Will India Get Its Own Answer Engine? Bhashini, Sarvam, and the Sovereign AI-Visibility Stack

The question, stated precisely

The phrase "India's own answer engine" can mean two very different things, and the difference matters for any business trying to plan around it. It can mean a consumer product, a chat box a shopper opens instead of ChatGPT or Google, that reads the Indian web and returns a synthesized recommendation. Or it can mean a stack, a set of government and private components, language models, compute, and commerce rails, that could power such a product whether or not any single branded app becomes the front door.

On the first meaning, the answer today is no. There is no India-built answer engine with the reach of the global ones. Sarvam has shipped a consumer chatbot, and Krutrim and others have their own assistants, but none functions as the default place Indians go to ask for a plumber, a coaching center, or a supplier. The global engines still do the reading and the choosing for most queries that reach an answer surface at all.

On the second meaning, the picture is more interesting. Every layer a sovereign answer stack would need is being built in public, with government money and stated intent behind it. That is why the question is worth taking seriously rather than dismissing. This article works through the layers one at a time, marks what is real against what is aspirational, and then turns to what a second answer layer would change for the small firm trying to be found.

Bhashini: the language layer

Bhashini is the part of the stack that already operates at national scale. Launched in July 2022 by the Ministry of Electronics and Information Technology, its name is a compression of "Bhasha Interface for India," and its job is to break the country's language barrier by exposing translation, speech, and text models through open APIs. It is run by the Digital India Bhashini Division under the Digital India Corporation.

The usage figures are not trivial. By March 2026 the platform had crossed six billion (600 crore) total requests, was processing on the order of fifteen million (1.5 crore) requests a day at sub-second response, and managed more than three hundred and fifty models serving over five hundred government websites. It supports text services in dozens of Indian languages and voice in roughly two dozen. In a notable convergence, the division has integrated Sarvam's open-source models into the platform, so the language layer and the model layer are already touching.

What Bhashini is not is an answer engine or a business directory. It translates and transcribes; it does not read the web, rank suppliers, or name a shop in response to a buyer's question. It is infrastructure, a way to move between languages, and its scale indicates that the language plumbing for an India-built discovery surface exists. It does not indicate that such a surface exists. Treating Bhashini's six billion requests as evidence that Indian MSMEs are already being discovered in-language would be a category error.

Sarvam: the sovereign model layer

If Bhashini is the language pipe, Sarvam is the clearest attempt at the model that would flow through it. Founded in August 2023 in Bengaluru by Vivek Raghavan and Pratyush Kumar, both from the AI4Bharat lab at IIT Madras, the company set out to build large models centered on Indian languages rather than adapted from English-first systems. It raised roughly forty-one million dollars in a combined seed and Series A round in December 2023, and by June 2026 had reached a reported valuation of about 1.5 billion dollars after a Series B led by HCLTech.

The government connection is the reason Sarvam belongs in a sovereignty discussion. In April 2025 it was selected under the IndiaAI Mission to build one of India's sovereign foundation models, a model meant to be trained, deployed, and optimized inside India, built for voice and fluent across Indian languages. Its model line has moved quickly: Sarvam-1 in October 2024, the 24-billion-parameter Sarvam-M in May 2025, and larger Sarvam-30B and Sarvam-105B models unveiled in early 2026, the largest of them reported at 105 billion parameters trained on trillions of tokens.

The strategic claim behind this work is not that a bigger model will beat the frontier labs at general intelligence. It is that a model built in India, on Indian languages and Indian governance, is better positioned for population-scale public deployment and for the vernacular queries a global model handles unevenly. Whether that translates into a discovery surface that reads the open web and names businesses is, again, unproven. A capable sovereign model is a necessary component of an India-built answer engine. It is not, by itself, the engine.

IndiaAI Mission: the compute and policy layer

Underneath both Bhashini and Sarvam sits the policy and compute layer that funds them. The IndiaAI Mission, approved by the Union Cabinet in March 2024 with an outlay of about 10,372 crore rupees (on the order of 1.1 to 1.25 billion dollars), is the government's framework for building an indigenous AI base. It is organized around several pillars that include domestic compute, indigenous foundation models, dataset access, and startup finance, and it selected a set of organizations and consortia, Sarvam among them, to build foundation models with subsidized compute.

