A telco and a ride-hailing app in Jakarta now ship their own language model. Indosat Ooredoo Hutchison and GoTo built Sahabat-AI, an open model that runs in Bahasa Indonesia and its local tongues, and neither company is an AI company in the sense the phrase usually means. That is the tell.

For two years the regional default was simple. You rented frontier access from an American lab, wrote a thin layer of product on top, and told your board the model was someone else’s problem. The default has quietly stopped being safe. When the access you rent can be gated, metered, or repriced by a policy you do not sit in the room for, it stops being a supply and starts being a loan.

The build-versus-rent line in Southeast Asia has moved, and it has moved toward build. Not because the region can suddenly train a better model than the American labs; it cannot. It has moved because a model you own outright, even a worse one, is now worth more to certain founders than a better one you borrow.

Watch who is crossing. AI Singapore has worked since 2023 on SEA-LION, an open family that reads eleven Southeast Asian languages the frontier labs treat as an afterthought. It was built to prove a plain point, one the people behind it state without hedging: the region does not have to use only OpenAI. Sahabat-AI in Indonesia is the same instinct with a telco’s balance sheet behind it. The model launched at eight and nine billion parameters, small by frontier standards, and runs on Indosat’s own GPU Merdeka sovereign cloud rather than a hyperscaler’s. Alibaba’s DAMO lab built SeaLLM on the same premise from the Chinese side. Three efforts, one shape.

The compute an operator can count on is now a function of whose flag its parent flies, and that is not a number a product team can control from a desk in Bandung.

The thing that changed the maths sits upstream of any of them. In January 2025 Washington rewrote its chip rules to police the ultimate owner of a company rather than the address of its servers, and in July Malaysia added its own permit for moving high-end US chips across its border. The compute an operator can count on is now a function of whose flag its parent flies, and that is not a number a product team can control from a desk in Bandung. Renting frontier access assumes the pipe stays open. Building assumes it might not.

None of this makes owning a model the right call for most founders. It is expensive, the talent is thin, and a self-hosted nine-billion-parameter model will lose to a rented frontier one on almost any benchmark you name. For a company selling to consumers in a language the frontier labs already serve well, renting is still correct, and will be for years. The founders crossing the line are the ones whose product breaks in a language, a jurisdiction, or a data rule the rented model was never going to respect. For them the worse model that answers is worth more than the better one that might one day be switched off.1

A model you own outright, even a worse one, is now worth more to certain founders than a better one you borrow.

The reflex to watch is the one that treats last year’s default as a law of nature. The operators still telling their boards the model is someone else’s problem are making a bet that access stays cheap and open. Sahabat-AI is what the other side of that bet looks like: a ride-hailing company deciding the safest model to depend on was the one it did not have to ask anyone for.

Footnotes

  1. The counterargument, which is real: most of the sovereign models shipping today are continue-trained on American or Chinese open weights, so the independence is partial. You own the model. You still borrowed the foundation. ↩