses electricity. In the sense that the scarce asset, the thing that will concentrate profit, has stopped being execution and become the capacity to read context and decide well, repeatedly, at a pace the competition can't match.

The price of a decision plummeted, and that changes everything

There's a historical asymmetry that organized the entire modern economy. Executing was cheap; deciding was expensive. An assembly line you scaled with capital. But knowing what to assemble, for whom, at what price, with what message, in what channel — that required senior people, scarce, slow, and expensive. The decision was the bottleneck. That's why strategy consulting charged fortunes, why a good analyst was worth ten operators, why the C-suite concentrated power: because judging well under uncertainty was rare.

That bottleneck is dissolving. Not because the machine decides better than the best human — it doesn't, yet, and it may take a while. But because it decides reasonably well, millions of times, at a marginal cost near zero. And in economics, what reorganizes sectors isn't the peak of quality. It's marginal cost. When reading a hundred-page contract and flagging the three dangerous clauses costs cents instead of eight hundred dollars of a lawyer's hour, it isn't the expensive lawyer who disappears — it's the entire structure around them that reorganizes. The scarce becomes abundant, and value slips toward what is still scarce.

Think about what happened to translation, to transcription, to the first draft of anything, to résumé screening, to first-tier support, to exploratory data analysis. All of it was low- and mid-level decision work, done by people, and all of it plummeted in price in eighteen months. The ones who made money weren't the ones doing those tasks. It was whoever was positioned to capture the surplus that the price drop released — provided they were in the right layer.

And here is where the trap lives. Most companies will use the falling price of decision to do exactly what they already did, only cheaper. They'll lay people off, cut costs, squeeze margin for one more quarter. It's the efficiency response applied to a world that no longer rewards efficiency. It works once, as a cost saving. Then the competitor does the same thing, the gain becomes the new floor, and everyone is back to square one — only with fewer people and the same lack of advantage. Efficiency distributed to everyone is an advantage to no one. It's the new cost of operating.

The interface was where the money lived. Not anymore

During the era of software as a service, value concentrated in the interface. The thesis was simple and correct for its time: the product is where the user works. Whoever controls the screen controls the relationship, the data, the lock-in, the price. Build an interface so good the user doesn't want to leave, and capture rent from their habit. That's how Salesforce was built, and Notion, and Figma, and Stripe itself in its developer-UX layer. The battle was over the best experience on top of a relatively standardized set of functions.

The interface is becoming the cheapest part of the stack. Today you describe the application you want and it appears — with database, authentication, deploy, everything. The screen stopped being the scarce asset because building a screen became trivial. What that means, concretely, is that the value-capture point is rising up the stack, from "where the user clicks" to "what the system decides before the user needs to click." The winning interface of the coming years is the one that disappears — because the system already solved it, already decided, already acted, and just shows the result for confirmation.

Look at what OpenAI and Anthropic are actually selling. It isn't an interface. ChatGPT has a text box; the product isn't the box. The product is the decision layer behind it — the capacity to read a vague request, infer intent, mobilize context, and return something useful. The interface is disposable and they know it; that's why they open up the API and let thousands of companies build their own screens on top. The value they keep for themselves isn't in the screen. It's in the model, in what it learned, in the infrastructure to serve it cheaply, and increasingly in the context they accumulate about each user and each task. The interface is the bait; the capture is underneath.

This reorganizes the rules of the game for everyone who builds product. If your advantage is the interface, you're building a sandcastle in the tidal zone. Someone will clone the screen over a weekend with the tools that are arriving. The question that matters has become: what does your product know and decide that a clone of the screen wouldn't know? If the answer is "nothing, the screen is the product," you're in a layer that is collapsing in price. If the answer is "it accumulates the client's history of decisions and gets better with each one," you're in a layer that rises in value over time.

Buying a tool versus building a system

There's a difference that seems subtle and is, in fact, the central fracture of this economy: the difference between buying a tool and building a system. Buying a tool is renting capacity. Building a system is accumulating capacity. And only one of the two compounds.

When you buy a tool — any SaaS, any API, any model by subscription — you buy the same power your competitor buys, at the same price, in the same week. The tool is, by definition, distributed. The vendor exists to sell it to as many people as possible; their economies of scale depend on it. So what you get is parity, not advantage. You raise the floor of your operation, and that's all. The instant it becomes a differentiator, the vendor sells it to everyone and the differentiator evaporates. You end up hostage to an input that, precisely because it's good and cheap, will never set you apart.

