Published October 2, 2026
If the best AI model became ten times cheaper and twice as capable tomorrow morning, would your moat shrink — or would your economics improve?
1
The question nobody asks out loud
Every company has an AI strategy. Fewer have an AI exposure.
The difference is direction. A strategy describes what you are doing with the technology — models adopted, features shipped, pilots running, people hired. An exposure describes what the technology does to you when it improves without your involvement, which it will, roughly every quarter, for the rest of your career. Only one of the two compounds.
Here is a way to find out, and it takes about four seconds. Suppose that tomorrow morning the best available model becomes ten times cheaper and twice as capable. No warning, no transition period, no time to prepare. What has happened to your business by lunchtime?
Only two answers are worth giving. The first: things get harder. The gap between what you sell and what your supplier sells has narrowed. A customer who was paying for your judgement discovers that most of it now arrives in the box. Your pricing comes under quiet pressure from three competitors who received the same upgrade on the same morning and did nothing to earn it. The second answer: things get easier. Your cost of production falls, the range of things you can afford to make widens, and one or two ideas you shelved as too expensive turn into arithmetic you can defend in front of a board.
Same event. Opposite consequence.
What is odd is not that companies sit in different positions. That is inevitable, and serious businesses exist on both sides of the line. What is odd is how rarely anyone says out loud which one they are in.
2
Two exposures, drawn plainly
Strip the question down to its two positions.
You sell the intelligence. Your product sits between a model and a customer who wants an outcome from it. The layer may be excellent — an interface, a workflow, an evaluation harness, a set of prompts refined over two years of contact with real users. But the value the customer is buying originates below you. Every release from your supplier does three things at once: it narrows what you add, it improves your competitors without their lifting a finger, and it teaches the customer that the capability is closer to a commodity than they had assumed. Improvements in the model arrive as pressure.
You use the intelligence. Intelligence is an input, in the same family as bandwidth, electricity and storage. You buy it, convert it into something a customer values for a different reason, and sell that. A cheaper input lowers your cost of goods. A better input widens the range of what you can make. Improvements arrive as slack.
Two dimensions rather than one, because models are getting cheaper and getting better, and the two movements do not act alike.
Figure 1: The two exposures. Everyone has some of both; the question is where the weight sits.
Read the top row against the bottom row. Sell the intelligence and you are squeezed on both dimensions: the falling price commoditises what you charge for, and the rising capability erodes what distinguishes you. Use the intelligence and you gain on both: the falling price improves your margin, and the rising capability improves your product without your having built anything. One position treats the technology’s progress as a headwind. The other treats it as a tailwind. The weather is identical.
This is a diagnostic, not an accusation. Almost every real company holds some of each, and serious businesses exist in every quadrant. The useful question is not which box you are in. It is where the weight sits, and which way it has been drifting for the last two years.
3
Why the second position is rarer than it looks
There is a simple tell, and it is worth applying without flattery, because most companies reach for the answer they would prefer rather than the one that is true.
When a better model ships, does your product get better, or does it get cheaper?
Figure 2: The tell. Which line moves on release day — the product, or the cost.
If a new release makes your demo more impressive, your outputs sharper, your accuracy higher, then the model is your product. You are in the first position, whatever the deck says. If a new release leaves the customer’s experience roughly where it was and moves a number in your cost line instead, the model is your input. You are in the second.
The first answer feels far better. Getting better for free is a pleasant morning, and there is a real temptation to describe it as momentum. But look at what it means. The improvement arrived without you — which is another way of saying that it can arrive for anybody else on the same morning, including the four companies whose demo now looks as good as yours. A roadmap written by a supplier is still a roadmap. It is not yours.
Most companies that describe themselves as AI-native hold the first exposure by construction, and it is not a failure of imagination. It is how you reach a market quickly. When capability is the scarce thing, packaging capability is the obvious business, and being early to package it is a real advantage for a while. The difficulty is that capability stops being scarce faster than anything else in the stack.
If your answer to the tell came back ambiguous, there is a sharper version of it. When the model improves, where does the benefit land first — in what the customer is willing to pay you, or in what it costs you to produce? Both are welcome. Only the second is yours to keep.
The same line is visible in what gets built. Giving every customer an assistant to ask questions of is a feature, and it will be common within the year, because the model supplies most of it. Giving every customer a service that watches their account continuously, notices what has changed, prepares the next action and stays quiet when there is nothing worth saying is an operating model — and it exists only because running it for every individual, every day, has stopped being expensive. Features arrive with the release. Operating models have to be built, which takes time your competitors also have to spend.
The second position is rare for a duller reason as well. It requires you to already have something the intelligence gets converted into — customers, a channel, a workflow, an obligation to fulfil — that existed before the intelligence was cheap and would survive if it stopped being cheap. Most young companies have not had time to accumulate one. Most older ones have, and have forgotten what it is.
