AI will put AI out of business
The frontier labs are racing to build the smartest models. But the open models are catching up. And the on-device ones are catching up too. The endgame isn't who wins the race. It's that the race stops mattering.
The gap is closing, not widening
A year ago, the gap between the best closed model and the best open one felt like a chasm. Today it feels like a few months. Llama, Qwen, DeepSeek, Kimi K2, MiniMax M2, GLM, Mistral. They trail the frontier, but they trail it close enough that for most real work, you can't tell. And the trailing distance keeps shrinking, not growing.
The same thing is happening with on-device. Apple's research keeps showing that small models, quantized properly, running on the silicon already in your pocket, can handle the bulk of what people actually do with AI. Summarize this, rewrite that, answer this question, draft that reply. The frontier keeps moving up, but the floor moves up faster.
Intelligence is becoming a commodity
When something becomes a commodity, the people selling it stop making money. The people using it do. Electricity didn't make power companies the winners of the 20th century. It made everyone who plugged something into the wall a winner. Bandwidth didn't make ISPs the winners of the internet. It made everyone who built on top of cheap bandwidth a winner.
Intelligence is on the same path. The price per token has been collapsing for two years straight. The capability per dollar has been climbing just as fast. At some point soon, the marginal cost of a smart answer is effectively zero, and the only question left is what you do with it.
The winners are not the AI companies
The winners are the people who can run a capable model on their phone, for free, without sending their data anywhere. The winners are the small teams who build products on top of cheap intelligence instead of paying rent to an API. The winners are the people who own the distribution. The device, the browser, the editor, the workflow. Not the people who own the model.
This is the part that should make the frontier labs nervous. They are spending tens of billions to build something that, six months later, someone else gives away for free. They have to keep running just to stay in the same place. The open ecosystem doesn't have to win the frontier. It just has to be good enough, often enough, cheaply enough, that the API business shrinks to a thin slice of hard problems.
The moat that isn't
For a while the story was that scale was the moat. Then it was data. Then it was reasoning. Then agents. Each moat held for about a year before someone open-sourced something that crossed it. Kimi K2 showed up with agentic capability that was supposed to be a frontier-only thing. MiniMax M2 did the same on long context and tool use. DeepSeek did it on reasoning. Qwen did it on multilingual and on coding. The pattern is consistent enough that you can stop being surprised by it. There is no moat in pure capability. There might be one in distribution, in product, in trust, in privacy. But not in the model itself.
This is also why Apple's angle is interesting. Privacy plus local execution is a moat the API players genuinely cannot cross, because crossing it means giving up their business model. You can't be a company that sells cloud inference and also a company that runs everything on the user's device. Apple can. Apple already does.
The frontier still matters, just less
None of this means the frontier stops mattering. The hardest 10% of problems. Deep research, long-horizon agentic work, novel science. Those will probably stay with the big labs for a while. That's real, and it's valuable. But it's a smaller market than people think, and it's a market where the customer is usually another business, not a person.
The other 90%, the part that touches daily life, the part most people will ever interact with, is being eaten by open weights and the silicon you already own. That's where the value moves to the user. Not because anyone planned it that way, but because that's what happens when a technology gets cheap enough.
The good ending
AI will put AI out of business. The frontier labs will keep pushing the ceiling, and the open ecosystem will keep raising the floor, and the floor will keep getting closer to the ceiling, and at some point the distinction stops being interesting. What's left is a world where smart software is cheap, local, and yours.
That's the best outcome we could hope for. And it's the one that's actually happening.