Subscribe
AI
3 min read

Sarvam's Token Factory: Six Chinese Models Out of Eight on Sovereign AI's Flagship

Sarvam AI launched Token Factory on November 16, an inference platform that serves eight AI models through a single OpenAI-compatible API, all hosted on GPU clusters inside India. Six of the eight models are Chinese. For the startup most closely associated with India's sovereign AI push, that ratio became the whole story. Bloomberg reported this week that the launch triggered online backlash, with critics arguing a company built on public support is monetizing other countries' research. The cat

Sarvam's Token Factory: Six Chinese Models Out of Eight on Sovereign AI's Flagship
Sarvam Token Factory sovereignty math

Sarvam AI launched Token Factory on November 16, an inference platform that serves eight AI models through a single OpenAI-compatible API, all hosted on GPU clusters inside India. Six of the eight models are Chinese. For the startup most closely associated with India's sovereign AI push, that ratio became the whole story.

Bloomberg reported this week that the launch triggered online backlash, with critics arguing a company built on public support is monetizing other countries' research. The catalogue pairs Sarvam's own 105B model with offerings from DeepSeek, Moonshot AI's Kimi, Z.ai's GLM and Google's Gemma. Pricing starts at 14 rupees per million input tokens for DeepSeek V4.1 Flash, and Sarvam offers zero-data-retention plans for enterprise customers.

Why the criticism stings

Sarvam is the flagship pick of the IndiaAI Mission, the government program meant to produce domestic frontier capability. The company received a subsidy worth 98.68 crore rupees covering 4,096 Nvidia H100 GPUs, and raised $234 million in June at a reported $1.5 billion valuation. Government money came with the expectation of Indian foundation models, not an Indian-hosted storefront for Chinese open weights.

The dispute has a prequel. In May 2025, Sarvam released Sarvam-M, a 24B open-weight hybrid built on top of Mistral Small, supporting 10 Indian languages. It drew 334 Hugging Face downloads in its first two days. Investor Deedy Das called the numbers "embarrassing" and compared them unfavorably to a model from two Korean college students that pulled roughly 200,000 downloads. BharatGen's Param-1, another government-backed model trained from scratch, managed 12 downloads on the national AIKosh repository at the time. India's sovereign models had an adoption problem before they had a hosting problem.

Lightspeed India partner Hemant Mohapatra, a Sarvam investor, pushed back on X, arguing critics underestimate how much international inference providers already earn hosting foreign models, including Chinese ones, for Indian customers. JarvisLabs, a smaller Indian provider, launched its own Token Factory about two months earlier and now claims to match or beat Sarvam on speed and latency across three open-weight models. Those are competitor benchmarks, but they raise the obvious question of how much of Sarvam's edge is technology versus subsidized capital.

Sarvam model count six of eight Chinese

The sovereignty question underneath

There is a defensible reading the backlash skips. Hosting DeepSeek on Indian GPUs does not make DeepSeek Indian, but it lets Indian enterprises run it on domestic infrastructure, keeps request data in India, and cuts dependence on overseas API providers. Sovereign AI can mean owning every layer, or it can mean controlling where your data gets processed and who can turn the compute off. India's early experience with domestic models suggests the market rewards the second definition: customers vote for the best available open-weight model at the lowest latency, and Chinese labs currently ship the best available open-weight models.

If Sarvam builds a profitable inference business, that cash flow could eventually fund its own frontier training run. If it becomes primarily a reseller of foreign weights, the original ambition of the IndiaAI Mission quietly becomes a hosting contract. The next few quarters, not the Twitter argument, will decide which one it is.

Sources

AIOpen WeightsInference