makhana mornings!

A little context. A lot to think about.

BUSINESS. PEOPLE. THE BIGGER PICTURE.

Clarity for a world that doesn’t slow down.

Sarvam’s Token Factory separates where an AI model runs from who built it

2 minutes

Sarvam’s India-hosted inference service combines its own and third-party models. For buyers, location, model choice and data-handling terms are separate questions.

Conceptual illustration of different AI models feeding a server in India and an application.

An AI model’s developer and the company running it for a customer do not have to be the same business. Sarvam’s Token Factory makes that distinction central to its proposition.

In a launch post dated November 16, the Indian AI company announced public access to a service that runs its own and third-party open-weight models on GPU clusters located in India. It offers a common OpenAI-compatible API, including access to models from the DeepSeek, Kimi, GLM and Gemma families alongside Sarvam’s own.

Local infrastructure, multiple model choices

Inference is the computing work that produces a model’s answer. Selling that service is different from developing the model itself. An India-hosted service can therefore offer models created elsewhere without turning them into Indian-developed models.

For a buyer, the useful separation is between the model selected for a task and the infrastructure used to run it. A common API can simplify the connection, but it does not make different models behave identically. An application still needs testing when its model changes.

What the launch does not establish

Sarvam publishes performance comparisons and pricing for the service. Its benchmarks are company-run tests, not independent proof that every customer workload will be faster or cheaper. Some comparisons use a client in Mumbai, defined prompt lengths and medians measured over 72 hours. Prices also vary by model and service tier, with introductory rates identified.

Those details matter because a quick answer to one test prompt is not the same as dependable performance under a business’s real workload. A sensible trial would include the actual documents, languages and concurrent requests the application needs to handle, rather than selecting a provider from one speed headline.

Sarvam’s catalogue says inference requests are not forwarded to the original developer’s overseas API, while external tools connected by an application have their own processing locations. It also says zero data retention is available on request under an agreed arrangement, with operational logs scoped separately. Hosting location therefore answers only one part of a buyer’s data-handling questions. It should not be treated as an automatic guarantee about retention, access, security or regulatory compliance. Those require their own checks against the service terms and the organisation’s requirements.

The business question

The offering gives customers another way to separate model choice from the work of operating the underlying computing infrastructure. Its practical value will depend on whether the chosen combination meets their quality, reliability and cost requirements.

That is a more useful reading than treating every locally hosted model as a locally developed one. Where the model runs is a purchase decision. Who built it is a different fact.

Illustration: original AI-generated editorial concept, not a depiction of Sarvam’s infrastructure or product interface.

Leave your thoughts

You might also like…