Building a system is a different nature of thing. A system is what you assemble around the tool that captures something you can't buy: the context of your operation, the decisions you've already made and their results, your client's specific flows, the tacit knowledge of your company codified into something that improves on its own. The tool is the same language model everyone has. The system is what you bolted onto it — the proprietary data, the memory of interactions, the rules distilled from ten thousand cases, the particular way you decide. That isn't for sale because it doesn't exist outside of you. And, above all, it compounds: each decision feeds the next, each case improves the system, the moat deepens while you sleep.

I live this in practice building what I'm building. The language model is a commodity — anyone can buy access to the same one. What isn't a commodity is the system that learns from each interaction and stores that learning in a way that doesn't leak into the product you hand to the client. The real moat isn't the code you ship; code is read, copied, rewritten. The moat is the learned state that stays on your side of the border and that the client never receives. You deliver the capacity; you retain what made that capacity good. Whoever understands this distinction builds an asset that appreciates. Whoever doesn't ends up eternally renting someone else's engine, calling it innovation, and wondering why the margin never rises.

Most companies will choose to buy the tool. It's faster, cheaper in the short term, doesn't require building internal competence, shows a result next quarter. And that's exactly why it will be a mass trap: because the comfortable road leads everyone to the same commoditized place, and by the time you get there, there's no room left to set yourself apart. The choice between buying and building isn't a technical choice. It's a choice about where you want to be when the dust settles — in the layer that captures value or in the layer that merely pays for it.

Architecture is destiny

Everything has an invisible architecture, and it's the architecture, not the surface, that determines what a system can become. In a company, the architecture is the answer to a question almost no one asks explicitly: what layer of the value chain do we live in, and what does that position let us accumulate over time. You can have the best product in your layer and still be doomed, if the entire layer is being commoditized from below. And you can have a mediocre product in a structurally privileged layer and prosper for a decade, capturing value you didn't build.

Cloudflare understood this with brutal clarity. They didn't sell the best server, nor the best interface, nor the best isolated feature. They positioned themselves in a layer — between the user and the entire internet — where each new client improves the network for everyone else: more traffic means more signal about attacks, more intelligence about threats, more capacity to mitigate the next attack before it happens. The architecture is the product. The position in the layer is what compounds. Competing with that isn't building a better product; it's having to rebuild the entire position, and the position was built with years of accumulated traffic you have no way to buy.

Stripe made a similar layer move with payments. The surface was a clean API for developers — and for years people thought that was the product, the famous "developer experience." It wasn't. The API was the door. The system was everything that accumulated behind it: the knowledge about fraud distilled from trillions in transactions, the relationships with banks and regulators across dozens of countries, the risk infrastructure that gets better with every payment processed. A competitor clones the API in a quarter. They don't clone twelve years of fraud signal. The chosen architecture — being the intelligence layer over money, not just the pipe it flows through — is what made the company what it is.

The lesson for anyone building now is uncomfortable, because it forces you to think before you execute, and startup culture rewards the opposite. The architectural question — what layer do I live in, what do I accumulate, what improves on its own with use — has to come before the execution question. Because you can execute a wrong-layer strategy to perfection and build, with enormous effort, a business that is doomed by structure. Speed in the wrong direction is just a more expensive way to arrive at the wrong place. Architecture is destiny, and destiny is decided at the start, when it still seems there's plenty of time to think about it later. There isn't.

What rises in value when deciding gets cheap

If deciding well gets cheap and abundant, what becomes scarce? Because value always migrates toward the scarce — that's perhaps the one economic law that never fails. And it's worth looking precisely at where it's going, because it isn't obvious and most people will bet on the wrong place.

First: proprietary context. When everyone has access to the same reasoning capacity, the difference between a generic decision and an excellent one is the context you feed into it. The data only you have — your client's history, the results of your past decisions, the tacit knowledge of your operation codified in a usable form. A language model with your context is worth orders of magnitude more than the same model without it. Context becomes the scarce input, and context isn't bought — it's accumulated, and only those who positioned themselves to accumulate early on do. That's why long-term memory, the capacity of a system to remember and build on what it has already lived, stopped being a technical detail and became the central asset.

Second: taste and judgment about what matters. When the machine generates a thousand reasonable options at the cost of cents, choosing the right one among the thousand becomes the highest-leverage work there is. The bottleneck shifts from production to curation, from "doing" to "knowing what's worth doing and what to discard without remorse." This is profoundly human and profoundly scarce — not because the machine can't judge, but because judging well requires a model of the world, of the client, of what truly matters, that still lives mostly in people with real scars. The person who knows which of the thousand options to ship is worth, again, orders of magnitude more than the one who produced the options one by one.

