Big AI companies prefer a louder story: a computer beats a chess champion, wins a quiz show, beats the world's best Go player, or lately, solves a hard math problem that stood for decades. Impressive stories. None of them tell you how to use the technology yourself.
There is an older, quieter story that does.
The coal problem Britain worried about in 1865
It comes from a British economist named William Stanley Jevons.
In 1865, Britain worried it was running out of coal. Jevons wrote a book about the problem. He looked at the new steam engine James Watt had built. Watt's engine used far less coal than the engines before it.
Most people expected coal use to fall. Jevons said the opposite would happen. A cheaper resource does not shrink demand, he said. It grows demand, because people find new uses for it once it is cheap.
He was right. Coal use did not fall after Watt's engine. It grew, fast. This pattern later got a name: Jevons' paradox. Make something more efficient, and you often do not save the resource. You use more of it.
Apple just quietly proved a 150-year-old prediction about AI
Apple did not set out to prove a 19th-century economist right. It set out to sell chips.
But Apples new unified memory structure is exactly the kind of jump Jevons described. It does not make AI a little cheaper to run. For anyone willing to run the model on their own machine instead of renting one, it removes the meter completely. And a resource with no meter does not get used a little more. It gets used far more, the same way cheap coal did.
That is Jevons' paradox, in silicon.
Metered AI is about to look like a strange idea
Right now, most AI work still happens on a meter. You open a tool in your browser. You send a request. A server somewhere else processes it and sends a result back. You pay for that, either as a monthly fee or a charge per use.
A meter changes how you behave. You ration a metered thing. You ask once, not five times. You accept the first result instead of trying three more. You do not experiment much, because experimenting costs money.
Move that same AI onto a Mac you already own, and the meter stops mattering. The model runs on your own hardware. Nobody bills you per request. The cost of the next question is close to zero: a little electricity, nothing more.
People do not use AI less when this happens. They use it far more. They try five versions of something instead of one. They let it run longer. They test ideas they would never have paid to test.
What this means for building a website
Most website tools work the way rented AI works. You open the tool in your browser. Every AI feature, every text suggestion, every image, runs on a server you do not own and cannot see. You pay per action, per seat, per month, sometimes per word.
The browser adds a second limit on top of the bill. It is built to keep a web page away from the real computer underneath it, the one with the chip we just talked about, sitting there doing nothing.
We have written before about why a browser is the wrong place to build a website. This article is about what happens next, once the AI itself comes home too.
Splitleaf is not a page you open in a browser. It is an app that runs on your own Mac or Windows PC, working on real files, in a real folder, that you own. Today, the app already does your design and content work locally, off any meter. Publishing and signing in still need the internet, because a live website needs a live host.
The AI agent is the part still catching up. On-device AI is still young. We are building toward it. We are not claiming it today.
That direction is the point of this article. Once an agent runs fully on the machine you already own, the cost of one more version, one more idea, one more fix, drops close to zero. You stop rationing your own AI. You start using it the way you use a tool you already bought, because that is exactly what it is.
Every person, their own agent
This matters even more for a team. Rented AI usually charges by the seat. The bill grows with every person you add. An agent that runs on each person's own machine does not work that way. Each designer and each writer on your team runs their own agent, on their own computer, without a shared meter counting every request against one bill.
Not a cheaper way to do the same thing
The real change is not that AI becomes a cheaper version of what you already do. It becomes affordable to do things that were never worth paying for before: the tenth headline, the version you would have skipped, the small fix nobody wanted to spend a token on.
We are not there yet on every part of this. But it is the direction we are building Splitleaf toward. It will not make a headline the way a chess match or a math problem does. It will just quietly change how much you build, starting with the chip already sitting on your desk.