Indosat and F5 launch Intelligence Suite for enterprise AI

Indosat and F5 launched Intelligence Suite on October 9, giving businesses a common interface for AI models, routing and usage accounting. The announcement targets expansion across Asia from Indosat’s Indonesian infrastructure. Fifty organizations are targeted at launch. That figure is a recruitment goal, rather than a disclosed count of customers already using the service.

Availability and model access

The immediate offering is the AI Token Marketplace, which the official website marks as live. It advertises access through an API compatible with OpenAI, with authentication, usage limits and spending controls. The site says more than 100 models at launch and more than 300 on the roadmap. That is narrower than the release’s claim of more than 300 models in one interface.

Even the smaller figure needs care. The public catalog listed 55 entries during our check on October 9. That list may differ from the approved catalog available to a customer, so buyers should verify the exact model names and versions before integration. The same page labels its dashboard and rate examples as illustrative screens with sample values. Those examples are not a usable price quote.

The broader AI Factory Platform is scheduled for November 2026. Indosat’s product page places private model serving, managed agents, retrieval from company documents, guardrails and model adaptation under that coming product, with an early access list. The token factory branding therefore spans a live marketplace and a planned set of application services. Teams seeking a hosted agent workflow should confirm its delivery date separately from endpoint access.

Pricing and data handling

The pricing page describes two commercial routes. Shared access is billed monthly by token without a commitment, while a dedicated private instance requires a volume commitment and offers reserved capacity, managed operations and named support. It gives no numeric token rates or minimum commitment. Its residency language is conditional, covering Sahabat AI and models hosted by Indosat locally while putting global model access behind routing policy.

For a buyer, the practical question is where each request and its associated records go. Confirm the approved processing location, the route used when a model is unavailable, retention of prompts and logs, and which people can change those rules. A regional sales ambition does not establish an available local hosting region in every Asian country. A single interface also does not make the contracts and data practices of different model providers interchangeable.

Performance and application testing

There is a concrete infrastructure proposition underneath the packaging. Indosat’s architecture description puts the gateway, caching, model routing, metering and compute under its operation, leaving the customer responsible for the application. It describes returning an exact cached response without another model call. That can avoid repeating the same computation, but a business still needs to measure its own cache hit rate and check whether reuse is appropriate for changing or sensitive information.

F5 brings Kubernetes traffic management and AI security capabilities to the partnership. Its load balancing documentation helps explain the performance goal. An analyzer observes queue depth, GPU memory use, inference latency, thermal conditions and errors, then changes how requests are distributed among serving systems. Sending work to an available server is a different decision from choosing the model best suited to the question. These controls can address infrastructure bottlenecks, while answer quality still needs separate evaluation.

Local language capability is another reason to examine the catalog closely. The existing Sahabat AI version 2 model card describes a 70 billion parameter model supporting English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese. GoTo and Indosat initiated the effort, with development credited to GoTo and AI Singapore. The model uses the Llama 3.1 Community License and its authors advise additional safety work. This is useful background for testing Indonesian workflows, although the exact model version served by the marketplace still needs confirmation.

Security controls also need application testing. OWASP’s prompt injection guidance explains how instructions hidden in retrieved documents or other outside content can redirect a model. It recommends limiting privileges, validating outputs, separating untrusted material and requiring human approval for consequential operations. A gateway that controls access and spending can be useful without settling whether an application should let a model send messages, modify records or access confidential documents.

The useful starting point is a small, representative workload with clear success criteria. Compare answer quality, full task cost, latency and the recorded route across the models actually available to the account. Include failed requests and fallback behavior in that test. For an enterprise choosing a managed platform, the value will come from reducing operational work while preserving control over those outcomes. The public launch material provides an initial scope, with several commercial and deployment details still needing a customer specific answer.

Archive photo of Indosat CEO Vikram Sinha with US Ambassador Kamala Shirin Lakhdhir in December 2024. Photo by Budi Sudarmo for the US Embassy Jakarta. Public domain via Wikimedia Commons.