The Knowledge Economy When a Model Becomes Capital: Cheap to Copy, Costly to Build

Write a programme once, then copy it onto a million devices without paying anything worth mentioning. An AI model combines two opposites: it is as cheap to copy as software, yet as costly to build the first time as a factory. In that combination lies the heart of what we might call the knowledge economy of AI.

Knowledge that is not used up, yet is not produced for free

Paul Romer won the Nobel Prize in Economics in 2018 "for integrating technological innovations into long-run macroeconomic analysis". One of his central ideas is that ideas are "non-rival" goods: my use of an idea does not stop you from using it, unlike a field or a machine. The prize committee describes the structure that follows: a first blueprint with a large fixed cost, then copies at a small marginal cost. That is why an innovator needs a margin of profit and some market power to recover the cost of innovating.

Cheap to run, costly to found

The figures show it. In a study published in June 2024, Epoch AI estimates that the cost of training frontier models has grown roughly two to three times a year over eight years, and that the largest training runs could exceed a billion dollars by 2027 if the trend continues. That is an estimate resting on the assumption of continuity, not a verdict on what will happen.

Against this, the Stanford AI Index (2025) reports that running a model at the level of GPT-3.5 cost 20 dollars per million tokens in November 2022 and about 0.07 dollars in October 2024. Use becomes cheaper while foundations become dearer. The door widens for those who use models and narrows for those who want to build a frontier one.

Wealth that cannot be touched

Ocean Tomo's study says intangible assets made up 17% of the market value of the S&P 500 companies in 1975 and about 92% by the end of 2025. The figure has its limits: it measures market value, not output or welfare, and it covers many assets besides AI. Yet it points in a clear direction: most of what is bought and sold is now what cannot be seen.

Research by Brynjolfsson, Rock and Syverson explains why the effects of major technologies reach the statistics late. They need complementary intangible investment, such as new processes, business models and skills, which national accounts capture poorly. The effect therefore looks weak at first and appears later, in what they call the "productivity J-curve". They find that the measured effect of AI is still small but growing.

What does one actually own?

Open code is available, but it does not suffice to build a frontier model. Power lives in the whole system: computing to run experiments, data to feed them, and people who know how to build. The same index shows large disparities in funding: private AI investment in 2024 reached about 109.1 billion dollars in the United States, against 9.3 billion in China and 4.5 billion in the United Kingdom. It excludes public spending, so it should not be read as a sole measure of strength. For the Arab picture, see the article "Arabic Artificial Intelligence: Do We Make or Consume? Sovereignty and Arabic Language Models".

Questions before enthusiasm

The big numbers describe nations and giant companies; a small institution's question is practical: what do we actually own of this system?

• Data: do we know what we have, and is it kept faithfully and in line with regulations?
• Ability to run: do we rent or build, and what does each choice cost over the medium term?
• Skills: do we rely on ready-made tools alone, or develop people who understand them and can question them?
• Distribution of benefit: does the technology's effect reach workers and beneficiaries, or stay in a narrow layer?

These questions offer knowledge, not financial advice; anyone facing an investment decision should consult a licensed specialist and observe the rules of Sharia. The principle remains that knowledge is a trust, for whose good use one will be asked, and that the blessing of learning calls for gratitude expressed in fairness in sharing its benefit.