The public AI conversation is about the frontier — the biggest models, the hardest benchmarks, the latest releases from three or four labs. The private conversation inside enterprises is different: it is about unit economics, data custody and vendor risk, and on all three axes, open-weight models keep winning deals.
The pattern shows up in deployment data, in cloud marketplace listings, and in the quiet fact that most corporate AI pilots that reach production are no longer calling a frontier API for every request.
Why it matters
If the volume layer of enterprise AI runs on open weights, the economics of the frontier labs depend on a narrow slice of high-value work — which pressures their pricing and changes what the multi-hundred-billion-dollar infrastructure buildout is actually for.
It also redistributes power in the stack: value accrues to whoever runs inference efficiently and to the tooling layer around open models, rather than to a single model provider.
How the split works in practice
The architecture that has become standard is a router: cheap, fast, self-hosted open models handle the bulk of requests; frontier APIs handle the exceptions — complex reasoning, ambiguous instructions, tasks where a failure costs more than the API call. Fine-tuning tilts the balance further: a small open model tuned on a company's own data frequently beats a frontier model prompted generically, at a tenth of the cost.
Data custody seals the deal in regulated industries. A hospital or bank that cannot send records to an external API can still run an open model inside its own cloud boundary — which means open weights are often not the cheaper option but the only option.
Evidence
Download counts for open-weight models run into the hundreds of millions; enterprise surveys consistently show a majority of organizations with production AI using at least one open-weight model; and the major clouds all now offer hosted open models as first-class products, which they would not do against their own proprietary partners' interests unless customers demanded it.
The capability gap has narrowed in step: each open-weight generation lands closer to the frontier of a year earlier, and for a large class of business tasks, last year's frontier is more than sufficient.
The competing read
Frontier labs argue the gap that matters is not today's benchmarks but the trajectory — that agentic, multi-step work rewards the best model by a compounding margin, and that enterprises routing everything to cheap models are optimizing for last year's workloads. Open-weight advocates answer that the trajectory cuts both ways: every frontier capability becomes an open-weight capability within a year, at a fraction of the cost.
What happens next
Watch the licensing terms — the definition of 'open' is contested and tightening — and watch whether the router architecture hardens into standard enterprise plumbing, which would make the model layer genuinely interchangeable for the first time.