On 13 March 2026, the Office of the Principal Scientific Adviser to the Government of India released a white paper titled "Advancing Indigenous Foundation Models." It is the latest in an ongoing AI Policy White Paper Series that has been building, document by document, the intellectual architecture of India's AI ambitions. Read alongside the India AI Impact Summit held in February — attended by delegations from over 100 countries, generating $200 billion in investment commitments, and producing the New Delhi Frontier AI Impact Commitments — it suggests that India is doing something more than catching up. It is attempting to shape the terms on which AI develops in the Global South.

The core argument of the white paper is structurally sound. Most AI models currently deployed in India are developed overseas, trained on datasets that do not fully represent India's twenty-two official languages, its agricultural conditions, its healthcare patterns, or its legal and governance frameworks. The models perform adequately in English and for use cases developed with Western users in mind. They perform poorly — or not at all — in Bhojpuri, Santali, or Dogri. They cannot navigate India's land records system or assist an MSME owner in Tamil Nadu drafting a compliance document.

Indigenous foundation models, the paper argues, correct this by design rather than by adaptation.

The Two-Tier Architecture

India's strategy is not simply to replicate what OpenAI or Google DeepMind have built. The white paper is explicit that large language models and small language models serve different functions and that both are necessary components of a coherent national AI ecosystem.

Large foundation models — trained on vast, diverse datasets — provide general capability across sectors. They are expensive to build and maintain, require substantial compute infrastructure, and the number of countries that can credibly develop them at the frontier is small. India's IndiaAI Mission has provisioned over 38,000 GPUs at approximately ₹65 per hour — roughly a third of global average costs — to make frontier-level training economically accessible to domestic institutions. Twelve startups and institutions, including Sarvam AI, IIT Bombay's BharatGen, and Fractal Analytics, have been selected under the mission for model development.

Small language models — specialised systems trained for narrow domains — are the more immediately practical component. Designed for agriculture advisory, healthcare triage, MSME compliance, and urban governance, they can be deployed on edge devices and run at a fraction of the computational cost of frontier models. For a country where broadband access is uneven and electricity supply is variable, the case for energy-efficient, deployable AI is not ideological — it is infrastructural.

The combination matters strategically. India is not trying to out-scale the United States or China in foundation model development. It is building a layered system in which a small number of capable national models sit atop a widely distributed ecosystem of task-specific models, connected to government data infrastructure and deployed through platforms like BHASHINI, the national language translation initiative. This is a realistic strategy — and it is more coherent than the purely aspirational AI sovereignty narratives that several other nations have produced.

The Governance Architecture

The policy scaffolding around India's AI ambitions has moved further and faster in the last eighteen months than at any previous point. MeitY's AI Governance Guidelines, released in November 2025 under the IndiaAI Mission, establish seven core principles — transparency, accountability, fairness, privacy, safety, and two others — and propose an inter-ministerial AI Governance Group, a Technology and Policy Expert Committee, and an IndiaAI Safety Institute for monitoring and capacity building.

The approach is deliberately light-touch. India is not enacting an AI Act in the European sense. It is not creating binding regulations that would slow deployment before use cases are established and risks are understood. The private member Artificial Intelligence (Ethics and Accountability) Bill introduced in the Lok Sabha in December 2025 — which proposes a statutory Ethics Committee, mandatory ethical reviews for high-risk systems, and bias audits — signals growing parliamentary interest in harder governance structures, but remains a Private Member's Bill without government backing.

This is the right sequencing. India's AI governance model — voluntary principles, institutional mechanisms, innovation-first framing — is calibrated to maximise deployment while building the evidence base for more targeted intervention where genuine harms emerge. The EU's approach, which codified risk classifications before deployment patterns were clear, has produced compliance costs without proportionate safety gains. India is watching that experiment carefully.

The Tension Worth Watching

The most consequential unresolved tension in India's AI framework sits at the intersection of AI development and the Digital Personal Data Protection Act. Training large language models at scale requires access to vast quantities of data — text, images, voice recordings, transactional data. The DPDP Act's consent-based architecture, designed to protect individuals' data rights, creates friction with the data-intensive requirements of foundation model training. An AI company building a Hindi LLM on Indian medical records, agricultural data, and legal judgements faces consent requirements that were designed for a different kind of data use.

India has not yet resolved this tension through regulation. The January 2026 white paper from the PSA's office advocates a "techno-legal" framework — embedding compliance directly into AI system design through watermarking, bias detection, and data provenance tracking — rather than resolving it through legislative carve-outs. This is intellectually interesting but institutionally untested. Whether it works in practice will depend on how courts interpret the DPDP Act as the first data disputes arrive.

The AIKosh Question

India's national AI dataset and model platform — AIKosh — is the foundation on which indigenous model development is supposed to rest. As of February 2026, it hosts over 7,500 datasets and 273 AI models across 20 sectors. This is a start. It is not yet a comprehensive training infrastructure.

The quality problem is as significant as the quantity problem. A large language model trained on AIKosh's current dataset would perform adequately for narrow government-service use cases and poorly for anything requiring nuanced, diverse, culturally specific content. Building the datasets that will make Indian AI genuinely competitive — in terms of quality and breadth, not just language coverage — requires the kind of sustained, unglamorous institutional effort that white papers tend to underweight and funding allocations tend to underprioritise.

The AI Impact Summit demonstrated that India can attract global attention and investment commitments. The PSA's white paper demonstrates that India can produce sophisticated policy thinking. What it has not yet demonstrated — at scale, with verifiable results — is execution: the patient, multi-year work of building training data, testing models, deploying systems in government workflows, iterating based on outcomes, and expanding coverage.

India has framed its AI ambitions correctly. It is building toward sovereignty, not dependence — toward systems that understand India, not systems that tolerate it. The architecture is sound. The investment is moving. The governance framework is taking shape. What happens next — whether the execution matches the aspiration — will determine whether India shapes the Global South's AI trajectory or simply watches it unfold.

The Hind covers policy, power, and strategic affairs from India's perspective.