Startup Deep Dive : Sarvam AI — from $41m seed to India’s sovereign AI bet

In December 2023, a five-month-old Bengaluru startup called Sarvam AI raised $41 million (₹393 crore at today’s rate) — a rounding error next to the billions OpenAI and Anthropic raise in a single quarter. Sixteen months later, the Indian government picked that same thinly funded startup, over 66 other applicants, to build the country’s sovereign large language model from scratch.

Then, within weeks of getting that mandate, Sarvam’s first big public model landed with a thud, criticised as a rebadged version of someone else’s technology. It took the company roughly nine months of silent rebuilding to answer that criticism and turn state trust into a $1.5 billion valuation. This is the story of how a research pivot, a public stumble and a government compute grant compounded into India’s newest AI unicorn.

Quick facts

What they do

Sarvam AI builds foundation models tuned for Indian languages and sells access to them three ways: as a pay-per-use API for developers, as licensed enterprise deployments for regulated industries such as banking, insurance and healthcare, and as government infrastructure for public-sector bodies. Its product line spans large language models (Sarvam-1, Sarvam-M, and the newer Sarvam-30B and Sarvam-105B), speech-to-text and text-to-speech systems (Saaras and Bulbul), a document-understanding tool called Sarvam Vision, and, since February 2026, a consumer-facing chat app called Indus that wraps the 105B model in a WhatsApp-simple interface supporting all 22 of India’s scheduled languages. The company’s pitch is that global models such as GPT and Gemini were built English-first and adapted afterwards for Indian languages, while Sarvam’s models are trained on Indian language data and Indian voice patterns from the start, and are meant to be cheaper to run for the same task because they need fewer computing tokens to say the same thing in Hindi, Tamil or Telugu.

The origin

Vivek Raghavan and Pratyush Kumar did not start as typical startup founders. Raghavan, an IIT Delhi engineering graduate, spent more than a decade at India’s Unique Identification Authority (UIDAI), the body that built Aadhaar, India’s national digital identity system used by more than a billion people. Kumar, an IIT Bombay graduate with a PhD from ETH Zurich, spent 15-plus years as a computer science researcher, including stints at IBM Research and Microsoft Research, before joining IIT Madras as faculty between 2018 and 2021. Together the two ran AI4Bharat, an open research initiative at IIT Madras that built some of India’s first large public datasets and speech corpora for Indian languages.

The founding insight came out of that research: global language models tokenise Indian languages far less efficiently than English, meaning the same sentence in Hindi or Telugu can cost several times more to process, by one industry estimate roughly 3.5 to 4.5 times more tokens for Hindi and 5.0 to 6.2 times more for Telugu than the English equivalent. Voice, not text, is also how most Indians would rather interact with a computer, given lower comfort with typing in English on a keyboard. Raghavan described the task at launch as one that “requires us to change the architecture of existing open models and to train them in custom ways to teach the new language.” The founders decided that academic research alone could not fix this at the scale they wanted; as one account of the founding put it, building a company “with commercial incentives, a product focus, and VC capital was the only way to take what AI4Bharat had learned and deploy it at population scale.” Sarvam AI was incorporated in July 2023.

The struggle years

Sarvam’s difficulties were not hidden behind polished announcements; they played out in public, twice.

The first was structural and lasted from the company’s founding through early 2025. Sarvam’s entire war chest through December 2023 was $41 million. Global rivals it was implicitly competing with for talent, compute and attention — OpenAI, Anthropic, Google — had raised, by comparison, eight to eighteen billion dollars or more. For roughly sixteen months, Sarvam had to train and ship models on a fraction of the compute budget available to the labs whose output it was trying to match or beat on Indian-language tasks, with no government compute support yet in place.

The second setback was sharper and better documented. On 23 May 2025, Sarvam released Sarvam-M, a 24-billion-parameter model built by fine-tuning Mistral Small, a foundation model licensed from the French lab Mistral AI, rather than training from scratch. The reception undercut the company’s own sovereign-AI positioning: critics called it a “wrapper” model, and by one account its Hugging Face downloads numbered only in the low hundreds against the roughly 200,000 downloads racked up by comparable open releases from rivals in the same window. Forbes India separately described the model’s reception as reflecting “significant adoption challenges” for the company. Rather than defend the release, the founders scrapped the fine-tuning approach. Sarvam went quiet publicly for close to nine months, restarting model development from the ground up — new data pipelines, new tokenisers, new base-model training on Indian infrastructure — before resurfacing with a very different kind of launch in February 2026.

