Sarvam has begun selling access to DeepSeek and other Chinese AI models through a new inference platform, drawing an online backlash against a startup that has been a leading advocate of India’s push for AI independence.

The Bangalore-based company this week launched Token Factory, which offers multiple AI models through a single API and runs on locally hosted computing infrastructure. Its catalog includes Sarvam’s own model as well as open-weight systems from Chinese labs DeepSeek, Moonshot AI and Z.AI Co.

The criticism stems from Sarvam’s role in India’s AI effort. It built its frontier model with backing from the government’s IndiaAI Mission and was among the first local startups to receive subsidized compute. The company was valued at $1.5 billion in April and has become a prominent example of India’s ambition to build AI services tailored to its languages and culture, a priority for Prime Minister Narendra Modi’s government.

Co-founder Pratyush Kumar defended the strategy after users questioned the company’s mission. “We haven’t changed our motto. We continue to build AI models for India. We will be a full stack AI company,” he said in an interview. He described the commercial logic as straightforward: selling AI tokens generates revenue and makes the capital-intensive model-building business more sustainable.

Sarvam said in a blog post that it already serves more than 450 billion tokens a day on the DeepSeek V4.1 Flash model. The startup is backed by Lightspeed and billionaire Vinod Khosla’s Khosla Ventures.

The founders emphasized that the infrastructure hosting Token Factory is separate and runs on computing power financed by investors. “Those chips are entirely funded by private money,” co-founder Vivek Raghavan said in the same interview.

Raghavan also addressed data concerns. He said that whenever a model, open or proprietary, is hosted on outside servers, data exfiltration is a possibility. Raghavan said the offering ensures that Indian AI users’ data stays within India’s borders. Without such local infrastructure, he argued, the economics for countries like India would leave users paying for tokens and then “double-paying” with data amassed by AI models.

The dispute highlights the tension facing Indian AI companies as they try to fund costly model development while maintaining a nationalist pitch. Sarvam says the Chinese models are part of a broader catalog and that its commitment to building AI for India remains unchanged.

Author