
Sarvam launches Token Factory for AI inference in India
The AMW Read
A model developer's local inference platform expands the infrastructure player map and highlights a sovereign AI policy tradeoff.
Sarvam launches Token Factory for AI inference in India
Sarvam launched Token Factory on November 16, giving developers and enterprises access to eight AI models through one OpenAI-compatible API. The models run on GPU infrastructure in India, so inference requests do not need to pass through overseas providers. The catalog includes Sarvam's own 105-billion-parameter model and models from DeepSeek, Kimi, GLM, and Google. Pricing starts at 14 Indian rupees per million input tokens for DeepSeek V4.1 Flash. Sarvam also offers zero-data-retention options for enterprises.
The launch places a foundation model developer in the market for inference services while reopening a debate about what sovereign AI means in India. Six of the eight listed models come from Chinese companies, according to Analytics India Magazine. Sarvam was selected under the IndiaAI Mission to build an indigenous foundation model and received a 986.8 million Indian rupee GPU subsidy for 4,096 Nvidia H100 GPUs. Hosting foreign models can give customers local processing without requiring them to choose an Indian-developed model. It also raises the question of how Sarvam will balance its infrastructure business with its model-development ambitions.
For builders, the practical proposition is access to several open-weight models through one API, with inference kept in India and an optional zero-data-retention arrangement. They will still need to assess which model and retention terms fit each workload. For investors, the distinction is whether local inference demand can support a business alongside the costly work of developing a domestic model. The launch alone does not establish customer uptake or show which business will take priority.


