The Architecting of Bharat’s Brain:India’s Sovereign AI Revolution

In the global theatre of Artificial Intelligence, a new narrative is being written — one where India is no longer just a consumer of Western code, but a primary architect of its own intelligence. With the India AI Impact Summit 2026 serving as a defining backdrop, it is clear that the nation is in a well-funded, rapidly accelerating race to build its own Sovereign AI: ensuring the digital future of 1.4 billion people is rooted in their own languages, values, and infrastructure.

₹10,372 Cr IndiaAI Mission Budget

38,000+ Public GPU Pool

9,500+ Datasets on AIKosh

13 Languages — Kruti AI

The Champions: Sarvam AI and Krutrim

At the heart of India’s sovereign AI race are two homegrown powerhouses — Sarvam AI and Krutrim. These companies are not building incremental chatbot wrappers on top of foreign models; they are creating foundational AI systems designed from the ground up for India’s scale, its languages, and its infrastructure constraints.

Sarvam AI, founded in 2023 by IIT Madras alumni Vivek Raghavan and Pratyush Kumar, made its most significant leap at the India AI Impact Summit in February 2026. The Bengaluru-based startup unveiled two frontier-class models: Sarvam-30B and Sarvam-105B — both trained from scratch entirely in India on compute provided under the IndiaAI Mission. The 30B model is built on a Mixture of Experts architecture, activating approximately 1 billion parameters per token, making it efficient enough to run on basic feature phones. The 105B model, named Indus, targets complex reasoning and enterprise applications with a 128,000-token context window — benchmarking competitively against leading global models including DeepSeek R1 and GPT-class systems on key Indian language tests.

The showstopper at the Summit was Sarvam Kaze — indigenous AI-powered smart glasses engineered to listen, understand, and capture what the wearer sees, supporting over 10 Indian languages in real time. Prime Minister Narendra Modi became the first person to try the device at Bharat Mandapam, with commercial availability targeted for May 2026. The move signals India’s ambitions extending beyond AI software into hardware sovereignty.

Krutrim, founded by Ola’s Bhavish Aggarwal and India’s fastest-ever AI unicorn, is pursuing a different but equally ambitious path. Rather than competing model-for-model with Silicon Valley, Krutrim is building the full-stack infrastructure that a national AI requires. In partnership with NVIDIA, it has deployed India’s first GB200 Grace-Blackwell supercomputer — initially four NVL72 racks delivering 5.6 exaflops of FP4 inference capacity — with a roadmap to scale this into one of the country’s most powerful AI infrastructure facilities. Its agentic assistant Kruti, which supports 13 Indian languages, can book cabs, pay bills, generate images, and perform multi-step tasks — positioning it as a genuinely useful product for everyday Indians rather than a technology demonstration.

" Sarvam’s 105B model was trained from scratch, entirely in India, on government-provided compute — and benchmarks competitively with leading global AI systems. "

The Government as the Ultimate Catalyst

While private startups provide the innovation, the Government of India is providing the infrastructure oxygen. Under the ₹10,372 crore IndiaAI Mission, the state has moved decisively from the role of regulator to that of a massive enabler. Building Sovereign AI requires a domestic chip supply chain — a goal that India is steadily advancing through ISM 2.0 and its semiconductor mission.

The government’s strategy mirrors the Digital Public Infrastructure (DPI) philosophy that produced UPI and Aadhaar. By making over 38,000 high-end GPUs available at a subsidised rate of approximately ₹65 per hour, the Ministry of Electronics and IT has effectively democratised compute — allowing a bootstrapped startup in Bhopal or a researcher at a state university to train AI models that were previously the exclusive domain of well-capitalised Silicon Valley labs.

The AIKosh platform completes this infrastructure equation. Hosting over 9,500 datasets and 273 sectoral models, AIKosh ensures that India’s AI models are trained on high-quality, culturally relevant Indian data — not simply whatever happened to be scraped from the English-language internet. Datasets range from agricultural records and medical histories to digitised government documents in regional scripts.

Sarvam AI is among 12 startups selected under the IndiaAI Mission to build sovereign foundational models, and received access to 4,000 high-end GPUs under the programme. Both its 30B and 105B models were trained entirely on this government-provisioned compute — a proof of concept that the Mission’s infrastructure strategy is working in practice, not just on paper.

