According to Hugging Face's CEO, around half of Fortune 500 companies now use open-weight models rather than renting API access. Three drivers are cited: cost, privacy, and the ability to customise. This shift had been predicted for two years; it appears to have happened.
We have been documenting the rise of open models for months. Here is the most significant adoption metric we have seen, with an important caveat about its source.
The caveat, first
This claim comes from the head of the platform that hosts these models. It is not from an independent study, and the person making it has a direct interest in it being true.
The figure should therefore be treated with caution. What remains solid is the direction, consistent with everything else we are observing.
The three drivers, and their real weight
Privacy. This is probably the decisive factor, and the most underestimated. For a bank, a law firm, a hospital, the question is not whether a model is good: it is whether data can leave the building. A self-hosted model removes the question. We saw this week with the Ox Alpha episode, where thousands of developers sent code to an anonymous provider that kept it.
Cost. At high volume, the comparison changes in nature. An API is paid for per request; a self-hosted model is paid for in hardware and electricity, with a marginal cost close to zero. The tipping point depends on volume, and many organisations have crossed it. The price increase taking effect today will not slow this reasoning down.
Stability. This is the least cited driver and one of the most solid. As we explained in our article on deprecation: a model accessed via API exists only as long as someone else decides to maintain it. A downloaded file does not retire.
This movement would not have happened eighteen months ago, for a simple reason: open models were not good enough. What changed, we quantified in our comparison: three points now separate the best closed model from the best open one. And crucially, serious models now run on accessible hardware, from Qwen 3.8 at 27 billion to Meta's agent on a single card. The quality gap no longer justifies the extra cost of dependency.
What this means for closed providers
If this trend holds, it poses a problem for the dominant business model.
The labs that sell access to closed models are targeting precisely this segment: large enterprises, which pay well and in volume. Watching them move to self-hosting removes the most profitable customer base.
This partly explains several recent moves. The infrastructure race, vertical models built on proprietary corpora, and Nvidia's acquisition of Hugging Face all answer the same question: where to capture value when the model itself is given away?
The important nuance
Using open models does not mean doing without providers. Many companies run them on a cloud host rather than on their own machines. They change the nature of the dependency without removing it.
And self-hosting has a real cost that is often forgotten: skills, hardware, maintenance, updates. It is only free on the "licence" line.
What to take away
This shift, if confirmed, validates a prediction few took seriously two years ago. It also shows that the criterion for enterprise choice was never maximum performance, but the ratio between what a tool brings and what it costs in control.
This is the advice we have always given, and it applies at every scale: do not choose a model, choose an architecture that lets you change it. Half of the Fortune 500 appears to have reached the same conclusion.