When the Frontier Goes Open

For most of the last three years, the frontier was a private address. If you wanted the best model, you rented it from one of a few labs, through their API, on their terms.

That address is now public.

The open-weight models — DeepSeek, Qwen, GLM, Kimi, Llama, Mistral, and a growing crowd behind them — have closed the gap to a rounding error. Not "good for open source." Not "fine if you can't afford the real thing." Within a benchmark point or two of the closed flagships on the tests most companies actually care about, under licenses permissive enough to run in your own data center, at a fraction of the price per token.

You can download frontier-class intelligence now. The weights fit on a hard drive.

The first-order panic

The obvious reaction is that the labs are cooked. If the thing they spent billions to build is available for free within a few months, where is the business? The moat was supposed to be the model. The model just became a commodity.

I think that read is correct and shallow at the same time. Correct, because the specific advantage of "we have the smartest weights and you can't have them" really is evaporating. Shallow, because it assumes the value was ever mostly in the weights.

This is the second-order series, so you already know where this goes. Make something abundant and the scarcity doesn't disappear. It moves next door. Cheap intelligence didn't kill the value in AI. It relocated it. Open weights are the same move, run one level deeper: they make the frontier itself abundant, and push the scarce work out to everything the frontier was quietly sitting on top of.

What a license doesn't give you

Download the weights and you have an engine. You do not have a running car.

Serving a frontier model well is its own discipline — latency budgets, uptime, autoscaling, security, guardrails, evaluation harnesses, and the unglamorous work of keeping a self-hosted checkpoint current while three new ones ship every month. The open model hands you the hardest part for free and leaves you the second-hardest part entirely. Plenty of companies are about to learn that "we self-host now" was a decision to hire an inference team.

Then there's the part no license can transfer: accountability. When the model is free and everyone can run the same checkpoint, the differentiator is not who has access to it — it's who will put their name on the output. Someone has to certify that this deployment is safe, that this answer is sound, that this system did what it was supposed to when it mattered. Open weights make the intelligence common and the ownership of the outcome rare. That's the trust scarcity again, wearing new clothes.

And underneath both sits the thing a competitor genuinely cannot clone by running your model: your data and your distribution. The proprietary context you feed it. The workflow it's embedded in. The place it already lives, in front of the users who already trust it. Two companies running the identical open checkpoint are not equally valuable — the gap between them is everything the checkpoint doesn't contain.

Who this actually helps

The winners of an open frontier are not who the "labs are dead" headline implies.

Regulated and sovereign buyers win big. A hospital, a bank, a defense agency, a government that could never send its most sensitive data to someone else's API can now run frontier-class intelligence entirely inside its own walls. That was a hard "no" a year ago. It's a "yes" now, and it unlocks a whole tier of demand that closed models structurally could not serve.

Small firms win too, in the way this series keeps predicting — they get capabilities that used to require a platform contract, and they can build on a foundation nobody can revoke or reprice out from under them. The floor under everyone just rose.

The people who lose are the ones whose entire pitch was intermediary access to a model — reselling tokens with a thin wrapper. When the model is free, "we'll call the API for you" is not a business.

The efficiency stack, again

There's a connection to the efficiency work worth naming. Open weights are not just cheaper — they're malleable in a way closed models never are. You can quantize them, distill them, harden them, certify them, and even etch them into silicon without waiting for a vendor to expose the right knob. The most compressible, deployable, embeddable intelligence in the world is now the intelligence anyone can take apart. That makes open weights the natural substrate for the industrial, delivery-economics phase of AI — not despite being free, but because being free is what lets you optimize them all the way down.

Abundant intelligence and open intelligence are the same story told at two altitudes. Both make the model cheap. Both move the money to the stack around it.

The reframe

So the frontier went open. The right question isn't "who still owns the best model" — increasingly, nobody does, and that's the point. The question this series has been asking all along still holds: once the impressive thing is everywhere, what's suddenly scarce?

The answer is the same as it's been every time. Not the intelligence. The judgment to deploy it, the accountability to stand behind it, and the data and distribution to make it matter.

You can download the frontier now. You still can't download the hard part.