Build vs buy just got a third option: own
Llama 2 is free for commercial use, with Azure as preferred home. What open-weight models change for enterprises — and what they conveniently don't.
🔓 Build vs buy just got a third option: own.
Last week Meta released Llama 2 — free for research and commercial use, in three sizes, with Microsoft as preferred partner and the models available on Azure. Whatever you think of the strategy behind it, the practical fact is new: a credible language model your organisation can run inside its own boundary, fine-tune on its own data, and pay no per-query licence for.
For enterprise AI decisions, that quietly rewrites a familiar equation. "Build vs buy" always really meant "buy vs buy" — you were choosing which vendor's model to rent through an API. Open weights add a genuinely different column: own. And owning changes specific things:
The data never has to leave. For workloads where the sensitivity conversation kills the project — legal, HR, anything with client names — a model running inside your perimeter changes what is even proposable.
The economics invert at scale. APIs are cheap to start and expensive at volume; self-hosted models are the opposite. For high-volume, narrow tasks, the cost curve now bends in your favour.
The dependency risk drops. A capability you run is one no vendor can reprice, deprecate or acquire out from under you.
But — and this is the part the excitement will skip — open weights transfer responsibility along with control. There is no vendor to call when output quality drifts. Hosting, tuning, evaluation, safety and upkeep land on your team, and that team now needs muscles most enterprises have not built. A free model with no MLOps capability behind it is not a cost saving; it is an unstaffed product.
So the honest decision framework, as I see it: rent for breadth and speed, own for depth and sensitivity — and let the workload, not the ideology, pick the column. The organisations that will use this well are those that already know their use cases, volumes and data boundaries precisely enough to do that math per-workload. Which is, quietly, the real test this release sets: not "should we use open models?" but "do we understand our own AI demand well enough to know?"
One exercise before anyone gets excited in your next strategy review: list your three highest-volume AI use cases and, for each, write down the monthly query volume, the data sensitivity, and what a vendor outage would cost you. If nobody can produce those three numbers, the open-vs-closed debate is premature — you are choosing between answers to a question you have not yet asked.
Reading: https://ai.meta.com/blog/llama-2/ https://blogs.microsoft.com/blog/2023/07/18/microsoft-and-meta-expand-their-ai-partnership-with-llama-2-on-azure-and-windows/