The turning point

The event that changed Sarvam’s trajectory arrived in the window between those two struggles. On 26 April 2025, according to Sarvam’s own announcement (some press coverage dates the news two to three days later, around 29 April 2025), India’s Ministry of Electronics and Information Technology selected Sarvam, from a pool of 67 shortlisted applicants, to build the country’s first indigenous sovereign large language model under the IndiaAI Mission. The selection came with dedicated, subsidised compute: access to roughly 4,096 Nvidia H100 GPUs (one government-technology outlet put the figure at 4,086) for a six-month training run.

The numbers on each side of that decision are stark. Before the selection, Sarvam had $41 million of equity funding in total and a flagship model, Sarvam-M, that had just landed badly with developers. After the selection, it had guaranteed access to a compute bill later reported at ₹246.71 crore for that six-month GPU allocation, of which the government subsidised ₹98.68 crore — about 40 percent — through compulsorily convertible debentures rather than a cash grant. That was compute worth more than double Sarvam’s entire prior funding, at a steep discount, precisely while the company was retraining its models from scratch. Sarvam used that window to build Sarvam-30B and Sarvam-105B, open-sourcing both in February 2026 at the India AI Impact Summit in New Delhi, alongside the Indus consumer app that put those models in front of ordinary users for the first time.

The money behind it

Sarvam’s funding has come in two very different shapes. The first, on 6 December 2023, was a combined seed and Series A of $41 million, led by Lightspeed Venture Partners, with Peak XV Partners (co-lead on the seed) and Khosla Ventures participating. At the time, TechCrunch called it the largest raise at that stage for an Indian AI startup; the capital funded the earliest versions of Sarvam-1 and the company’s core language-model research team.

The second, announced 15 June 2026, was the first close of a Series B: $234 million of a targeted $300 million round, at a post-money valuation of $1.5 billion (₹1,44,000 crore approximately, using ₹96 to the dollar) that made Sarvam India’s newest AI unicorn. HCLTech, the IT services arm of the HCL Group, led with ₹1,427 crore (about $150 million) for a reported 10.46 percent stake, bringing an enterprise distribution channel alongside its capital; Bessemer Venture Partners joined as a new investor; Khosla Ventures and Peak XV Partners returned. Total funding raised to date comes to roughly $275 million across both rounds. Layered on top of that private capital is the government’s IndiaAI Mission compute subsidy — structured, according to one financial-markets report, as compulsorily convertible debentures that convert into equity at the Series B valuation, reportedly giving India’s central government a stake of one to two percent in a private AI company, said to be a first for the country.

How it makes money

Sarvam runs three revenue lines side by side. The first is a consumption-based developer API marketplace, billed per token or per call, which the company says now handles more than 10 million API calls a day. The second is enterprise business: multi-year software licenses and on-premise or air-gapped deployments sold to regulated sectors such as banking, insurance and healthcare, often bundled with Sarvam engineers who fine-tune models on a client’s own data. The third is government contracting: infrastructure projects with bodies such as UIDAI and the Ministry of Agriculture, and state governments including Odisha and Tamil Nadu, typically funded in part through the broader ₹10,000 crore IndiaAI Mission budget.

The margin story is really the tokenisation story. Because Sarvam’s tokenisers are built for Indian languages, the company can process a Hindi or Tamil query using meaningfully fewer computing tokens than a Western model built for English would need for the same query — the efficiency gap that made Indian-language use expensive on foreign models in the first place. That lets Sarvam price competitively against global rivals while keeping more of each rupee of revenue as margin, at least in principle. What people tend to get wrong is treating a government-anointed “sovereign AI” company as effectively a public-sector body cushioned from commercial pressure. By the company’s own framing as of February 2026, Sarvam was still “primarily pre-revenue at scale,” meaning the pilots — 17 million farmers reached through a Ministry of Agriculture voice programme, 45 million insurance policy renewals supported for one insurer, a 350,000-person sales force enabled on an agentic platform for a financial-services client — had not yet converted into revenue proportionate to that reach. The mission mandate buys compute and credibility; it does not itself buy paying customers.

