New Delhi: Sarvam AI's Token Factory has allowed businesses to use China's DeepSeek and have their requests processed in India. Running DeepSeek alongside its own model on Indian servers raises a question over India's push for sovereign AI.
How much control does local hosting give businesses over models trained overseas? Token Factory connects a company's software to several AI models through a single interface, an API. Businesses can choose which model handles each request.
Sarvam offers its own 105B model alongside China's DeepSeek, Kimi and GLM, and Google's Gemma. The company says it runs the models on clusters of graphics processors in India and matches or beats leading global providers on speed and cost for most of the models it offers.
"A model's country of origin deserves scrutiny, particularly from banks and insurers. Businesses should ask whether they can audit and secure the model, control how their data passes through it and understand where it falls short," said Atul Arya, founder and CEO of AI services firm Blackstraw.
"If I'm honest, a model from a 'trusted' country that you can't inspect or govern isn't much safer either," Arya said, adding that running a model in India can help regulated businesses address data security and compliance requirements.
The discussion has also touched the use of publicly supported computing infrastructure and the role of foreign-developed models in India's AI ecosystem. Sarvam has said the platform runs on privately procured GPUs and does not use capacity subsidised under the IndiaAI Mission.
For a bank or an insurer weighing the switch, the practical test is narrower than the sovereignty argument. Can the model be audited, can its data path be controlled, and does it fail in ways the compliance team can see.