This is a MapMyTimes opinion piece. On 19 November 2026, Bengaluru-based Sarvam AI found itself at the centre of an online backlash after reports, first carried by Bloomberg, that its new Token Factory inference platform sells access to Chinese open-weight AI models, including DeepSeek's, alongside its own. MapMyTimes reported the row earlier today. The anger is understandable. But the more useful question is not whether Sarvam should host Chinese models; it is what India actually means when it says "sovereign AI".

What Sarvam did

Token Factory lets startups, enterprises and developers run AI models on Sarvam's computing infrastructure in India through a single interface, according to MediaNama. Its catalogue lists 12 models, six of them Chinese, from developers including DeepSeek, Moonshot AI and Z.AI, next to Google's Gemma and Sarvam's own Sarvam 105B, the outlet reported. Sarvam announced the product on its company blog.

The company has given three answers to its critics. Co-founder Vivek Raghavan said the platform runs on chips "entirely funded by private money", according to the Bloomberg report carried by Deccan Chronicle. Sarvam's FAQ states that inference runs on Sarvam-operated GPUs in India and that requests are not forwarded to the original developers' overseas servers, MediaNama reported. And co-founder Pratyush Kumar said: "We haven't changed our motto. We continue to build AI models for India."

The strongest case against Sarvam

Critics deserve a fair hearing, because their argument is not simply about China. Sarvam was among the first Indian startups to receive subsidised computing under the government's IndiaAI Mission. Kumar himself has said, as quoted by MediaNama from The Hindu BusinessLine, that the company got GPU access at a 40 per cent discount to train models, and that this access ended in March 2026. When public support goes to a company in the name of national capability, taxpayers can reasonably expect that capability to be built, not rented out.

Sanchit Monga, chief executive of US-based RunAnywhere, put it bluntly, calling the platform "a hosting business" and saying he had hoped Sarvam would be "the lab that ships frontier research out of India". Srikanth Lakshmanan, founder of CashlessConsumer, raised a different worry, that a single company acting as an inference provider risks concentrating critical infrastructure, MediaNama reported. These are serious points about public money and market power, and they should not be dismissed as online noise.

Why the business logic still holds

Yet there is a fair case on the other side. Building foundation models is expensive, and Kumar has argued that selling tokens is how the company sustains that work. Hemant Mohapatra of Lightspeed India, an investor in Sarvam, described the launch as a timely move amid a global shortage of GPUs. MediaNama founder Nikhil Pahwa noted that Sarvam's obligations are set by its government contracts and that running an inference business does not mean abandoning model research.

There is also a sovereignty argument that favours Token Factory. Raghavan's point is that any model hosted abroad carries a risk of data leaving the country. If Indian developers are going to use open-weight models anyway, running them on Indian servers, priced in rupees, arguably gives India more control, not less. Open weights can be inspected and run independently of the company that released them.

What should happen now

In our view, both sides are partly right, and the gap between them can be closed with clarity rather than outrage. Three steps would help.

  • Clear separation. Sarvam should publish, in plain terms, how its subsidised training work under the IndiaAI Mission is kept separate from its commercial hosting business, so that the "private money" claim can be checked rather than taken on trust.
  • Model provenance. Every model on platforms like Token Factory should carry a clear label of who built it, where, and under what licence, so that businesses and government buyers can make informed choices.
  • Defined expectations. The IndiaAI Mission should spell out publicly what outcomes it expects from companies that receive subsidised compute, so that debates like this one are judged against agreed benchmarks.

Sovereign AI should not mean refusing every foreign tool. It should mean that India controls where its data sits, understands what runs on its infrastructure, and keeps building its own capability. Sarvam's job now is to show, with evidence, that hosting other people's models funds that last goal rather than replacing it.

Photo source · Carl Lender, CC BY 2.0