The numbers

Sarvam does not publish audited profit-and-loss statements, and none could be verified from public filings trackers for this piece; only top-line revenue estimates, compiled from regulatory filings by business-data trackers, are publicly available. Two trackers diverge sharply on FY25: Inc42’s Datalabs puts FY25 revenue at ₹29.1 crore, up from ₹11.1 crore in FY24, while other public trackers cite a much smaller ₹1.5 crore for the same year. FY26 revenue (year ended March 2026) is more consistently reported, at ₹45.10 crore, unaudited.

The absence of published loss figures is itself informative for an early-stage, capital-intensive AI lab: model training at the 105-billion-parameter scale is widely described in industry commentary as a multi-million-dollar undertaking per run, and Sarvam has relied on a time-limited, discounted government GPU allocation rather than steady-state paid infrastructure for its most recent flagship training run. That points to real cash burn even though no verified loss figure exists in the public record at the time of writing.

Where the money comes from

Three broad channels account for Sarvam’s reach, by the company’s own disclosures around its June 2026 funding round. Government and public infrastructure: UIDAI’s Aadhaar system uses Sarvam’s stack for voice interactions across ten Indian languages, and a Ministry of Agriculture voice programme reached 17 million farmers. Enterprise, chiefly banking, insurance and financial services: one insurance client used Sarvam’s voice AI to support 45 million policy renewals, and one financial-services client runs a 350,000-person sales force on Sarvam’s agentic platform; the company also reports more than 500,000 hours of audio transcribed and 35 million document pages digitised monthly across enterprise clients. Developers and the API long tail: more than 10 million API calls a day and 2 million-plus conversational interactions daily across Sarvam-powered products.

The surprise, given how much of Sarvam’s public narrative is about sovereign, government-backed AI, is that its commercial engine and its highest-profile new investor both sit on the private enterprise side: HCLTech’s $150 million, for a 10.46 percent stake, buys Sarvam a route into HCLTech’s existing enterprise software client base, a distribution channel no government mandate could replicate. Government partnerships supply cheap compute, legitimacy and scale pilots; enterprise deals, increasingly, look set to supply the recurring revenue.

The risks

Three risks stand out, each with a fairly specific mechanism rather than a generic warning.

Compute dependency: Sarvam’s most important recent model family was trained substantially on a temporary, subsidised GPU allocation — 4,096 H100 GPUs for six months, worth a reported ₹246.71 crore of which the government covered roughly 40 percent. When that allocation lapses, Sarvam has to fund training runs of similar scale from its own balance sheet or its next compute partnership, against just $275 million raised to date versus rivals with multiples of that in reserve.

Revenue-to-valuation gap: even taking the higher, more recent FY26 revenue figure of ₹45.10 crore against a $1.5 billion valuation (roughly ₹1,44,000 crore at ₹96 to the dollar), Sarvam trades at a revenue multiple in the hundreds — a bet on future scaling of enterprise and API revenue rather than on current earnings, and one that assumes the pilots described above convert into multi-year contracts.

Competitive and governance pressure: global model providers are closing the Indian-language performance gap faster than many expected as of early 2026, narrowing Sarvam’s core differentiation, while domestic rivals are pursuing the same market from different angles. Separately, the government’s compute-for-equity structure — converting its GPU subsidy into a reported one to two percent equity stake via compulsorily convertible debentures — creates a governance question about how a company balances public-interest, low-margin deployments (identity systems, farmer outreach) against the higher-margin enterprise contracts its investors are counting on.

The takeaway

Being chosen is not the same as having won. A government mandate, a subsidised GPU allocation and a marquee strategic investor can buy a young company runway, credibility and distribution that its balance sheet alone could never afford — Sarvam’s rise from a $41 million startup to a state-anointed sovereign-AI builder to a $1.5 billion unicorn in under three years is proof of that. But none of it insulated the company from the moment, in May 2025, when it shipped a model that developers judged, in public, on its own technical merits and found wanting. The recovery from that moment — not the mandate that preceded it — is what actually rebuilt Sarvam’s credibility. Institutional backing can open doors; only the product decides whether anyone stays once they walk through.