The compute subsidy is the concrete lever. Reporting on the mission's first phase indicates Sarvam was granted access to thousands of Nvidia H100 GPUs for a fixed period to train its sovereign model, one of the larger single allocations in the program. The mission also stresses a sovereign-cloud posture, processing on infrastructure under Indian control rather than dependent on foreign hyperscalers. These are the ingredients of technological sovereignty: domestic compute, domestically built models, and national governance over the data that trains them.

It is worth being precise about what this layer does and does not settle. Public money and GPU allocations make an India-built model layer viable; they do not commit the state to building a consumer answer engine or a business directory on top of it. The mission's stated priorities lean toward public services, language access, and foundational capability. An answer surface that mediates commercial discovery for MSMEs is a plausible downstream product of this investment, not a declared deliverable of it.

ONDC: the discovery layer that already exists

The layer most often left out of the answer-engine conversation is the one that most directly touches MSME visibility: discovery. India already runs a government-backed, open discovery network for commerce. The Open Network for Digital Commerce, incorporated as a non-profit Section 8 company in December 2021 under the Department for Promotion of Industry and Internal Trade, is built on the open Beckn Protocol and is designed as a network rather than a platform, so buyers and sellers on different apps can find and transact with each other.

Its scale is already material. By late 2024 ONDC was reported handling on the order of fourteen million transactions a month, with roughly three hundred and seventy thousand sellers and service providers onboarded across more than eight hundred cities. And the language layer is wired in: in September 2024, ONDC and Bhashini jointly released Saarthi, a reference app that lets a shopper discover and buy across the network in Indian languages, starting with a handful and aiming at all of the country's scheduled languages.

This matters because ONDC is, in effect, a structured discovery graph of Indian sellers that the state controls, expressed in an open protocol, with catalogs and attributes an engine could read. If any India-built answer layer wanted a native source of business data to draw on, it would not have to scrape the open web the way global engines do; a machine-readable commerce network already exists. The counterpoint matters just as much: ONDC ranks and routes within its own network by its own protocol logic, not by synthesizing an answer to a natural-language question, and its reach is still a fraction of the broader Indian retail economy. It is a discovery layer, not an answer engine, but it is the layer where sovereignty and MSME visibility most concretely meet.

What a sovereign answer stack would mean for MSME visibility

Line the layers up and the shape of a possible stack is visible: Bhashini for language, Sarvam and the mission's other model-builders for reasoning, IndiaAI compute and sovereign cloud for the substrate, and ONDC for a native, structured source of who sells what. None of these was built to be an answer engine, and assembling them into one that reads the open web and names small businesses is a design choice nobody has publicly committed to. But the components exist, they are government-backed, and they touch each other already.

For a micro, small, or medium enterprise, the useful question is not whether this stack will be built, which no one can yet answer, but what it would change if it were. Three implications follow from the public record without requiring any speculation about specific products.

What the record actually supports

The temptation is to pick an extreme. One says India will imminently launch a sovereign answer engine that reshuffles who gets found, and small firms must scramble. The other says this is national-pride vaporware and nothing will change. The public record supports neither. It supports a middle reading: the individual layers are real, well-funded, and already interconnected, while their convergence into a consumer answer engine that names MSMEs is a possibility rather than a plan.

That middle reading has a clear operational takeaway, and it does not depend on predicting which future arrives. The properties that would make a business legible to an India-built answer layer, a consistent verified identity, an offering described in the languages its customers use, corroboration across independent sources, and a complete structured presence in the networks that matter, are the same properties that make it legible to the global engines operating in India right now. Machine readability is not a bet on one stack. It is the common requirement across all of them.

So the answer to the headline is: not yet, and not certainly, but the pieces are being laid, and the response that survives every branch of the fork is the same. A firm that measures where it actually stands across the surfaces that decide visibility today, and fixes what is unreadable, is prepared whichever engine, global or sovereign, ends up doing the choosing. That is a discipline, not a prediction, and it is the one worth adopting while the question stays open.