Sarvam’s Token Factory is finally live. The new offering will let enterprises access Indian and global AI models through a single API while running workloads on GPUs in India. With the unicorn making a bold play for the local AI inference market, can it outpace global giants?
Unified Models Access: Token Factory hosts Sarvam’s own 105B model alongside overseas open-weight models from Google, DeepSeek, GLM and Kimi. Developers can switch between models without setting up their own compute. Leaning on three-tier caching architecture, Sarvam is claiming latency and throughput speeds that rival global inference leaders.
The Sovereign Pitch: Beyond raw speed and cost, Token Factory’s central pitch rests on sovereign infrastructure. Because every model runs directly on Sarvam’s India-based GPU clusters, the AI unicorn aims to woo enterprises that have to satisfy strict data residency mandates. For the BFSI customers, the company is also offering dedicated GPU capacity and on-premises inference deployments.
A Crowded Arena: However, Sarvam is not alone in offering local inference. Krutrim has been hosting DeepSeek models on Indian servers since last year, while OpenAI recently introduced India-based API inference for GPT-6 Luna workloads. Anthropic also offers in-country inference for select Claude models through Amazon Bedrock. As such, the jury is still out on whether Sarvam’s local inference playbook is coming too little and too late.
Questions Linger: Critics have also flagged that Token Factory hosts overseas models despite Sarvam positioning itself as a champion of sovereign Indian AI. While Sarvam says hosting global models locally supports data residency, questions remain over whether reliance on foreign open-weight models dilutes its push for homegrown AI capability.
Source link