Which is why the unit that matters here is not the token, the model call or the benchmark score. It is the previously uneconomic experience.
4
What qualifies for the second position
The second position needs a particular kind of opportunity: something already known to be better, which nobody could afford to supply at scale.
Three examples. Take the pattern rather than the industries.
Teaching. The best pedagogy ever documented is a patient tutor who will explain the same idea a fourth time in a different way. Nobody had to discover this; it has been understood for two thousand years and confirmed by every study since. It has also been unavailable to almost everyone, because the marginal cost of the fourth explanation was an hour of a skilled person’s life. That cost has collapsed. The idea is unchanged. The price of supplying it is not.
Service in a customer’s own language. In a country with two dozen major languages, serving each one properly used to be a capital project: a corpus to translate, agents to hire, quality to hold across all of it, forever. So most companies picked two languages and asked everyone else to cope. Translation and voice have moved from capital project to running cost. What changed is not the aspiration. It is who can afford it.
Writing to one person. A daily piece of writing composed for a single reader — their history, their situation, their interests, today’s context — was arithmetically impossible beyond a few hundred subscribers. It required an editorial desk per reader. This is the example closest to my own work, so I will hold it to the same test as the other two: the editorial cost has fallen far enough that the arithmetic works, which is a statement about supply and says nothing yet about whether anybody wants the result.
Figure 3: The crossing. Demand did not move. The cost of supplying it did.
The shape is the same in all three. Nobody invented a new desire. The desire was documented, obvious and rationed — supply was restricted by cost, and after a few decades the restriction was mistaken for the state of the world. Cheap intelligence does not create demand. It removes the reason the supply was rationed.
So the question to put to your own business is narrower than “what can AI do for us?” It is this: what did we always know customers would prefer, and refuse to build because the unit economics were impossible? That list is usually short, specific, and already written down somewhere in a plan that was rejected years ago for reasons that have quietly expired.
5
Cheap AI is not itself a moat
Now the part that arguments of this kind tend to skip.
If cheap intelligence is available to you, it is available to everybody. Your cost of production falls; so does your competitor’s, on the same day, by the same amount, from the same supplier. A cost advantage that everyone receives is not an advantage. It is a new price level, and markets find new price levels quickly.
Which means the experience that cheap intelligence makes possible cannot, on its own, defend anything. Something scarce has to sit above it.
The candidates are unglamorous and mostly old: distribution you own rather than rent; a workflow customers have arranged their own operations around; an identity or account relationship that would be tedious to recreate; accumulated attention, meaning people who choose to come back; data that arises from running a business rather than from buying a file. None of these are AI assets. That is the point of them. They are the things that do not get cheaper when models do.
Figure 4: Cheap intelligence makes the experience possible; the scarce asset defends it.
The instruction is not “own the model”. It is: own something the model makes more valuable. The model can be rented. The scarce asset should not be.
This is also where the most frequently claimed moat needs examining. Proprietary data is offered as automatic defence more often than any other asset, and most of the time it is not one. Data bought from a broker is available to anyone else who pays for it. Data that trains the same feature your competitors are already shipping defends nothing. What counts is data thrown off by a relationship or a workflow that only you occupy, and which gets denser every time a customer acts — because that is the only kind a rival cannot simply order in.
So there are two questions, not one:
— What does cheap intelligence let me build that I could not build before?
— What do I already hold that cannot be bought at the same falling price?
The strongest positions answer both, and the two answers are often found in the same place: an asset that was neglected precisely because activating it was too expensive. A dormant customer base. An archive. A licence. A physical footprint. A channel written off as tired by people who had never been able to afford to make it interesting. In each case the asset never stopped being scarce. The cost of doing anything useful with it was the problem, and that cost is exactly what has changed.
6
The honest limit
Two things this argument does not hand you.
Falling production cost protects margin, not position. Somebody with no history in your market can assemble the same experience from the same cheap models inside a quarter, and somebody will. Whatever defends you has to be the scarce layer — and most owners overestimate their own scarcity. A customer list that does not open your mail is not distribution. A workflow nobody has arranged their operations around is not a switching cost. An archive nobody searches is not an asset.
And cheap does not mean wanted. Being able to produce something for a hundredth of its former cost says nothing about whether a single person will spend attention on it. That is a separate experiment, run in the market rather than the spreadsheet, and it is the one that fails more often. Both tests have to pass, in order: can it be made, and will it be valued. Only the first has been solved for you, and it was solved by somebody else.
Which is why the question at the top of this essay needs a second line. If the best model became ten times cheaper and twice as capable tomorrow morning, would your economics improve? And if they did — what would stop everyone else’s economics improving in exactly the same way, by the same amount, on the same morning? Whatever survives the second question is the part of the business that is yours.
Figure 5: The test in four steps. The fourth is the one that decides anything.
The contrarian move in an AI boom may not be building the next model. It may be taking a neglected asset you already own and using cheap intelligence to make it valuable again.
