Singtel token service sells AI usage by the token, inside Singapore

The Singtel token service is a telco selling model calls the way it once sold minutes, and the October 1, 2026 release is careful about what it does not print. There is no public price per token in the release that loaded. There is a stack: models, routing, a sovereign cloud, and a door to non-sovereign clouds for work that is less sensitive. TelecomTV, opened beside the release, quotes the same pitch and then asks whether other telcos will copy it. This is not a procurement note and not a trade.

TL;DR - RE:AI, the sovereign AI cloud unit inside Singtel's Digital InfraCo, launched AI Token-as-a-Service on October 1, 2026. TelecomTV calls Singtel one of the first telcos into this kind of billing, and says China's three main telcos are already ahead. - Customers pick a token volume and use it across queries, software development, content, voice and video, and agentic workflows. Named open-weight options include Qwen, Mistral, GLM, Kimi, and MiniMax, plus selected frontier models, plus a customer's own model. - Data for the sovereign path stays in Singapore, the release says. Less sensitive workloads can use non-sovereign AI clouds. TelecomTV notes China's three main telcos already sell AI token subscriptions.

What actually happened

The release is datelined Singapore, 1 October 2026. RE:AI launched TaaS as a fully managed service: access to AI models, scalable infrastructure, orchestration, and governance, billed as a flexible token subscription. TelecomTV quotes the operator calling it a fully managed service: models, scalable infrastructure, orchestration, and governance, through a flexible token subscription. The company page's own bullets did not load in this check, so this note does not repeat a "first in Singapore" label it could not read.

The problem paragraph is about cost and borders. As companies spread AI through their operations, token costs rise, and sending sensitive data to models hosted overseas gets in the way. Agentic AI, the release says, makes both problems worse, because agents burn more tokens and because several models are harder to manage. The proposed fix is hosting models on sovereign infrastructure in Singapore, on the RE:AI cloud, with AI data centers, terrestrial and subsea connectivity, and the Paragon orchestration platform.

Bill Chang, CEO of RE:AI and Digital InfraCo, is quoted twice in substance. The opportunity, he says, is scale: enterprises want one integrated way to deploy AI instead of buying and managing models separately. He says Singtel has been on its own path to become an AI-first organization, and that TaaS bundles the stack so customers get flexibility, visibility, and control, and can move from ambition to impact while strengthening Singapore as an AI hub. TelecomTV uses that scale quote and the same title for Chang. The wording matches. TelecomTV is not inventing a second CEO.

Subscriptions let a customer select the token volumes they need. The release lists internal queries, software development, and agentic AI, and it says agents can automate workflows, conduct market research, and review documents. Models-as-a-service means the customer does not have to buy and run each model. Intelligent routing "automatically matches workloads to the most suitable AI models to optimise cost and performance." The release uses the British spelling. Applications it names: software development, content creation, and voice and video. The open-weight names are Qwen, Mistral, GLM, Kimi, and MiniMax. Frontier models are "selected" and unnamed. Customers can onboard their own models.

Hosting on the sovereign cloud is supposed to keep data inside Singapore so sovereignty and regulatory requirements are easier to meet. The next sentence is the escape hatch: for less sensitive workloads, enterprises can use non-sovereign AI clouds to optimize cost and performance. Token-level visibility is the governance hook: policies, guardrails, and monitoring. The release also says the service shortens deployment because the customer does not have to keep refreshing a complex AI stack, and that TaaS complements an existing GPU-as-a-Service offer. A link points at re-ai.ai for more on the product. The release does not print a rate card.

What TelecomTV adds

TelecomTV frames the launch as a telco question: can a carrier bill AI token consumption instead of data or voice minutes? It says the jury is out, and that Singtel thinks yes. It places RE:AI as the sovereign cloud unit of Digital InfraCo. It says Singtel is not the first telco to chase revenue from AI tokens. China's three main telcos are ahead, and TelecomTV points at a separate note on China Telecom's AI token subscriptions. That China comparison is not in the Singtel release. It is the newsroom's context.

TelecomTV then quotes the release on flexible subscriptions, on models-as-a-service, and on routing to Qwen, Mistral, GLM, Kimi, MiniMax, and frontier models. It quotes the sovereignty pitch and the non-sovereign option for less sensitive work. It quotes Chang on scale. It ends by saying other telcos that already run AI factories and sell GPU-as-a-service will watch, and it asks whether "tokenomics" is a profitable model for carriers. The release does not ask that question. It asserts the product. The doubt is TelecomTV's, and it does not come with a revenue figure. Neither page prints one.

So the confirmed core of the Singtel token service is the October 1 launch, the token subscription, the named open-weight set, Singapore hosting for the sovereign path, and the non-sovereign option. The unconfirmed part, if you were hoping for a price, is the price. "Greater control over cost" is a claim about visibility. It is not a number.

What the pages do not show

Neither page shows a token price, a minimum commit, a latency number, or a customer logo. Neither page shows which frontier models are in the "selected" bucket. A reader who needs to know whether a specific lab's model is on the router will not get that from these two texts. Neither page shows how routing chooses "most suitable." The word is in the release. The method is not.

The Singtel token service is also not a consumer chatbot and not an open-weight drop you can download. The weights it names are other companies' models, reached as a service. Onboarding your own model is a sentence in the release, not a tutorial. GPU-as-a-Service is a sibling product the release says TaaS complements. The pages do not say the GPU product changed on October 1.

People comparing model launches can read the Grok 4.7 note or the MAI voice models note. Those are model posts with scores or prices. This is a telco wrapper around other models, and the wrapper's price is not on the page.

What to do as a reader (not a trade)

This is not investment advice and not a suggestion to buy Singtel shares or any token. If you are scoping the product, write down what the release actually promises: a token volume you select, routing across a named open-weight list plus unnamed frontier models, data in Singapore for the sovereign path, and a non-sovereign path for less sensitive work. Then write down what it does not promise: a public rate, a named frontier model, or a proof that token billing makes money. TelecomTV's question about profit is a question. Do not turn it into a finding.

Treat the Singtel token service as a product announcement, not a price list. If you cite Chinese telcos being earlier, or Singtel as one of the first telcos into token billing, cite TelecomTV. Do not upgrade that to "first in Singapore" unless you have the release bullet in front of you. If you cite Chang, the quote is on both pages. If someone asks what a token costs, the honest answer from these pages is that they do not say.

The pages are Singtel's release and TelecomTV's report. Where the newsroom adds China or a profit question, keep that attribution. Where the release is silent on price, leave the silence.