Six months ago, every enterprise customer asked us which frontier model to use. Now half of them are asking how to run post-training on their own data, host their own weights, and keep everything inside their network.
Once the capability gap between open source and frontier closed, the cost argument became overwhelming. You can run a fine-tuned open model for your specific use case at a fraction of the cost of API calls to a frontier lab.
All the big cos with enough data to fine-tune are moving in this direction.