AI may not reward the builders after all

If artificial intelligence is truly a general economic technology, why is so much of the money still being made by selling software built on top of it? The answer may be uncomfortable: AI could lower the value of creation itself and shift even more power to those who own capital, distribution and infrastructure.

Artificial intelligence is sold, above all, as a story of empowerment. A lone founder can build more. A small team can move faster. A worker can do the job of three. The promise is familiar enough: technology lowers the cost of creation and spreads economic power more widely.

But what if AI points in the opposite direction?

For all the grand claims made on its behalf, the commercial logic of the industry remains oddly prosaic. If these systems are as economically powerful as enthusiasts suggest, why is so much of the business model still based on wrapping models in software and charging subscriptions? Why the endless crop of copilots, assistants and application-layer tools, rather than a more direct deployment of machine judgement into capital itself?

Put bluntly, if tokens are so valuable, why are they not being turned more directly into money?

This is not a frivolous challenge. It goes to the distinction between a useful tool and a genuinely general economic intelligence. A model can be dazzling in demonstration and highly productive in practice without being able to generate durable returns in a competitive environment. It can lower the cost of writing, coding, designing and summarising without yet possessing the kind of judgement that matters most in markets: the ability to make better decisions than rivals when uncertainty, cost and consequence are real.

If AI were already that capable, one might expect more of the industry’s centre of gravity to shift towards direct capital allocation — trading, underwriting, lending, pricing, procurement, insurance. These are all domains in which superior judgement compounds quickly into profit. Instead, most AI monetisation still sits at a safer distance. The models propose; humans decide. The revenue comes not from autonomous economic agency but from software fees.

There are obvious objections. Public markets, in particular, are a punishing test of intelligence. They are adversarial, reflexive and crowded with actors trying to extract the same signals. Any visible edge is quickly competed away. Success depends not only on judgement but on infrastructure, data, speed, risk management and institutional trust. The fact that large language models have not replaced portfolio managers overnight proves very little.

True. But neither does it prove the opposite. Markets are useful precisely because they are unforgiving. They strip away some of the ambiguity that clouds AI’s economic claims elsewhere. In enterprise software, it is easy to confuse adoption with dependence, novelty with lasting value, convenience with genuine economic advantage. In capital allocation, the scoreboard is much harder to manipulate. Either the machine makes better decisions than the alternative, after costs, or it does not.

That is why the pattern of AI monetisation matters. It suggests that the technology may be highly valuable while still being less economically general than the rhetoric implies. Today’s models look more like systems for compressing the cost of cognition than like autonomous engines of judgement. They augment labour brilliantly. They do not yet obviously substitute for capital allocation itself.

But there is a more consequential interpretation. Perhaps the real significance of AI is not that it will immediately turn models into traders. Perhaps it is that, by making intelligence cheaper, it will make ownership more important.

This is the possibility the industry is less eager to discuss. If AI reduces the cost of producing code, text, analysis, design and even strategic recommendations, then the act of building becomes less scarce. And when a crucial input becomes abundant, rents do not disappear. They migrate. They flow to the owners of whatever remains scarce.

In the age of AI, those scarce assets are unlikely to be intelligence alone. They are more likely to be capital, proprietary data, distribution, compute, customer relationships, regulatory permissions, legal rights and trusted brands. In other words, the winners may not be those who create the most, but those who own the channels through which cheap intelligence is converted into cash.

That would mark a deep break with the self-image of the digital economy. For years, the software era nourished the idea that small teams and talented builders could routinely challenge incumbents. AI appears, at first glance, to supercharge that promise. Yet abundant intelligence may do to cognition what industrial machinery did to artisanal production: raise output, lower unit value and shift bargaining power towards owners of scale.

This no longer looks especially hypothetical. Much of the durable value in AI has so far accrued not to the legion of application-layer experiments, but to those controlling the bottlenecks around them — chipmakers, cloud platforms, incumbent software distributors and enterprises with proprietary workflows and customer access. The glamour sits with the interface. The leverage sits deeper in the stack.

That does not mean application-layer companies are trivial, or that AI products will not create substantial wealth. Many already do. But it does suggest that the current boom may be misread if it is seen only as a renaissance for builders. It may also be the beginning of a harsher economic regime, one in which creation becomes easier, cheaper and more abundant — and therefore captures a smaller share of value than before.

If that is right, the profusion of AI wrappers is not just a temporary phase before something more profound arrives. It is itself a clue. It suggests that the easiest way to monetise AI today is still through familiar software channels, because direct autonomous deployment remains technically fragile, institutionally constrained and economically contested. But if the technology continues to improve, the larger prize may go not to those with the cleverest product, but to those with the balance sheet and operational reach to embed machine judgement directly into decisions that move capital.

The romantic view of AI is that it will democratise genius. The more plausible one is that it will commoditise much of what passes for knowledge work and raise the premium on ownership.

That is the question lurking beneath the hype. Not whether AI can write an email, draft a memo or produce passable code. But whether, once intelligence becomes cheap, the rewards will flow to the builders — or to those who own the assets that matter when building no longer does.

The most important economic story in AI may turn out to be a very old one: when a vital input becomes abundant, capital wins.