Third: trust and accountability. When deciding gets cheap and abundant, wrong decisions also get cheap and abundant — and someone has to answer for them. The scarce thing becomes the capacity to guarantee that the automated decision is right, is safe, is auditable, won't blow up in your face or your client's. Governance stops being a compliance cost and becomes a competitive advantage, because in a world of cheap automatic decision-making, the company you can trust is the rare one. Whoever builds the trust layer — auditing, traceability, the capacity to explain why the system decided what it decided — captures value that the company which only automates does not.

Fourth, and deepest: the capacity to ask the right questions. When answering gets cheap, the lever migrates entirely to the quality of the question. A system that answers the wrong question brilliantly is worse than useless, because it gives the false sense of progress while you accelerate toward the cliff. Asking the right question — about what to build, what to measure, what matters in this decade — is the work that doesn't commoditize, because it requires a vision of the world, and vision is the last scarce input. It's also the hardest to outsource, precisely because there's no way to.

The reorganization: fewer people, more leverage, a deeper moat

The economics of intelligence reorganizes the shape of the company, not just what it uses. The company of the previous cycle was a machine for coordinating human labor: layers of management existing so that many people could do many things in coordination. A good part of that structure was coordination cost — people managing people, people translating decision into execution, people verifying the work of people. When execution and a good part of low-level decision become cheap and abundant, that whole structure becomes too fat for what it delivers.

What emerges is a different shape: small teams with enormous leverage. Not because "AI replaces people" in the lazy sense of the headline, but because the unit of production changes. One person with a well-built system does what once required a department — not because they work more, but because leverage per person rose an order of magnitude. The ten-person company that competes with the thousand-person one isn't fiction; it's the direct result of decision and execution becoming cheap for whoever built the right system. The advantage stopped being the quantity of people coordinated and became the quality of the system that leverages each person.

This has a counterintuitive consequence: value per person in a well-architected company will rise, not fall. When each person operates a system that does the work of many, each person carries far more leverage, and the right talent — the kind that knows how to build and steer those systems, ask the right questions, judge among the thousand options — becomes more valuable, not less. The war for talent doesn't end; it concentrates. It's the middle of the labor market, the generic mid-level decision work, that evaporates — compressed between the machine below and scarce judgment above.

And the moat behaves differently. In the efficiency cycle, the moat was static: you built a cost or scale advantage and defended it. Now the moat is dynamic — it deepens with use, or it doesn't exist. Either your system gets better with each client, each decision, each interaction, accumulating context no one can replicate; or it's just a tool anyone can rent, and you have no moat at all, just a race in which you're always tied. The question that defines whether you have a defensible business has become: what in my system improves on its own while I sleep? If the answer is nothing, you don't have a moat. You have a monthly subscription that your competitor also has.

The window is open now, and it closes

There's a moment, in every economic reorganization, when the layer positions are still up for grabs and you can position yourself. Then that moment closes, the layers consolidate, and whoever stayed out pays rent forever to whoever got in. It happened with the internet — there was a window in which you could be the search layer, the social layer, the payment layer, and whoever got in at the right time collects a toll to this day from everyone who arrived later. It happened with the cloud. It's happening now with intelligence, and the window is narrower than it looks because the speed of consolidation is greater.

Most will spend that window wrong. They'll use it to cut costs, to do the same as always more cheaply, to buy the tool of the moment and glue it on top of the old process. They'll confuse adopting AI with positioning themselves in the economics of intelligence, and those are opposite things: adopting the tool puts you in the same commoditized boat as everyone; positioning yourself requires the hard layer question, requires building instead of renting, requires accumulating instead of consuming. The comfort of buying the ready-made solution is exactly what will leave most behind, because the easy road leads to the crowded place.

What I would do — what I am doing — is simple to say and hard to execute, which is how everything that matters tends to be. Decide what layer to live in before deciding what to build. Build a system instead of renting a tool whenever what's at stake is your differentiator, and rent a tool without guilt for everything that's commodity anyway. Accumulate proprietary context from day one, treating the data of your decisions as the asset it is and not as operational exhaust. Retain the learned state on your side of the border, delivering capacity to the client without delivering what made that capacity good. And keep human, scarce, and well paid the work that doesn't commoditize: the right question, the judgment about what matters, the accountability for the decision.

The economics of intelligence doesn't reward whoever uses AI. Almost everyone will use it, the way almost everyone uses electric light. It rewards whoever understands that deciding well got cheap, and that therefore advantage migrated to what cheap deciding doesn't solve: the layer you live in, the context only you have, the system that improves on its own, the question no one else is asking. Efficiency was the advantage of fifty years and became the cost of getting a seat at the table. Architecture is the advantage of the next fifty. The difference between whoever will capture the value of this turn and whoever will pay rent for it isn't in the tool they bought. It's in the question they had the courage to ask before they started to build — and in the discipline to build, instead of just renting someone else's engine and calling that the future.