A few days ago I was talking with a friend who works on-site at one of Indonesia's big tech companies. That conversation made something click: plenty of companies run local models on their own on-premise servers instead of calling the OpenAI or Anthropic API, and they have a long workflow before anything ships to production.
Why local models at all?
The reasoning is pretty reasonable: privacy and security. When the data touches identities, internal documents, or sensitive customer information, a company is going to be careful about where that data travels and where it gets stored.
With a local model, everything runs directly on servers the company owns. The data never has to leave for a third-party AI provider, the infrastructure stays more controlled, and for heavy long-term usage the cost tends to be more predictable.
On-premise isn't free of trade-offs
Running your own hardware means buying and maintaining the servers, dealing with GPUs, power, networking, security patches, and monitoring — plus an infra team ready to respond when something breaks. You get more control, but you also carry more responsibility.
The deployment is deliberately slow
I've also heard stories from other companies about how long their deployment process is. The code is pushed, the PR is merged, and it still doesn't go straight to staging or production. There's a security review, testing, sometimes even penetration testing in between.
So at a large company, slow deployment isn't necessarily a sign of a bad workflow. Sometimes it's intentionally layered, because the priority isn't only shipping fast. It's keeping data private, keeping the system secure, and making sure a change doesn't quietly introduce new risk.