Singtel launches managed AI token service for enterprises
Singtel combines model access, routing and governance in a token service with Singapore hosting and other cloud options for less sensitive work.
Singtel launched AI Token as a Service on October 1, giving enterprises managed access to multiple models through its RE AI cloud business. Customers can choose token volumes instead of managing each model separately.
The announcement names Qwen, Mistral, GLM, Kimi and MiniMax alongside selected frontier models. Enterprises can also bring their own models.
One connection with central controls
The official product page describes an API compatible with OpenAI, with a common endpoint for supported models. Singtel says requests are authenticated, checked against access permissions, limited by request rates and metered before reaching a model. Teams can manage model permissions, consumption and spending centrally.
Customers can set routing rules around cost, performance, model choice and sovereignty. Usage and audit logs are intended to make that activity visible across employees, teams and applications.
Singapore hosting depends on the workload
Singtel says its sovereign environment keeps data within Singapore. It also permits clouds outside that sovereign environment for less sensitive work. The subscription supports uses including internal queries, software development and agent workflows.
For an enterprise buyer, the practical test is the route a particular request takes. A procurement team should establish which models can process protected data locally, how fallback destinations are restricted and whether logs or cached responses follow the same residency requirements. A common interface does not answer those questions by itself.
The bill still needs a workload test
The service page describes billing by token with a provider margin, rather than charging customers for idle GPU capacity. It offers private pricing for committed consumption volumes and directs prospective customers to contact Singtel. The pages reviewed do not provide a public numerical rate card.
That leaves buyers without enough public information to calculate savings for their own traffic. A useful pilot would compare the full bill, response quality and completion time on representative tasks. It should also measure what happens when a cheaper model needs another attempt or a more capable model takes over.
A service built on a wider infrastructure business
The launch sits within Singtel Digital InfraCo, the infrastructure unit formed in 2023. Its portfolio combines Nxera data centres, network connectivity and Paragon orchestration. RE AI brings those capabilities together as a cloud platform. Token consumption adds a managed access layer to that infrastructure business, alongside its existing GPU services.
ByteForward previously covered NVIDIA NeMo Switchyard, a library for routing work between models. Singtel packages model access and operational controls into a managed commercial service, making the buying decision as much about service terms as model selection.
Featured image is an original AI generated conceptual illustration of routing between computing environments.
