LowV Technology Geopolitics & AI18 November 2026, Wednesday
India's sovereign AI champion Sarvam mostly serves Chinese models
Sarvam, which the government picked to build a foundation model, has launched Token Factory, an inference platform hosted in India. According to Analytics India Magazine, six of the eight models on the platform are Chinese.
Token Factory lets developers run open-weight models on GPUs in India through a single interface. The line-up includes the DeepSeek, Kimi and GLM families, a Google Gemma model and Sarvam's own 105B model. Entry pricing is 14 rupees per million input tokens and 55 rupees per million output tokens. Sources disagree on the launch date: Analytics India Magazine says 16 November, MediaNama 18 November.
The criticism centres on subsidies. Analytics India Magazine reports that Sarvam received 98.68 crore rupees in GPU support for 4,096 Nvidia H100s. The company raised 234 million dollars at a 1.5 billion dollar valuation in June. Sarvam says the platform runs on privately purchased GPUs without subsidy. According to MediaNama, the 40% GPU discount under the IndiaAI Mission expired in March 2026.
The constraint is that sovereignty is split across two layers. The hardware depends on American chips, the model layer largely on China's open-weight models. Sarvam says requests never reach the servers of foreign model developers, so the data stays in India. Its own models have drawn limited interest: Sarvam-M was downloaded 334 times on Hugging Face in its first two days.
Talay assessment
Bottom line
Token Factory shows India's sovereign AI ambition advancing at the data layer while leaning on China at the model layer. India can keep data onshore but cannot yet build competitive models of its own. As the US–China technology split hardens, that gap leaves New Delhi dependent on both sides at once. The most likely path is continued debate without the government restricting Chinese models in public procurement.
Likely effects
- Model supply riskNegative1–6 months
Most models hosted in India are Chinese-built; if New Delhi restricts their public use, the domestic inference market shrinks.
- Data localisationPositive1–6 months
Keeping requests onshore meets data localisation goals; companies may shift from foreign APIs to domestic hosting.
- Public supportUncertain1–6 months
The row over 98.68 crore rupees in GPU support adds pressure to tie future IndiaAI Mission allocations to performance conditions.
Possibilities, ranked
- 1Debate continues, no rules60%
The government imposes no formal limit on Chinese models and the platform grows commercially.
Watch: MeitY or IndiaAI refraining from guidance on foreign open-weight models
- 2Limits in public procurement25%
The government restricts Chinese-origin models in public-sector and defence use.
Watch: Model-origin clauses appearing in public tender specifications
- 3Domestic model breakthrough15%
Sarvam or another Indian model overtakes the Chinese models in usage.
Watch: Company data on Sarvam 105B's share of usage on Token Factory
Probabilities are calibrated judgement based on the sources, not measurement, and are revised as new information arrives. Not investment advice.
Market reaction
Indicators affected
- Chinese models on the platform▲ 6 / 8
- GPU support received by Sarvam▲ 98.68 crore rupees
- Entry price (per million input tokens)▼ 14 rupees