What is being rented?
After renting an AI chip, where does an app send its question? On July 20, 2026, FuriosaAI announced that Samsung SDS had launched an RNGD service on Samsung Cloud Platform, with subscriptions for 1, 2, 4 or 8 cards. RNGD is an inference accelerator that runs trained models. The rental supplies computing capacity.
Weights and the execution bundle
A model must be prepared to answer on that capacity. Its trained weights and the compiled execution bundle, FXB, play different roles. Furiosa’s current documentation describes an FXB as code compiled for a model architecture plus the information needed to run it. Having the weights alone does not finish the preparation.
What can be reused?
A ready-made bundle may work. Furiosa publishes FXBs for popular models and documents reuse with changed weights when architecture compatibility is preserved. Otherwise, a suitable bundle must be prepared. Automatic reuse also requires a matching compiler version. The model and its execution bundle must fit together.
Where the app’s question arrives
Furiosa-LLM server software receives the app’s questions. It hosts one prepared model, accepts chat requests and returns answers. Its documented OpenAI-compatible API describes how requests and responses are exchanged. The model answering is the one loaded on the server; the API name does not determine its identity. The app sends requests to this server’s address.
Support closer to customers
The support organization around deployment is expanding too. On September 10, Furiosa announced a Singapore subsidiary for Asia-Pacific business development, sales and technical support. Cloud rental offers access to computing capacity; regional support gives customers a point of contact for adopting it.
This explains the execution structure using company announcements and developer documentation 2026.3.0; it is not a tested application or installation procedure for a Samsung SDS account.