Sarvam AI's Indus Wins Maharashtra, Putting Sovereign AI to Work for 2,500 Officials

Sarvam AI has signed a formal partnership with the Government of Maharashtra to deploy its Indus sovereign AI workspace across state departments. Around 2,500 officials will use the platform in their daily work, making this one of the largest government AI rollouts by an Indian-built model anywhere in the country.

What Indus actually does on the job

Indus is not a chatbot bolted onto a government portal. The platform provides drafting and file-noting assistance, RTI response generation, document intelligence, optical character recognition across Indian scripts, multilingual interaction in 22 languages including Marathi, project workspaces, data analysis, and integration with eOffice and NIC email wherever API access is available. In plain terms: an official can drop a scanned Marathi document into Indus, get it converted to searchable text, summarise it, and draft a reply, all without switching tools.

Beyond Indus, the package also includes Samvaad, voice and WhatsApp-based citizen outreach agents, and Saaras and Bulbul speech models for integration into the state's MahaAI platform. The full suite covers the bureaucratic workflow from internal file management all the way to citizen-facing voice calls.

The numbers behind the deal

The package covers the deployment of Indus for around 2,500 state government users at an indicative first-year cost of ₹5.63 crore and a two-year outlay of ₹11.26 crore. That works out to roughly $1.35 million over two years, a modest figure by enterprise software standards, but the pricing structure tells the real story.

• A Foundation Partner subscription rate of ₹1,200 per user per month for the first 12 months, versus the standard rate of ₹2,500 per user per month.
• A one-time professional services fee of ₹50 lakh, plus setup and adoption support costing ₹42 lakh for three months.
• API usage charges of ₹0.50 per minute for Saaras and ₹30 per 10,000 characters for Bulbul, excluding GST.
• A one-month pilot of the Samvaad AI agent for tuberculosis awareness, with a ₹6 lakh agent development charge and ₹3 per minute usage.

The discounted entry pricing is a classic land-and-expand play. Sarvam gets a reference deployment at scale; Maharashtra gets a heavily subsidised first year to prove value before committing to full rates.

Who built this, and why it matters that they did

Sarvam AI was founded in August 2023 by Vivek Raghavan and Pratyush Kumar, who were previously associated with AI4Bharat at IIT Madras. The launch follows the company's selection under the Government of India's IndiaAI Mission to build a home-grown large language model. That government mandate gave Sarvam access to serious compute: 4,096 NVIDIA H100 SXM GPUs via Yotta Data Services, roughly ₹99 crore in GPU subsidies under the IndiaAI Mission.

In February 2026, Sarvam open-sourced two large language models: Sarvam-30B and Sarvam-105B, the latter using a Mixture-of-Experts architecture with 10.3 billion active parameters. Mixture-of-Experts (MoE) is an architecture where only a fraction of the model's total parameters are activated for any given input, making a 105-billion-parameter model run with the compute cost closer to a 10-billion-parameter one. Both models were trained from scratch, not fine-tuned from English models, on sovereign Indian data.

The company has since become a unicorn. Sarvam raised $234 million in the first close of its $300 million Series B at a post-money valuation of $1.5 billion, with HCLTech and Bessemer Venture Partners investing, alongside existing backers Khosla Ventures and Peak XV Partners.

The real story: sovereign AI is no longer a slogan

The word "sovereign" has been thrown around Indian AI policy circles for years. The Maharashtra deployment is one of the clearest examples yet of what it means in practice. Everything runs on secure, audited infrastructure located entirely within India. Government data stays in the country, on systems the state can inspect. The proposal was cleared by a High-Power Committee headed by the Chief Secretary, signalling that this is a state-level policy commitment, not a departmental experiment.

The global LLM giants, OpenAI, Google, Anthropic, Meta, are excellent at English and adequate at a handful of European languages. For a farmer in Maharashtra asking a government AI about crop subsidies in Marathi, these models are essentially useless. That gap is exactly the market Sarvam is filling, and the Maharashtra deal is proof it can close contracts at the state level on that basis.

There is a broader geopolitical current here too. Sarvam's fundraising comes amid rising concerns over access to advanced AI technologies. Recent U.S. restrictions on certain advanced AI models have intensified discussions around AI sovereignty globally, reinforcing India's ambition to build domestically controlled AI infrastructure. Nations are building domestic AI compute capacity not because it is the most economically efficient way to access AI capability, but because they have concluded that relying on foreign-owned infrastructure for AI workloads that touch sensitive government data represents an unacceptable strategic dependency.

Who wins, who watches nervously

The clearest winner is Sarvam itself. A live deployment across 2,500 officials in India's second most populous state is the kind of reference that opens doors in every other state government. Sarvam has already signed up with HCL Technologies, the Government of Odisha, and the Institute of Chartered Accountants of India, suggesting the Maharashtra deal is part of a deliberate pattern of institutional anchors rather than a one-off win.

The more uncomfortable position belongs to global AI providers. Microsoft, Google, and Amazon all offer in-country data processing options for India, but there is no legal standard or technical requirement to be classified as sovereign, so hyperscalers and Indian companies self-certify. Sarvam's pitch is that its models were built in India, trained on Indian data, and run on Indian infrastructure, a claim the global players structurally cannot match at the model layer.

What comes next

The Indus rollout is the first step in a wider programme of work between Sarvam and the state, spanning voice agents for citizen outreach and speech technologies for public services. The TB awareness pilot using Samvaad is the first concrete step in that direction, a small but telling signal of where this goes: AI that calls citizens in their own language, not just assists officials at a desk.

Sarvam's conversational AI platform now processes more than 2 million interactions daily, with usage reportedly doubling over the previous two months. If the Maharashtra deployment adds meaningfully to that number, it will also strengthen the feedback loop that makes Indian-language models better over time. The data generated by 2,500 officials working in Marathi, across real government documents, is training signal that no synthetic dataset can replicate.

For teams building in the Indian public sector or enterprise space, the practical takeaway is that Indus is now a production-grade option with a live government reference. Sarvam's API platform exposes the same speech, translation, and document intelligence capabilities that underpin the Maharashtra deployment, available to developers today. The sovereign AI stack India has been promising for years is, at least in Maharashtra, actually running.