9
The AI conversation is quietly shifting. It used to be about which frontier model was smartest. Now the sharper question is: which weights are yours to run?
The shift from capability to control
For years, the debate was dominated by a single metric: which model scored highest on the latest benchmark. Firms chased the smartest engine, as though picking a model were the whole of their AI strategy. That framing is fading — and it is fading for a reason. The capability gap between the frontier and the open-weights families has narrowed dramatically, while the question of who controls the data has become the one that actually matters.
Public AI is powerful — but it processes your data on someone else’s infrastructure, under someone else’s terms, with no chain of custody you can inspect. For a firm that handles privileged material, tax returns, or controlled unclassified information, that’s not a feature. It’s a liability.
What open-weights actually gives you
Open-weights models — the Llama, Nemotron, Qwen, Mistral, and DeepSeek-class families — have closed most of the capability gap. And because the weights are yours, you can run them on infrastructure you own or lease, behind your own controls, with a data trail you can actually audit.
That is the difference between using AI and controlling AI. The capability is nearly the same. What changes is who holds the data, and who can prove it. When the model runs on your hardware, your documents never leave your perimeter. The answers are generated where the data lives, from the data itself, and every query leaves a record you can produce.
Why the distinction is the whole ballgame
For law firms protecting privilege, CPAs bound by §7216, and government contractors under CMMC, that single distinction is the whole ballgame. It’s why “just use ChatGPT” is the wrong answer — and why a sovereign setup pays for itself the first time an insurer, a client, or a regulator asks the question.
The firms that will lead are not the ones with the cleverest model. They are the ones that can answer, with confidence and a record, where their data runs and who can prove what happened to it. Open-weights is the mechanism. Sovereignty is the outcome.
What this means for your firm’s AI roadmap
Adopting open-weights is not a downgrade in ambition. It is a reordering of priorities — capability first, then control. The practical path is straightforward: choose a model family that meets your needs, run it on hardware you own or lease, connect it to your own corpus, and log what it does. The result is an AI system that behaves like every other system in your firm: governed, auditable, and inside the perimeter.
This is also the answer to the question every firm will eventually face: what happens when a client, insurer, or regulator asks where your data went? With open-weights on your own infrastructure, the answer is a record you can produce. With public AI, the answer is a promise you cannot verify.
Know where you stand
Not sure where your firm stands? The Confidential AI Exposure Checklist takes five minutes and tells you exactly that. Nine lives. Zero leaks.