A year ago, when an operator asked whether they could run their AI work on weights they owned, the honest answer was usually no. The open models lagged the frontier on the tasks that mattered, and the gap was wide enough that the trade was hard to justify.
That answer is no longer true.
In the last month alone, five frontier-class open-weight models have shipped — Llama 4, Qwen 3.5, DeepSeek V4, Gemma 4, and Mistral Medium 3.5. On specialized benchmarks the open models are competitive with the closed ones. On general capability the gap is narrow enough that, for most operational workloads, it is no longer the deciding factor. Enterprise deployment of open weights moved from twenty-three percent to sixty-seven percent of organizations in a year. The buyer's question flipped from "can we get away with open?" to "which open-weight model do we deploy?"
We have written separately about the shape this opens up — the broom-closet shift on the hardware side, the barbell on the broader infrastructure split. The argument we want to make here is the operational one.
Running your own weights is not a purity test. It is not a political statement. It is an operational choice with three real consequences. Your data does not leave your environment. Your costs are a depreciation schedule on hardware you own, not a per-token bill on someone else's model. Your dependency surface — the set of vendors whose decisions can break your operation overnight — gets smaller.
The cost of the choice is real. You take on the work of running, updating, evaluating, and securing the model yourself. The frontier closed models will, for some time, hold a lead on the hardest tasks. If your business depends on that last mile of capability, you will keep paying for it.
Most operations do not depend on that last mile. Most operations depend on a model that is good enough, predictable, owned, and cheap at volume. That model now exists, and it does not require a frontier lab's permission to use.
Sovereignty over your AI stack used to be a luxury and an aspiration. This year, it became a procurement decision.