Frequently asked questions

What does Sarvam AI do?

Sarvam AI builds foundation language and speech models tuned for Indian languages and sells access through developer APIs, enterprise licenses for sectors such as banking and insurance, and government infrastructure contracts. Its products include the Sarvam-30B and Sarvam-105B language models, the Saaras and Bulbul speech systems, and the Indus consumer chat app launched in February 2026.

Who founded Sarvam AI and when?

Sarvam AI was founded in July 2023 by Vivek Raghavan, formerly of India’s UIDAI (Aadhaar) authority, and Pratyush Kumar, a computer scientist who previously worked at IBM Research and Microsoft Research and taught at IIT Madras. Both had run the AI4Bharat research initiative before starting the company.

How much funding has Sarvam AI raised and at what valuation?

Sarvam raised a combined seed and Series A of $41 million in December 2023, led by Lightspeed Venture Partners with Peak XV Partners and Khosla Ventures. In June 2026 it closed the first tranche of a Series B, $234 million of a targeted $300 million, at a $1.5 billion post-money valuation led by HCLTech, taking total funding to roughly $275 million.

What is the IndiaAI Mission and how is Sarvam involved?

The IndiaAI Mission is the Indian government’s programme to build domestic AI capability, including subsidised access to GPU computing for selected companies. In April 2025, Sarvam was selected, from 67 shortlisted applicants, to build India’s first indigenous sovereign large language model, receiving access to roughly 4,096 Nvidia H100 GPUs for a six-month training run at a government-subsidised rate.

Is Sarvam AI profitable, and is it listed?

Sarvam AI is privately held and not listed on any stock exchange. No audited profit or loss figures are publicly available; reported unaudited revenue for the year ended March 2026 stood at ₹45.10 crore, and the company has described itself as still largely pre-revenue relative to its scale of deployment as of early 2026.

Sources

Figures are as of September 2026. Currency converted at $1 ≈ ₹96.0 as of 18 September 2026 (Trading Economics).

• TechCrunch, “India’s Sarvam AI raises $41 million from Lightspeed, Khosla, Peak XV”, December 2023
• TechCrunch, “Sarvam becomes India’s newest AI unicorn with $234 million funding round led by HCLTech”, June 2026
• Wikipedia, “Sarvam AI”, accessed September 2026
• ConstructionWorld.in, “MeitY selects Sarvam AI to build India’s first sovereign LLM”, April 2025
• VARINDIA, “Sarvam selected to build sovereign AI model under IndiaAI Mission”, April 2025
• Sarvam AI company blog, “India’s sovereign LLM” announcement, April 2025
• Sarvam AI company blog, “Announcing Series B”, June 2026
• Inc42, “Sarvam AI — Funding, Revenue & Investors” company profile, accessed September 2026
• Entrackr, “Sarvam turns unicorn after $234 Mn round led by HCLTech”, June 2026
• MoneyFlow India (Substack), “Sarvam AI’s $234M unicorn round” briefing, June 2026
• HCLTech press release, “Sarvam raises $234 million first close of $300 million Series B at $1.5 billion valuation”, June 2026
• Forbes India, “Sarvam AI’s Pratyush Kumar and Vivek Raghavan on building AI for Bharat”, 2025
• Tech.Yahoo.com (TechCrunch), “India’s Sarvam launches Indus AI chat app as competition heats up”, February 2026
• NVIDIA, “Sarvam AI brings sovereign AI to 1.4 billion people with NVIDIA” case study, 2026
• Niftytrader.in, “Sarvam AI deal: how ₹98 crore GPU support became equity”, June 2026
• ValueForStartups.in, “Sarvam AI — India’s sovereign LLM, funding & business model 2026”, 2026
• BusinessOutreach.in, “Sarvam AI business model: how India’s AI unicorn makes money in 2026”, 2026
• A Junior VC, “Can $1.5B Sarvam be India’s intelligence infrastructure?”, 2026

Found an error? Write to us and we’ll correct it in the open, dated, on the piece.