Last week a company in Beijing you've probably never heard of gave away a frontier AI model.

Kimi K3: 2.8 trillion parameters, benchmarks nipping at the heels of the best models money can buy, and weights you'll be able to download for free by the end of the month. The Western labs spent years and tens of billions building a lead. Moonshot AI just announced the finish line is now a public park.

The obvious reaction is "the closed labs are cooked." I think that's the boring reading. The interesting question is the one every founder and investor should be asking instead: when the smart part gets free, where does the money go?

Because it always goes somewhere. It just stops living where you're looking.

The core always gets cheap. The value just moves to the edges.

We have done this three times already

This is not a new movie. We've watched it at least three times, and it ends the same way each time.

Electricity. In 1900, if you ran a factory, you generated your own power on-site, and having a good generator was a real competitive advantage. Then the grid arrived, and within a couple of decades electricity became something you pulled from a socket without a single thought about which plant produced it. Nicholas Carr wrote a whole book about this: "The Big Switch" arguing computing would go the same way. The part that matters: the money was never really in the electron itself. Selling raw power into a competitive market is thin, cyclical work. The reliable money sat up the stack, in everything cheap power made possible, and down the stack, in owning the grid, which is a regulated near-monopoly on the wires. That's why a company like State Grid is one of the largest on earth by revenue. Not because generic electricity is a great business, but because owning the grid is. The money went up, into every appliance the cheap power suddenly made possible, and down, into owning the grid at scale. Electricity didn't change the world by being expensive. It changed the world the year it got boring.

Bandwidth. In the late '90s, everyone laid fiber like the internet would need all of it by Christmas. It didn't, not yet, and the price of moving a bit collapsed toward zero. The companies left selling raw bandwidth got a humiliating new name: dumb pipes. The money they thought they'd make evaporated up the stack, into the services cheap bandwidth suddenly made possible: Google, YouTube, Netflix, the whole internet economy, and down, into the data centers and CDNs that actually carried the traffic. The carriers owned the pipe and captured almost none of what flowed through it. Which is the exact fate a frontier lab should be losing sleep over: being the dumb pipe for intelligence.

The PC. My favorite, because it's the cleanest lesson in the history of tech. When IBM built the original PC, it didn't build the two parts that mattered. It licensed the processor from Intel and the operating system from a small outfit called Microsoft. Within a few years, essentially all the value had migrated away from IBM, the company actually assembling the box, and into the two suppliers sitting one layer above and one layer below it. IBM built the house. Microsoft and Intel collected the rent for twenty years. IBM eventually sold the PC business off entirely.

Three industries, one pattern: the core commoditizes, and value flees to the layers above and below it.

Free models, expensive edges

Now map it onto AI, because the layers line up almost suspiciously well.

The core: raw intelligence, tokens, the model itself, is the crude, the kilowatt-hour, the beige box. Kimi K3 may be the moment we get to watch it commoditize in real time.

Down from the core sits the capacity layer: chips, data centers, energy. And here's my favorite detail; a "free" open-weight model is a bit like a free puppy. The download costs nothing, and then the thing proceeds to eat your entire budget for the next three years. You cannot run a 2.8-trillion-parameter model on your laptop, or your gaming rig, or honestly anything you own. It wants a rack of GPUs that costs more than an apartment. The weights are free. The freedom is sold separately, priced by the kilowatt.

Up from the core sits the application layer: the workflows, the agents, the vertical products that turn cheap intelligence into an outcome someone will actually pay for. This is where "I have an LLM" quietly loses to "I have a thing that does your month-end close and never sleeps."

The money went up, and it went down. It skipped the middle entirely.

The middle is the trap

The middle is exactly where most AI companies are currently standing.

Here's where it gets more complicated, though: the labs understand this better than anyone, because they can see the commoditization happening in their own pricing. So rather than sit in the middle and become the dumb pipe, my bet is they keep moving up into the application layer themselves. And they get there with three advantages a startup building on top of them doesn't have.

They own the distribution. Hundreds of millions of people already use their apps by default, so they don't have to pay to acquire the customer you'd be spending heavily to win.

They own the data on what people are building. Every prompt and API call shows them which use cases are taking off on top of their models, which is a fairly direct signal of what to build next. When your product starts working, you've basically told them it was worth building.

And they can afford to underprice you. They can offer the same product at or below cost and sustain it for a long time, because the tokens that cost you real money cost them very little. That's hard to compete with head-on.

So "move up to the application layer" is necessary but not sufficient on its own, because the labs are moving there too. The more defensible spot is the part of it they're least likely to chase: a specific vertical, a complex workflow, proprietary data, or a customer relationship in a market too small or too specialized for a general-purpose lab to prioritize.

It's roughly the same thing I argued in my last post, from a different angle. I don't think the value was ever really in the intelligence itself. It's in what the intelligence gets situated in: the data, the workflow, the accumulated context that a free model doesn't have about your business and can't get just by cutting its price. The frontier is a public park now. Parks don't have moats. Whatever you build around them does, especially the parts that are hard for the park's owner to build too.

Kimi K3 didn't make intelligence worthless. It made it ambient. That's good news if you have a real business sitting on top of it, and a problem if the model was the business.

The token is heading toward zero.