深潮TechFlow|11月 19, 2026 04:39
[Sarvam Simultaneously Pursues Sovereign AI While Reselling DeepSeek Models] According to a Bloomberg report on November 19th, Indian AI company Sarvam is simultaneously pursuing its "sovereign AI" goals and selling access to Chinese models through its newly launched inference platform, sparking debate about the autonomy of Indian AI.
This week, Sarvam opened its Token Factory to developers and enterprises, providing a batch of open-source weighted models on GPUs within India, including DeepSeek V4.1 Flash, Kimi K3.1 Pro from Dark Side of the Moon, GLM 5.3 from Z.ai, Alibaba Qwen 3.8 27B, Google Gemma, and its own Sarvam 105B; some are available on demand, while others require dedicated computing power. The company claims that DeepSeek V4.1 Flash processes over 450 billion tokens daily, with input prices starting at 14 rupees per million tokens.
The controversy lies in the overlapping identities: Sarvam is a beneficiary of the Indian government's IndiaAI Mission, having previously obtained domestic GPUs at a 40% discount for training its own models. Co-founder Pratyush Kumar responded that Token Factory runs on GPUs purchased by the company itself, that the subsidized computing power ended in March 2026, and that the open-source models generated from the training have been released under license on HuggingFace and AIKosh. Analysts are divided: some believe inference is a viable business for Indian companies, while others argue that hosting foreign models on Indian infrastructure is only "partial sovereignty," and that related revenue should be used to support domestic model development.
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink