The venue said as much as the technology itself. Mid-February 2026, New Delhi's India AI Impact Summit—a gathering thick with government ministers, venture capitalists, and the kind of tech executives who shape national digital policy. And there, among the panels and policy proclamations, Sarvam AI chose to unveil something audacious: a 105-billion-parameter language model trained entirely from scratch on Indian languages. No fine-tuning borrowed from Western foundations. No shortcuts.
For a startup not yet three years old, it was a bold play. Perhaps more than the founders expected, the announcement landed in the middle of a broader conversation about what it means for a nation of 1.4 billion people—speaking 22 officially recognized languages—to rely almost entirely on AI systems trained primarily on English, hosted on foreign servers, governed by regulatory frameworks written in Silicon Valley and Beijing. Sarvam's model, alongside a smaller 30-billion-parameter sibling, arrived as something tangible: an answer, or at least the beginning of one.
Building From the Ground Up
Both models use Mixture-of-Experts architecture, a technical choice that reflects the sheer complexity of handling multiple languages simultaneously without collapsing into incoherence. The 30B version supports a 32,000-token context window; the 105B stretches that to 128,000 tokens, positioning it for the kind of enterprise workflows—long-document reasoning, extended customer interactions—that demand more than chatbot banter.
According to company statements published in February 2026, the 30B model was pre-trained on approximately 16 trillion tokens. The 105B consumed "trillions of tokens" during its training run, a vague formulation that suggests either caution about revealing competitive details or the messy reality of iterative training regimes where exact counts shift as datasets evolve.
The infrastructure tells its own story. Sarvam leveraged resources from the IndiaAI Mission, a government-backed program that had deployed somewhere between 14,000 and 18,000 GPUs as of early 2026, with plans to scale toward 29,000 to 38,000 units. Data center capacity came from Yotta; technical support from Nvidia. Multiple reports from summit week cited subsidized GPU pricing around ₹116 to ₹150 per hour, though some briefings mentioned rates as low as ₹65 per hour—the kind of aggressive discounting designed to lower barriers for domestic model development and, not incidentally, build a constituency of startups dependent on government compute.
Sarvam claims native support for all 22 Schedule 8 Indian languages, plus code-mixed varieties like Hinglish. The company stated in mid-February 2026 that it plans to open-source both models, though specifics on training data provenance and code release timelines remained conspicuously unspecified as of the launch dates.
The Case for Sovereignty—and Its Complications
The argument for sovereign AI in India rests on three pillars that seem increasingly difficult to ignore: linguistic diversity, regulatory autonomy, and market opportunity.
Language first. English proficiency in India remains moderate and deeply uneven across regions, according to the EF English Proficiency Index released in November 2025. For an AI system to genuinely serve India's insurance agents, government service centers, and rural healthcare workers, it must operate fluently in Tamil, Bengali, Marathi, and a dozen other languages. Global models treat these as afterthoughts, add-ons to be bolted on later. Indian-built models—in theory—treat them as the foundation.
Then there's the matter of data sovereignty, which has moved from abstract principle to regulatory requirement. The Digital Personal Data Protection Act of 2023 saw its implementing rules notified on November 13, 2025, with phased enforcement timelines extending through 2026 and into 2027. As privacy regulations tighten, enterprises evaluating AI deployments increasingly ask uncomfortable questions: Where does training data reside? Who controls model weights? Are local alternatives even viable? Sarvam's pitch—that its models are trained in India, hosted in India, governed by Indian law—resonates in this environment, whether or not it ultimately delivers on technical performance.
The market opportunity is equally compelling, if adoption studies are to be believed. Throughout 2025, research painted India as an outlier in generative AI uptake. A Forrester report cited in August 2025 found over 56% of urban Indians using GenAI tools, leading the Asia-Pacific region. BCG research from 2025 noted that 92% of Indian employees reported regular GenAI use, well above global averages. Deloitte's May 2024 study showed 83% of Indian employees and 93% of students engaging with the technology, figures that reflect the early adoption wave even as the landscape continues to evolve rapidly.
Yet most of that adoption still relies on foreign platforms. Sarvam's founders—Pratyush Kumar, formerly of AI4Bharat and IIT Madras, and Vivek Raghavan, who helped build the technical infrastructure behind Aadhaar—frame the dependency in stark, almost stark terms. In public statements during the summit period, Raghavan emphasized the imperative of digital sovereignty and India's need to lead rather than follow in AI development. Kumar, in a Business Today interview published February 19, emphasized voice-first interfaces and deliberate scaling over what he called a "mindless" parameter race, a critique that seems aimed as much at Western AI labs as at domestic competitors rushing to chase benchmark numbers.
From Pilot to Production?
Sarvam's commercial trajectory suggests ambitions beyond academic demos. On February 26, 2026, the company announced a partnership with SBI Life Insurance to deploy AI tools—voice agents, claim bots, sales co-pilots—across a distribution network serving 80 million customers and roughly 350,000 agents. That's the kind of scale that either proves the technology or exposes its limitations quickly. Earlier in the month, on February 20, Bosch Global Software Technologies signed a memorandum of understanding to co-develop secure AI solutions for automotive, medical devices, and wearables—sectors where data sensitivity and regulatory compliance matter enormously.
The Indus chat app, launched February 20 on iOS, Android, and web, provides a consumer-facing interface to the new models. It emphasizes voice interaction, reflecting India's massive feature-phone base—estimated at around 230 million users in a September 2024 Counterpoint report. During summit week, Sarvam announced partnerships with HMD (Nokia's feature-phone brand) and Qualcomm to bring AI assistants to devices that lack touchscreens or robust internet connectivity, a market segment largely ignored by Silicon Valley.
Government collaborations add weight, if not always transparency. A March 19, 2025 press release from UIDAI detailed Sarvam's role in enabling multilingual, voice-driven interactions for Aadhaar services—potentially touching hundreds of millions of citizens, though details on deployment timelines and actual usage remained sparse. Tata Capital deployed Sarvam's "Samvaad" voice AI platform in January 2025 for customer interactions, according to a company case study that, like most case studies, emphasized benefits while glossing over challenges.
On February 5, 2026, Sarvam unveiled Sarvam Vision, a 3-billion-parameter vision-language model designed for optical character recognition, table extraction, and document understanding across English and 22 Indian languages. The company cited 84.3% performance on olmOCR-Bench, targeting enterprises drowning in multilingual paperwork—a real pain point in a country where bureaucracy operates in dozens of scripts.
A Crowded Field
Sarvam is hardly alone in this arena. BharatGen, a government-backed consortium anchored at IIT Bombay and IIT Madras, announced plans in August 2025 to cover all 22 scheduled languages by June 2026 with multimodal capabilities—a timeline that may or may not survive contact with reality. Gnani.ai launched "Vachana STT" in February 2026, trained on over one million hours of speech data. Krutrim, backed by Ola's Bhavish Aggarwal, published research in February 2025 detailing a two-trillion-token multilingual training run, though specifics on model availability remained vague. CoRover's BharatGPT family has offered smaller, offline-capable Indic models since 2024, targeting government and enterprise clients.
Further afield, Reliance Jio signaled a ₹10 trillion investment (~$120 billion) over seven years during the summit, with initial 120-megawatt capacity expected online in the second half of 2026. Mukesh Ambani's remarks on February 19, 2026 framed this as a gigawatt-scale buildout to support sovereign compute and services, the kind of infrastructure play that could reshape India's AI landscape—or become a cautionary tale about overinvestment in rapidly evolving technology.
Global players remain formidable baselines. Meta's Llama 3.1 (405B) became available in FP8 variants on Oracle Cloud Infrastructure between February and July 2025. DeepSeek-R1, a 671-billion-parameter mixture-of-experts reasoning model released January 20, 2025, disrupted cost-performance expectations across math and coding benchmarks, setting a new bar for what open-source models could achieve. Mistral's Mixtral 8x22B served as a common multilingual baseline through 2024 and early 2025 before its retirement in March 2025, replaced by more efficient successors.
Sarvam's performance claims—that the 105B model outperforms DeepSeek-R1 and Google's Gemini Flash on certain India-language benchmarks—have appeared in multiple media reports from the February 18-21 launch window. No independent, peer-reviewed leaderboard has yet confirmed these assertions, a gap that matters. The emergence of Indic-specific evaluation frameworks—IndicMMLU-Pro (January 2025), ParamBench (August 2025), BharatBench for vision-language models (February 2025)—suggests the benchmarking infrastructure is still maturing, still figuring out what "good" actually means in these contexts.
