A New Strategy in Inference

On November 16, Sarvam officially entered the commercial inference arena with the launch of Token Factory. This platform hosts the company’s proprietary Sarvam 105B model alongside a curated selection of open-weight models from Google, Zhipu, Moonshot AI, and DeepSeek. The most striking element of this rollout is the company’s pricing strategy: Sarvam has positioned its flagship model at a lower price point for both input and output tokens than any of the third-party alternatives featured in its debut lineup.

Under the current introductory rates, which are set to remain in place through the end of 2026, Sarvam 105B is priced at ₹10 per million uncached input tokens and ₹25 per million output tokens. In comparison, the standard tier of DeepSeek V4.1 Flash is priced at ₹14 for input and ₹55 for output. While both services charge ₹0.25 for cached input tokens, DeepSeek’s interactive tier—designed for lower latency—jumps significantly to ₹25 for input and ₹105 for output. Customers are advised to verify current rates, as Sarvam’s general API documentation has shown fluctuations in recent weeks.

Shifting Market Priorities

This aggressive pricing marks a sharp pivot from the company’s strategy just months ago. In August, Sarvam had raised the input cost for its 105B model to ₹29.28 per million tokens. Furthermore, when Sashikumar Sreedharan joined as president in October, the firm was still offering the older DeepSeek V4 Flash at a lower price point than its own domestic model. By launching Token Factory, Sarvam is clearly attempting to incentivize developers to adopt its proprietary technology over foreign alternatives.

Infrastructure and Technical Edge

Token Factory, which builds upon the preview showcased at the July Epoch developer conference, provides access to eight distinct models via an OpenAI-compatible API. While standard developers utilize a serverless API, high-demand models like DeepSeek V4 Pro and Kimi K3.1 Pro are restricted to provisioned throughput at the organizational level.

The platform is powered by what Sarvam describes as India’s largest cluster of Nvidia B200 GPUs. To maximize efficiency, the architecture decouples the prefill and decode phases of inference, assigning them to separate GPU pools. Additionally, the system employs a three-tier prompt cache to handle repetitive context in agentic applications, which Sarvam claims allows for a 95.7% cache hit rate for DeepSeek V4.1 Flash workloads—though this figure is specific to that model’s traffic patterns and may not reflect universal performance.

The Sovereignty Debate

Cofounder Vivek Raghavan has championed the concept of "token sovereignty," aiming to shift the processing of AI tokens from overseas cloud providers to domestic infrastructure. However, the current catalog reveals a paradox: while the infrastructure is Indian, the majority of the models—including those from Zhipu, DeepSeek, and Moonshot AI—are Chinese. While running these models on local hardware addresses the physical location of inference, it does not necessarily resolve concerns regarding data residency, model provenance, or the safety of the underlying weights.

This strategic direction is particularly relevant given Sarvam’s recent memorandum of understanding with India’s defense establishment. By pricing its own 105B model competitively, Sarvam is better positioned to serve government and defense clients who prioritize domestic development. Yet, as the company expands its footprint, it faces the challenge of balancing its identity as a frontier AI research lab with the commoditized, high-volume demands of an inference provider.