India's AI Ecosystem Hits $2.9B: Sarvam Open-Sources Its 105B Foundational Model

Sarvam's open-source bet: what Sarvam-30B and Sarvam-105B actually are

In February 2026, Bengaluru-based Sarvam AI released two foundational large language models under an open licence: Sarvam-30B and Sarvam-105B. These are not the speech recognition or transliteration tools Sarvam built its early reputation on. They are full-stack generative language models — trained from scratch to handle the breadth of India's official language landscape, including Hindi, Tamil, Telugu, Kannada, Bengali, Marathi, and a further 16 of the country's 22 scheduled languages.

The distinction matters. Earlier open-source Indian AI releases tended to be fine-tuned adaptations of Western base models — essentially Llama or Mistral with additional multilingual data layered on top. Sarvam's 2026 release claims a different origin: foundational pre-training on curated Indian-language corpora, with architecture decisions made for cost-efficient inference on domestic infrastructure rather than for maximising Western leaderboard scores.

"Open-source" in this context means weights released publicly under a commercial-use licence, with the stated intent that any developer, startup, or enterprise can self-host or deploy via API without routing queries through a US-controlled cloud. That is the practical sovereignty argument: sensitive Indian-language data — government records, healthcare notes, legal documents, educational content — stays within the jurisdictional bounds its users expect.

The model sizes are meaningful. At 105B parameters, Sarvam-105B is competitive in raw scale with models that global labs charge significant API fees to access. At 30B, the smaller sibling is targeted at edge deployment and cost-sensitive inference scenarios — the kind of workloads that dominate India's AI adoption curve, where margins are thin and compute costs are scrutinised far more closely than in Western enterprise settings.

The ecosystem behind the models: $2.9B and 1,700 companies

Sarvam AI's open-source release did not emerge from a vacuum. It is one signal within a broader Indian AI ecosystem that has quietly become one of the world's most active — and most distinctively structured — concentrations of AI development.

India now hosts more than 1,700 AI-focused companies. Total venture capital and strategic investment flowing into the ecosystem has exceeded $2.9B across the top players. A single-year figure from MeriShiksha put total Indian AI startup funding at $1.48B in 2025 alone, suggesting the pace of capital deployment is accelerating rather than plateauing.

Microsoft's commitment is the largest single external signal of conviction: the company announced a $17.5B investment in India's AI infrastructure between 2026 and 2029, covering data centres, cloud capacity, and developer ecosystem programmes. For context, this exceeds what Microsoft committed to any single European market in the same period.

The IndiaAI Mission sits at the centre of the government's response to this moment. Backed by public funding, the Mission has formalised partnerships with 12 LLM development organisations — Sarvam and BharatGen among them — and is actively building out GPU cluster access and AI research centre infrastructure. The intent is to prevent India's AI talent from being absorbed exclusively into US lab pipelines and to ensure domestic models exist for sensitive, regulated, or linguistically complex use cases.

The pattern across these companies is consistent: each occupies a different layer of the stack — models, silicon, compute, data — and the overall picture is of an ecosystem building vertical depth rather than horizontal breadth. That is a deliberate strategic choice, and one that reflects the specific constraints and opportunities of the Indian market.

Krutrim's parallel play: unicorn, hardware, and sovereign silicon

No discussion of India's AI moment is complete without Krutrim. Founded by Bhavish Aggarwal — the entrepreneur behind Ola, one of India's most recognised consumer technology companies — Krutrim became India's first AI unicorn when its valuation crossed $1B. The company has since secured $50M in equity and $230M in committed financing, making it the best-capitalised pure-play AI company on the subcontinent.

What distinguishes Krutrim from Sarvam is its hardware ambition. The company's Bodhi-1 chip programme is India's most visible effort to build sovereign AI silicon — the inference accelerators that underpin modern LLM deployment at scale. The strategic logic mirrors what China's Huawei and Cambricon have pursued: if you depend on Nvidia GPUs and US export licences for your AI compute, your sovereignty is conditional. Krutrim is attempting to close that gap.

The software and hardware efforts are complementary rather than competing. Krutrim's models run on its own infrastructure; the Bodhi-1 chip, if it reaches commercial volumes, could become the substrate on which Indian AI runs at national scale. Krutrim's path to profitability depends on this vertical integration delivering unit economics that Western-hosted alternatives cannot match for Indian enterprise workloads.

Aggarwal's stated vision is explicitly sovereign: AI infrastructure for India, built and owned in India, without dependence on foreign supply chains at the critical junctures. Whether Bodhi-1 achieves production viability is a separate question from whether the strategy is sound — and most analysts who cover the Indian market believe the strategy is.

The Sarvam Startup Programme: who it's for and what it offers

A foundational model release is only as useful as the developer ecosystem that forms around it. In March 2026 — one month after releasing Sarvam-30B and Sarvam-105B — Sarvam AI launched the Sarvam Startup Programme, a structured support offering for early-stage builders wanting to build on sovereign AI infrastructure.

