How do you see airgapped/self-hosted AI fitting into enterprise security requirements?

For context I work with teams in regulated industries where cloud AI services are often blocked by security/compliance (think SOC 2, HIPAA, FedRAMP, or just strict internal policies about code leaving the network). The usual workaround is... nothing. Teams just don't get AI assistance.

So discussing here on:
- What's driving the demand for airgapped AI? Is it primarily compliance requirements, or are you seeing general enterprise nervousness about proprietary code going to third-party APIs?

- How does local Llama 4 perform for real coding assistance compared to Claude/GPT-5 etc..? Is the tradeoff acceptable, or do most self-hosted users eventually connect to cloud providers anyway?

Curious to discuss both the product vision and real-world experiences.