Sarvam Opened a Token Factory. The Point Is Where the Racks Are.

Sarvam AI has opened its Token Factory to outside developers and enterprises, turning what was an internal inference stack into a commercial service. The pitch is straightforward: an OpenAI-compatible API, billed in rupees, running on GPU clusters physically located in India. The models on offer include Sarvam’s own 105-billion-parameter system alongside Gemma 4 31B, DeepSeek V4.1 Flash, GLM 5.3 and Kimi K3.1 Pro.
The company says the service reached general availability earlier in November and was presented at its Sarvam Epoch event. For Sarvam 105B, it is charging ₹15 per million input tokens, ₹60 per million output tokens and ₹5 per million cached input tokens. It claims a 200-millisecond first token and 330 tokens per second for a single user.
What a token factory is
Most people meet artificial intelligence through a chat box. Most businesses meet it through an invoice. A token is the unit of that invoice: roughly three-quarters of a word, produced by a model that has to be run on a GPU somewhere. A token factory is the somewhere, packaged as a utility. You send text, you get text back, and you pay per million tokens.
The interesting part is not the API shape, which is deliberately identical to OpenAI’s so that code can be moved across with a URL change. It is the physical location. Every request served from a rack in India is a request that does not leave the country, does not cross a submarine cable and does not depend on a foreign company’s capacity allocation in a foreign data centre.
Sarvam says one of its hosted models, DeepSeek V4.1 Flash, is already handling about 450 billion tokens a day, with 95.7 per cent of prompt tokens served from cache. That cache number is the least glamorous and most important figure in the announcement, because caching is how inference gets cheap enough to sell.
The model can be American or Chinese. The electricity, the racks, the latency and the invoice can be Indian. That is the entire argument.
The Crescent World
Why hosting foreign models is the point
Three of the five models on the platform are not Sarvam’s. That looks odd for a company that raised money on the promise of building Indian foundation models. It is actually the commercially sensible move. Enterprises do not want one model; they want a menu, a price list, a data residency guarantee and a single contract. A platform that offers only its own model loses every deal where a customer wants a cheaper small model for classification and a large one for reasoning.
It also positions Sarvam as the neutral layer. If a bank in Mumbai wants to run a Chinese open-weight model, an American one and an Indian one in the same workflow, with logs that stay inside India and a rupee invoice that its finance team can process, the alternative today is to negotiate separately with three providers and route through Singapore. That is a real friction, and friction is where platforms make money.
The competition arrived first
Krutrim opened an Indian inference platform in January 2025. OpenAI began offering India-based inference for GPT-6 Luna on 12 November. Anthropic’s models are available through Amazon Bedrock regions that Indian enterprises already use. The gap Sarvam is attacking is not raw capability; it is price, latency and paperwork.
On price, ₹15 per million input tokens is aggressive but not astonishing; the open-weight models it hosts are available far cheaper from global aggregators, and some are effectively free to self-host. On latency, 200 milliseconds to first token is competitive for a domestic deployment and unbeatable for anything crossing an ocean twice. On paperwork, an Indian entity, Indian data residency and a rupee invoice are advantages that no foreign provider can fully replicate without building here.
The economics nobody advertises
Inference is a utilisation business. A GPU cluster costs the same whether it is busy or idle, so the marginal cost of a token collapses once a rack is full and the marginal loss is brutal when it is not. That is why providers chase high-volume, boring workloads: customer support summaries, document extraction, code completion, call centre transcription. Those jobs do not need the smartest model. They need a cheap one that never goes down.
Caching is the lever. If 95 per cent of the prompt tokens in a repeated workflow can be served from memory rather than recomputed, the effective price of a long-context request falls by an order of magnitude. Sarvam is advertising that ratio because it is the difference between a demo and a product for any customer with a 40-page contract to read a thousand times a day.
What this means for Indian builders
The practical effect is that a startup in Bengaluru can now build an application without a foreign credit card, without a data transfer agreement and without a dollar exposure. That sounds administrative. It is actually the difference between shipping and not shipping for a large number of small teams, especially in regulated sectors where the compliance officer has veto power.
What to watch
Whether the 450-billion-token figure is mostly internal traffic or paying customers. Whether the company can keep a menu of five models current as their upstream versions change, which is an operational grind rather than a research problem. Whether the price holds once a competitor with cheaper power and more GPUs decides to buy market share. And whether Indian enterprises actually move workloads, or sign pilot agreements and keep sending their data abroad.
India’s technology ambitions have mostly been told through applications: payments, identity, commerce. The harder layer is the one underneath, where electricity becomes computation and computation becomes an invoice. A token factory is the first Indian attempt to own that layer at commercial scale. It will be judged on utilisation, not on launch day.
Illustration generated with AI (Z-Image Turbo, rendered locally) for The Crescent World.