The programme provides two things that matter most to startups in their first two years: API credits and engineering support. API credits reduce the cost of experimentation and prototyping; engineering support from Sarvam's team accelerates integration and troubleshooting for builders who are working with Indian-language data for the first time.

Who should apply

The programme is most obviously suited to builders working on use cases that require genuine multilingual coverage — healthcare platforms serving non-English-speaking patients, agricultural advisory tools for rural farmers, legal aid applications, vernacular education platforms, and government service delivery applications where operating in the user's first language is not a feature but a requirement.

It is equally relevant to UK-based builders expanding into India. A UK fintech, insurtech, or edtech that wants to serve Indian markets credibly — rather than treating English as a proxy for accessibility — now has a structured path to sovereign-model API access without building its own multilingual stack from scratch.

The broader implication for India's builder community

The Sarvam Startup Programme signals something more structural than an API credit offer. It reflects a maturing playbook: open-source the model, build a developer community around it, establish a startup ecosystem that creates switching costs organically. This is precisely how Hugging Face built its gravity in the Western open-source AI world, and how cloud providers have long locked in developers through generous early-stage credits.

For Indian builders, the choice is now concrete: build on US-hosted APIs, build on Chinese open-weight models, or build on sovereign Indian infrastructure. Sarvam's programme is the most explicit attempt yet to make the third option as frictionless as the first two.

India versus the world: what "sovereign AI" actually means in 2026

The phrase "sovereign AI" risks becoming marketing noise — invoked by every national government and tech company that wants to sound strategically serious. In India's case, the substance behind the phrase is more specific and more defensible than the rhetoric suggests.

India has 22 official languages. Hundreds of millions of its residents are functionally more literate in their mother tongue than in English. Any AI application that matters at population scale — government services, healthcare, agriculture, education, financial inclusion — must operate in those languages. The existing global open-weight models, including Meta's Llama 4 and Alibaba's Qwen family, are capable of some multilingual handling but were not designed around India's specific linguistic distribution, script diversity, or the particular data sparsity challenges that come with less-resourced Indian languages.

This is not a benchmark competition. India's sovereign AI push is not an attempt to outperform DeepSeek on HumanEval or beat GPT-4o on MMLU. It is an attempt to build models that are fit for the actual use cases that matter in the world's most populous country.

The Digital Personal Data Protection Act (DPDP) Phase 2 compliance requirements add a regulatory dimension. Data localisation norms under DPDP create legal friction for routing sensitive Indian data through foreign APIs — friction that sovereign models eliminate by design. For builders operating in regulated sectors, this is not a philosophical preference but a compliance necessity.

The context for India's AI ecosystem reaching a $50B total addressable market milestone is precisely this: a market large enough, distinctive enough, and now well-enough capitalised to sustain its own AI stack rather than perpetually importing it.

What this means for UK AI builders

The UK has a specific and underappreciated stake in India's AI moment. Indian AI engineers are among the most recruited globally — UK AI teams in London, Manchester, and Edinburgh regularly hire from the same talent pool that feeds Sarvam, Krutrim, and Neysa. The builders most likely to understand how to deploy sovereign AI models for Indian markets are often already working in British companies.

UK enterprises expanding into India face a choice that did not exist two years ago. Previously, the default was to use the same US-hosted LLM APIs they deployed domestically and add a translation layer. Sarvam-105B and the broader sovereign AI ecosystem now offer an alternative: deploy models trained specifically for Indian linguistic and regulatory requirements, with the open-weight access needed for on-premises or sovereign-cloud deployment.

UK professional services firms, fintechs, and edtechs with India operations should be watching Sarvam's model releases as closely as they watch GPT or Gemini updates. The trajectory of Sarvam's funding rounds suggests the company will continue to scale its model capabilities — and the Sarvam Startup Programme is explicitly designed to be accessible to international builders, not just domestic ones.

There is also a talent arbitrage angle. UK AI teams that develop genuine expertise in Indian-language AI — understanding the data pipelines, the evaluation frameworks, the specific failure modes of models on low-resource scripts — will be well positioned as more UK enterprises seek to serve Indian markets meaningfully rather than superficially. Profiles in our Verified Builders directory already include engineers with exactly this background, and demand is increasing.

The broader lesson from India's 2026 AI surge is one that applies to the UK's own AI strategy: open-source foundational models reduce the dependency of entire national economies on a handful of US-based labs. India is demonstrating that a country with sufficient talent density and strategic intent can build and sustain its own AI stack. The UK, watching from a shared-language advantage, has every reason to engage with that ecosystem rather than simply observe it.

For builders ready to act now, the practical starting points are clear: explore the IndiaAI Mission's LLM partner network, apply to the Sarvam Startup Programme for API access, and add your profile to AI Tech Connect to be found by teams building at this intersection. The infrastructure is here; the question is who moves first.