Bengaluru-based AI startup Sarvam AI has launched Token Factory, an inference platform allowing developers to access global models via local Indian infrastructure. MediaNama reported that the service, launched on November 18, faces backlash for including six Chinese AI models alongside Google’s Gemma and Sarvam’s own models. Critics argue this contradicts the company’s indigenous ambitions, especially given recent military contracts. The platform offers introductory pricing and aims to provide a domestic alternative to foreign inference services.

Sanchit Monga, CEO of US-based RunAnywhere, criticized Sarvam for "mostly serving Chinese models from Indian GPUs". He noted that Sarvam received a Rs 98.68 crore subsidy for 4,096 H100s in April 2025. "Now they’ve opened Token Factory to everyone," Monga stated, adding that "8 models on it" include only one from Sarvam. He argued that after such significant capital and subsidized compute, the current output does not represent frontier research. Monga described the venture as a hosting business rather than a lab shipping indigenous research.

Sarvam AI co-founder Pratyush Kumar clarified that Token Factory runs on privately procured GPUs, not government subsidies. Kumar told The Hindu businessline that subsidized GPU access through the IndiaAI Mission ended in March 2026. He explained that the subsidy was strictly for training open-weight models published on HuggingFace and AIKosh. Kumar emphasized that inference requests are not forwarded to overseas APIs, ensuring data sovereignty and rupee pricing for users.

Industry observers offered mixed reactions. MediaNama founder Nikhil Pahwa argued Sarvam must compete in the open market, noting that government contracts define obligations, not public sentiment. He stated that nationalism secured the GPUs, but the company is not expected to deliver beyond its contract. Lightspeed India partner Hemant Mohapatra highlighted that global companies generate revenue running models on Indian infrastructure. He suggested Token Factory capitalizes on a global GPU shortage, rather than abandoning research goals.

The controversy follows a deal with the Headquarters Integrated Defence Staff for military AI applications. Srikanth Lakshmanan of CashlessConsumer warned that Sarvam’s role risks centralizing critical infrastructure. He described the company as a de facto national inference corporation, advocating for decentralized growth. Despite the criticism, Sarvam maintains the platform serves as a domestic alternative, offering faster access to leading open-weight models through a single API.