The Sovereignty Paradox
"Sovereign AI" crystallized as a concept during the India AI Impact Summit. Le Monde, reporting February 18, described India's approach as a "third way" between American platforms and Chinese state-directed models. Forbes, writing February 23, characterized it as a pitch to reduce dependence on big tech, citing aggressive pricing and domestic compute capacity as strategic levers. A Brookings Institution report published February 17 mapped AI sovereignty strategies across developing economies, positioning India's model as an emerging template—though whether other nations follow remains far from certain.
The rhetoric carries political weight. It aligns with India's broader push for digital autonomy, evident in payments infrastructure (UPI), identity systems (Aadhaar), and regulatory frameworks (DPDP Act). But sovereignty in AI is messier than sovereignty in payments. Training large models still depends on Nvidia chips, often deployed in data centers using hyperscaler cloud stacks. Open-sourcing weights democratizes access but doesn't eliminate dependency on the underlying silicon and systems, a paradox that no amount of nationalist rhetoric can resolve.
Sarvam's stated intention to release the 30B and 105B models as open-source may test those tensions. If weights become public, enterprises and researchers worldwide could deploy and modify the models—advancing the technology while complicating narratives about exclusive Indian control. The balance between openness and sovereignty remains, for now, unresolved.
What Happens Next
For CTOs evaluating multilingual AI, the arrival of credible India-trained alternatives shifts the procurement calculus, at least marginally. Enterprises no longer face a binary choice between global general-purpose models and expensive custom development. A third option—regional foundation models optimized for local languages and compliance regimes—is becoming viable, though "viable" doesn't yet mean "proven."
For venture investors, the sovereign AI thesis presents both opportunity and the usual risks. Government support lowers infrastructure costs, but market fragmentation across languages and use cases complicates go-to-market strategies in ways that don't always show up in pitch decks. Sarvam's partnerships with SBI Life, Bosch, and Nokia suggest pathways to scale, but proof will come in retention metrics and expansion beyond pilot programs—metrics that tend to look less impressive six months in.
For AI researchers, the proliferation of Indic benchmarks and datasets (RASMALAI for speech translation in May 2025, BhasaAnuvaad in November 2024, Indic-QA in the NAACL 2025 findings) signals a maturing ecosystem, one that's moving beyond simply adapting Western tools. The technical questions around mixture-of-experts routing in multilingual contexts, documented in January 2026 research, remain open problems that could yield insights applicable far beyond India—if the research community bothers to pay attention.
Regulatory clarity will matter enormously. The phased rollout of DPDP rules through 2026 and 2027 will establish boundaries for data usage, cross-border transfers, and algorithmic accountability. MeitY's ongoing advisories on deepfakes and synthetic media, evolving since March 2024, add another layer of compliance burden—or competitive advantage for local players who navigate them early and well.
Sarvam raised $41 million in December 2023 from Lightspeed, Khosla Ventures, and Peak XV. Media estimates from February 2026 suggest total capital raised exceeds $50 million, with a valuation around $200 million, though these figures are not company-filed and should be treated as approximations at best. Whether that funding suffices to sustain development, recruitment, and enterprise sales at scale remains an open question, the kind that separates promising startups from durable companies.
The broader context is difficult to ignore. India's AI infrastructure is scaling faster than most observers expected even a year ago. The IndiaAI Mission's GPU capacity, combined with Jio's planned gigawatt-scale compute facilities, represents a significant expansion of domestic compute resources, though the full economic implications for startups will depend on sustained government support and pricing policies that remain subject to change. If subsidized compute persists and quality datasets accumulate, the sovereign AI stack becomes more than rhetoric—it becomes infrastructure, the kind that shapes entire industries.
Sarvam's 105B model is not the endpoint of this trajectory. It's an early marker, a data point in a story still being written. Whether it delivers on performance claims, achieves meaningful open-source adoption, and sustains enterprise deployments will determine if India's sovereign AI ambitions translate into durable competitive advantage—or remain aspirational positioning in a market still dominated by Western and Chinese incumbents.
The next twelve months will clarify which outcome prevails. Maybe longer.
