A Moat Made of Tokens Is Not a Moat

A business colleague told me on a call recently that he planned to patent a process, to build a moat around a particular information-processing pipeline. I do not know the details of his pipeline, and they do not change the answer. The question is what the moat would be made of, and a pipeline that turns information into a product is made of words and logic. Tokens, in the current jargon: the word-sized pieces of text these systems read and write.

The last piece in this series argued that attention is the one thing that stays scarce. This one asks what that does to the value of whatever is made of words.

What a moat is

Warren Buffett took the image from castles. In his 2007 letter to Berkshire Hathaway shareholders he wrote that a great business needs an enduring moat, because competitors will keep assaulting any castle that earns high returns. The moat is whatever makes the assault fail.

Morningstar, which rates companies this way, sorts moats into five kinds: intangible assets such as patents, licenses, and brands; switching costs; network effects; cost advantages; and efficient scale, meaning a market with room for only a few players. Underneath all five is a single idea. Copying you is slow, expensive, or both.

The price of copying

That idea now has a price tag. In 2021, the only model that could reach a given score on a standard test of language ability cost $60 per million tokens to run. By 2024, a model with the same score cost six cents, a thousandfold drop in three years, according to an analysis by the venture firm a16z.

In 1865, William Stanley Jevons noticed something odd about coal. James Watt's engine burned it far more efficiently, and Britain responded by burning far more of it. The pattern carries his name as Jevons paradox, though it is less a paradox than the demand curve behaving: when something gets cheaper, people want more of it, and if they want enough more, total use rises even though each unit of work needs less. Tokens have been on that curve. Cheaper tokens mean more things made of tokens, in more places, by more people, and a pipeline is one of those things. People with a tool like Claude Desktop are prompting new ones into existence all day, and each is a possible copy of somebody else's.

The curve does not run forever, and Jevons's own coal shows how it ends. British coal output peaked in 1913 and has fallen ever since, as oil, gas, and electricity took its place. Demand for any one thing fills up, or a substitute arrives. Seawater is the limiting case. Plenty of people have a use for it: ships sail on it and coastal plants cool with it. It sells for almost nothing, because there is no shortage of it. Jevons explained that too, in his 1871 Theory of Political Economy: what the next unit of something is worth depends on how much of it you already have.

Bandwidth is the closest relative tokens have: a metered commodity for moving information, sold by the unit. I watched it run its course. In 1989 I worked at Timeplex, which made equipment for T1 lines, the 1.5-megabit digital links that carried a company's voice and data between sites. I sat across from the accounts-receivable clerk, who processed monthly invoices of $30,000 to $40,000 to the enterprise and government customers who could afford that much bandwidth. Prices are about 2.7 times higher now, so that is roughly $80,000 to $110,000 a month in today's money, a number a company running tokens at scale can reach in a month.

Now I stream animated 4K café scenes with computer-generated jazz to my TV all day, over gigabit fiber that costs €39 a month. A single 4K stream needs ten or more T1 lines' worth of speed, and the fiber carries about 650 of them. In 1989 only enterprises and governments could pay for that much bandwidth. Now it runs a make-believe café. Many of the fortunes made on top of it went to companies that held attention, not to the ones that owned the pipes.

Tokens of any given ability are headed the same way. Demand for them will stay large, and it still will not support a price, because supply will no longer be the constraint.

The early signs are on the spending side. Companies are learning what it costs to aim the biggest model at a small job, such as balancing a checkbook, which a spreadsheet does for nothing and a small model does for pennies. A 2023 Stanford paper called FrugalGPT showed that sending each question to the cheapest model able to answer it matched GPT-4's accuracy on some tasks at up to 98 percent lower cost. In July 2026, J.P. Morgan wrote that companies were moving away from blindly scaling AI usage and toward optimizing where it pays, and described one that scrapped an internal leaderboard after employees began burning tokens to climb it. Headlines about million-dollar token bills are common, and my bet is that token use and price are at an inflection point, where a steepening curve starts to bend. The bend may not look the way I expect: the same report put token prices up more than 60 percent since December 2025, as demand spiked for data-center capacity.

My friend Doug writes about the old line that if it were easy, everybody would do it. It got easy, and the doing is no longer rare.

A moat made of tokens is not a moat.

The patent

The legal moat is slow and narrow. In 2025 a US utility patent took about 26 months, on average, to get from filing to a decision, which is a long wait in a field where the price of the main ingredient fell a thousandfold in three years. In 2014, the Supreme Court ruled in Alice v. CLS Bank that an abstract idea does not become patentable because it runs on a generic computer, which made patents on information-handling processes harder to get and to keep. A patent also has to describe the invention well enough for others to build it, which makes the recipe public, in words.

What is left

Hold Morningstar's list up to one question: which of these is made of tokens? Patents are, and they are the slowest and narrowest of the five. Brands, network effects, and most switching costs are made of people: what they recognize, who else is there, what they have already invested. Licenses are made of permission, granted by a regulator rather than written by a model. Cost advantages and efficient scale are made of atoms: factories, warehouses, power. The software that runs a warehouse can be copied; the building, the conveyors, and the customer's trust in the people who run it cannot be prompted into existence. Almost nothing on the list can be produced by a prompt.

The ones made of people are attention in another form. A brand is attention already earned and stored. A network is attention that keeps bringing more attention. They hold up because the arithmetic from the last piece holds: attention is the one input a prompt cannot manufacture, and every competitor draws on the same fixed supply.

My own bet is that the value of intellectual property, at least the kind made of words and logic, is trending toward zero for these reasons. A drug patent protects a molecule that took a lab and years to find and a regulator to approve, which is atoms and permission, and I would not bet against it. A pipeline that a stranger with a chatbot could plausibly rebuild by Friday is a different thing.

The question a moat answers is how long a stranger would need to build the same thing. For anything made of tokens, the answer keeps getting shorter.