Civilisational Sovereignty: The BharatiyaGPT Milestone

Perhaps the most philosophically distinctive dimension of India’s sovereign AI story is BharatiyaGPT — developed by Nagpur-based ImmverseAI and formally showcased at the India AI Impact Summit 2026. While Sarvam and Krutrim focus on contemporary Indian language and enterprise use cases, BharatiyaGPT addresses a different, equally important challenge: what happens to the knowledge that existed long before the internet?

India holds the world’s largest known collection of ancient and medieval manuscripts — an estimated 10 million texts written across 40-plus distinct scripts and covering fields including medicine (Ayurveda), mathematics, astronomy, architecture, philosophy, and governance. The overwhelming majority have never been digitised, let alone made accessible. ImmverseAI has built a complete data pipeline from the ground up: collecting ancient Indic manuscripts, digitising them, and training AI models on the resulting corpus — over 50 terabytes of data encompassing more than 10 lakh manuscripts across 23 languages. The core model is trained on 8 billion parameters and draws on texts including the Vedas, Upanishads, Ramayana, Mahabharata, and Bhagavad Gita.

What makes BharatiyaGPT distinct is not merely text retrieval. The system returns not just answers, but original shlokas with contextual explanation, conceptual linkage within the source scripture, and domain-specific reasoning — whether the query is on ancient surgical technique, Vedic mathematics, or early metallurgy. Specialised vertical models include AyurvedaGPT, GanitGPT (mathematics), and SthapatyaGPT (architecture). This represents a form of civilisational sovereignty in AI: ensuring that India’s pre-colonial knowledge systems are not simply lost to history, but become active, queryable, and useful in the digital age.

The Race Is Real — and the Pace Is Accelerating

The shift in India’s AI story in 2026 is qualitative, not just quantitative. The transition is from adaptation — deploying foreign AI in Indian contexts — to sovereignty: building AI that is Indian by design, not by adaptation. Union Minister Ashwini Vaishnaw has framed this as India’s participation in the Fifth Industrial Revolution, with AI as the defining technology.

The real-world applications are already live. ISRO is piloting offline, auditable AI systems for space observation — models that cannot transmit sensitive data to foreign servers. The National Health Authority is evaluating medical-specific AI for digitising fragmented clinical records across district hospitals. The Unique Identification Authority of India has announced a collaboration with Sarvam AI to integrate voice-based multilingual AI into Aadhaar services — potentially the largest-scale deployment of indigenous AI in history, reaching over a billion enrolled users.

How Close Is the Sovereign Stack?

India is closer than ever to a fully realised sovereign AI stack, but honest assessment demands acknowledging the gaps. The milestones are real: two frontier open-source models trained entirely in India, a functioning GB200 supercomputer, over 38,000 GPUs of public compute, BharatiyaGPT’s 10 lakh manuscripts, and live government deployments. These are not roadmap slides — they are operational realities. At the grassroots level, AI adoption is already accelerating — read how AI agents are transforming Indian SMBs in 2026.

The hurdles that remain are primarily hardware and energy. India’s AI compute infrastructure still relies on imported NVIDIA chips. While NVIDIA has deepened its partnership with the IndiaAI Mission — collaborating with L&T, Yotta, and E2E Networks to build sovereign AI factory infrastructure across Chennai, Bengaluru, and Mumbai — domestic chip design and fabrication at AI-relevant scales remains an unsolved challenge. Krutrim’s roadmap includes indigenous ‘Bodhi’ AI chips targeted for 2026 deployment, but these have yet to be publicly demonstrated.

Energy is the second frontier. Training large AI models is extraordinarily power-intensive. The green-energy-powered AI data centre model — championed by groups like Adani and Reliance — is critical to making India’s AI ambitions both scalable and sustainable. India’s renewable energy buildout gives it a genuine long-term advantage here, but the current infrastructure gap between ambition and operational capacity remains significant.

The Verdict: India Is Architecting, Not Just Adopting

India is no longer asking whether it should adopt AI; it is deciding how to architect it. The combination of the IndiaAI Mission’s compute and data infrastructure, a maturing ecosystem of technically credible startups, BharatiyaGPT’s civilisational knowledge base, and a population of over 800 million non-English speakers who need AI that actually understands them — creates a genuinely distinctive opportunity.

By 2027, India may not simply be recognised as an AI services hub. If the current trajectory holds, it could emerge as the pioneer of a new model of population-scale, multilingual, sovereign AI — one built not for the few who speak English and own GPUs, but for the many who speak Bhojpuri, Tamil, and Marathi and carry a feature phone. That is a story worth watching closely.