3 Sources
[1]
Sarvam AI launches Token Factory to offer cheaper AI model inference from India
Sarvam said Token Factory matches or outperforms leading global inference providers on speed and pricing across most of the models it serves. Sarvam AI has launched Token Factory, an inference stack that serves its own models and global open-weight AI models from GPU clusters located in India.
[2]
Sarvam AI Opens Token Factory To Run Global AI Models On Indian Servers
Token Factory charges by usage, with Sarvam's catalogue listing Sarvam 105B at ₹15 per Mn input tokens and ₹60 per Mn output tokens AI unicorn Sarvam AI has opened Token Factory to developers and enterprises, giving them access to its own and global AI models through a single application
[3]
Sarvam AI Opens Token Factory, Challenges OpenRouter in India
Sarvam AI has opened its Token Factory platform to developers and businesses, offering multiple AI models through one API with rupee-based pricing and computing infrastructure hosted in India. Indian artificial intelligence startup Sarvam AI has opened its Token Factory platform to developers and
Share
Copy Link
Sarvam AI has opened Token Factory, running its own and global open-weight models on GPUs in India. The Indian startup says it matches global providers on speed and cost.
Sarvam AI, an Indian AI startup that builds its own models, has opened Token Factory, a platform that runs those models and a list of global open-weight ones on GPU clusters located in India, all through a single OpenAI-compatible API
1
2
.
Source: Inc42
The pitch is that processing stays in India rather than being routed overseas. The company says requests run on its own GPUs rather than being routed to model developers' APIs abroad. It also offers dedicated GPU capacity plus on-premises deployment for institutions that cannot send data out at all
2
. On cost and speed, Sarvam says the platform matches or beats leading global inference providers for most of the models it hosts1
2
.The launch catalogue spans Sarvam's own 105B model plus seven named global open-weight models, including DeepSeek V4.1 Flash, Kimi K3.1 Pro and Google's Gemma 4 31B
1
2
. A wider list sits beyond that3
. Developers switch between them with one API, one software development kit and one credit balance; teams already on OpenAI's API can connect by changing the endpoint2
3
.Pricing is per token, in rupees. Inc42 lists Sarvam 105B at 15 rupees per million input tokens and 60 per million output tokens, with cached input at 5 rupees
2
. Analytics Insight reports introductory rates of 10 and 25 rupees per million tokens for the same model; the two outlets do not reconcile the difference3
.In the company's own tests, Sarvam 105B took 200 milliseconds to return its opening token and generated 330 tokens a second for one user
2
. The company also says it serves over 450 billion tokens a day on DeepSeek V4.1 Flash, and that 95.7% of prompt tokens come from a cache that skips recomputing repeated text2
. Those are the company's own figures, not independent benchmarks, and they are what the speed and cost claims rest on2
. Sarvam says it runs India's largest cluster of Nvidia B200 chips and is adding capacity, per Analytics Insight3
. The plan itself is not new: Sarvam previewed it at its Sarvam Epoch event earlier in the year and opened general availability this month2
.
Source: Analytics Insight
Related Stories
Sarvam is not the first to offer this. Per Inc42, Krutrim has hosted DeepSeek models on Indian servers since January 2025, and Anthropic serves select Claude models from within India through Amazon Bedrock
2
. Beyond India, Sarvam will compete with inference specialists Together AI, Fireworks AI and OpenRouter, plus Amazon Web Services, Microsoft Azure and Google Cloud2
. OpenAI began serving eligible GPT-6 Luna workloads from India on November 12, days before Token Factory's general availability2
.
Source: ET
Summarized by
Navi
[3]