Sarvam defends hosting Chinese AI models on Token Factory
What's the story
Sarvam AI, a key player in India's sovereign AI initiatives, has defended its decision to host foreign models on its new inference platform, Token Factory. The company clarified that the operations are powered by privately procured GPUs and not government-subsidized infrastructure. This comes after criticism over the inclusion of six Chinese-developed AI models on the platform.
Clarification
Sarvam clarifies GPU resource usage amid criticism
In response to the criticism, Sarvam co-founder Pratyush Kumar said, "The token factory is running on GPUs privately procured by Sarvam."
The company also clarified that it had previously used subsidized GPU access under the IndiaAI Mission for training its own models.
This program provided a 40% discount on GPU access, which was used to develop open-weight models later published on Hugging Face and AIKosh under permissive licenses.
Kumar added that Sarvam's access to these GPUs ended in March 2026.
Strategic approach
Sarvam compares foreign-developed models to international AI infrastructure providers
Defending its decision to offer foreign-developed models, Sarvam said open-weight AI models with capabilities approaching those of leading proprietary systems are now available at significantly lower costs.
The company highlighted their increasing adoption among global enterprises and compared them with international AI infrastructure providers offering inference services for open-weight models at scale.
Domestic solution
Sarvam emphasizes focus on building indigenous AI capabilities
Sarvam said its Token Factory aims to provide Indian enterprises with a domestic alternative to these foreign platforms through competitive rupee-denominated pricing and data sovereignty.
The company also emphasized that serving foreign-developed models won't dilute its focus on building indigenous AI capabilities.
"While we have begun serving open models through Sarvam Token Factory, the core mission of Sarvam remains to innovate and build our own full stack including our own models trained from scratch," the company said.