Skip to content
MukhyamantriMukhyamantri
TechnologyKarnataka

Six of eight models on Sarvam's India stack are Chinese. The AI mission row is on

Sarvam AI's Token Factory, launched on November 16, serves eight models from Indian GPUs through one API. Six are Chinese, which has triggered a debate on what sovereign AI means and whether public backing is being used to resell foreign models.

· · 1 min read

Bengaluru: Sarvam AI, picked under the IndiaAI Mission to build India's own foundation models, launched an inference platform called Token Factory on November 16. The argument since has been about what sovereign AI is supposed to mean.

Token Factory gives developers and enterprises access to eight models through a single OpenAI-compatible API. Sarvam says it serves them from GPU clusters in India.

The catalogue: Sarvam 105B, DeepSeek V4.1 Flash and V4 Pro, Kimi K3.1 Pro, GLM 5.3 and Gemma 4 31B. They come from DeepSeek, Moonshot AI, Z.ai and Alibaba, along with Google's Gemma and Sarvam's own 105-billion-parameter model. Six of the eight are Chinese, a point critics have seized on.

Pricing runs by usage. Sarvam 105B costs ₹15 per million input tokens and ₹60 per million output tokens.

The criticism has two strands. First, why is a company with a national mission mandate reselling foreign models, several of them Chinese? Second, is any public money flowing into a commercial platform?

Sarvam says Token Factory does not use government-subsidised GPUs, that it runs on its own infrastructure in India, and that customer prompts are not sent to the overseas developers. The company says its core mission remains building a full stack, including models trained from scratch, and that Token Factory gives Indian enterprises rupee pricing and domestic data handling.

Supporters make a practical case. Plenty of Indian firms want to use strong open models such as DeepSeek or Kimi but do not have the GPUs or the engineers to run them. Serving those models inside India keeps the data here and the bill in rupees.

The conceptual objection survives that. Running a model in India is not the same as building one in India. Where inference happens answers a data-location question. It does not answer who designed the model, what it was trained on, or what bias it carries.

Building frontier models from scratch is expensive, and inference revenue is one way to pay for it. Whether that trade is acceptable is now a policy question, not just a company one.