Your context is the moat

The base model is turning into a utility you rent by the token. The part that's actually yours is what you feed it, and feeding it is how it leaves.

Watch where the money physically goes and the story tells itself. Not into cleverness, into concrete. Gigawatts, substations, data centres that went from roughly 1.5% of the world's electricity in 2024 to a base case that about doubles by 2030, the accelerated part of it growing near 30% a year. Capital moves like that toward one kind of thing: a commodity someone has decided the whole economy is going to run on. Steel, bandwidth, power, and now raw inference. And commodities behave in a known way. The price grinds down toward the cost of producing one more unit, and it keeps going. The lead the best model holds over the second best is real, and it is also on the clock. Capability converges. It has converged in every field it has ever touched. What all this concrete is quietly manufacturing is cheap, uniform, metered intelligence, rented by the token, a little cheaper each quarter.

A general intelligence is, by its nature, one that knows a great deal in general and nothing about you in particular. It has read most of what has been written and it has never seen your book of business, the way your operation actually fails, the procedure that lives in three people's heads and on no document anywhere. A friend of mine has a name for that missing piece. He runs money for a living, wealth management, a man who could not tell you what a parameter is and would be mildly offended if you tried to explain it, but who can walk into a room and know inside a minute where the value is sitting and who is about to lose it. A year ago he told me the only thing that has ever mattered in his trade is domain knowledge, and that no machine was going to have his. He is right about what it is. Domain knowledge is the specific, slowly earned, mostly unwritten understanding of how one corner of the world really works. It is costly precisely because it takes years to accumulate, it belongs to you, and for most of history there was no way to transfer it to anyone who had not also spent the years.

That last part is where the ground has moved under him. The reasoning he was sure the machine would never match is the part turning cheap and interchangeable. The domain knowledge he was sure it could never hold is the part he now gives away himself, in small amounts, every time he asks the cheap intelligence to do something useful with it. A general model earns its keep only on the particulars, and the particulars were the whole of what he owned. You cannot consult the machine about your world without describing your world to the machine. He would point out, correctly, that the contract forbids the vendor from keeping any of it, and on the business tiers that holds: the inputs train no model, the retention is measured in days or switched off outright. But the confidentiality clause guards the wrong door. The reading was never the danger. The judgment that used to require him is now done, well enough, by a rented model holding his context for the length of a session, and a capability you rent is a capability you are quietly agreeing to stop having.

That is the careful version, the one who reads what he signs. Most of it is not careful. Most of it is someone pasting the quarter's numbers into whatever tab is open at nine at night because the tool is right there and the answer is good, and nothing happens, which is the whole trouble. A cost that never lands on the day you incur it is one people keep incurring, and more of us each year, because the edge is unmarked and the last step felt fine. There will always be someone who simply hands it over. The honest thing to admit is that the someone is increasingly us.

It never appears on the bill, which is why it goes down so smoothly. The bill says tokens, and tokens get cheaper. The figure no one writes down sits right next to it: what the work earns you, minus what you give away producing it. There is a cruder test that arrives at the same place. If a tool could truly make someone rich, the people selling it would not be metering it out by the word. They would want a piece of the outcome. When something supposedly transformative is priced like electricity instead of sold for a share of what it creates, it is worth asking, without heat, what the sellers have quietly concluded about where the value actually ends up, and on which side of the deal.

I am not building toward the conclusion that you should switch it off. Electricity is worth the bill. But no one who runs a serious operation confuses paying the utility with letting the utility onto the floor to learn how the plant is run. You pay at the meter and you keep the process behind the wall. Intelligence is about to be the utility everything runs on, and the only question that carries real weight is whether the thing that made your operation worth running stays behind the wall, or crosses it a little at a time, in the ordinary course of using the thing.

So the argument worth having was never really about how clever the models become. Clever is the part being commoditized, on schedule and in plain view. Underneath it sits the question of who ends up holding the layer where a general model meets a particular reality, because that is where the domain knowledge lives and where the value has always been. Hold it, and the model beneath you is a supplier you can change on price. Rent it on the seller's terms, and over enough time the one thing you knew that they did not becomes something the tool simply does, on hardware that is not yours. I am not certain this holds, and it is the part I keep turning over when it gets late. It may be that the knowledge stops being scarce as well, that a general model and an hour of your context come to do what the years did, and then there is no wall left because there is little behind it worth guarding. Perhaps that is where it goes. It is late, and I don't know. What I am fairly sure of is that today the common mistake is to treat how smart the machine is as the thing at stake, when the thing at stake is who gets to keep what only you knew.