Open Weights Are Winning the Parts That Matter

The benchmark leaderboards tell one story: closed frontier models lead. The invoice data tells another: the fastest-growing share of real-world AI workloads runs on open weights.

The reasons are structural. Fine-tuning, on-premises deployment, predictable costs and freedom from a single vendor's pricing decisions are not nice-to-haves — they are purchase requirements in finance, healthcare, manufacturing and the public sector. Meta's Llama family and a wave of smaller specialist models have made "good enough, fully controlled" the default answer for a growing list of workloads.

Closed models keep their edge at the frontier: the hardest reasoning, the longest contexts, the newest modalities. That edge is real and valuable. But the frontier is a thin slice of demand. Most business problems are classification, extraction and summarization at scale — precisely the jobs where a tuned seven-billion-parameter model on your own hardware is indistinguishable from a frontier API, at a fraction of the cost.

The strategic consequence is that "open vs. closed" is the wrong question. The right one is which parts of the stack become commodities — and the answer, year after year, has been the same: one more layer than the market expected.