समाचार Samāchār · Tech & AI · 19 November 2026
Sarvam now serves DeepSeek from Indian GPUs. Sovereign AI needs the shop and the lab
Sarvam’s Token Factory runs its own 105B model alongside Chinese open-weight models on clusters in India. Critics ask if that fits a sovereign AI mission. Sarvam says its subsidised GPU access ended in March.
Sarvam has opened a second business, and it has started an argument. Its Token Factory, launched on 16 November, sells access to AI models from GPU clusters in India. This week the question of what it serves became a national debate.
The menu is the point. Alongside its own Sarvam 105B, the platform serves open-weight models including DeepSeek V4.1 Flash, DeepSeek V4 Pro, Kimi K3.1 Pro, GLM 5.3 and Google’s Gemma 4 31B. Most come from Chinese labs. Sarvam’s FAQ says requests run on Sarvam-operated GPUs in India and are not forwarded to the developer’s overseas API. Introductory pricing for DeepSeek V4.1 Flash is ₹14 per million input tokens and ₹55 per million output tokens. Sarvam says it serves more than 450 billion tokens of that model a day.
Critics asked whether a company backed by the IndiaAI Mission should be reselling foreign models. Sanchit Monga, CEO of RunAnywhere, wrote that Sarvam was “mostly serving Chinese models from India GPUs”. Forbes noted that the IndiaAI subsidy of ₹98.68 crore covered 4,096 H100 GPUs for training, not this inference stack. Co-founder Pratyush Kumar told BusinessLine that Sarvam’s subsidised access to those GPUs ended in March 2026, MediaNama reported.
Both sides have a point. Indian-hosted inference keeps prompts and company data on Indian soil. That matters for banks, hospitals and the state. Hosting is not building. Running a Chinese model in India does not tell us how it was trained or what it will decline to say.
Credit to Vivek Raghavan, Pratyush Kumar and the Sarvam team for inference that competes on price and speed. The test is whether Sarvam 105B wins the traffic on Sarvam’s own platform. Revenue from the shop can fund the lab, as long as the lab stays first.
What a founder takes from it: Sovereign AI has three layers: compute, models and data. India is building the first one fast. The second decides who sets the rules, so founders should build on Indian-hosted inference and back Indian models to win traffic on it.