Executive Summary
Executive Summary
This white paper, produced by India's Principal Scientific Adviser's office, examines the rise of foundation models in AI and India's strategy around them. In a nutshell, foundation models (large, pre-trained AI models) are transformative because the same model can be adapted to many tasks instead of training separate systems for each. This gives huge power but also concentrates influence - design choices in these base models affect many downstream applications and sectors. The paper emphasizes India's approach: build homegrown foundation models aligned with Indian languages, data and values, backed by public compute and datasets, while also establishing governance (guidelines, laws, safety) around their use.
Key points include:
Key points include:
- Innovation Pillar: Under the IndiaAI Mission, India launched calls for proposals and selected teams (industry plus academia) to develop indigenous large language and multimodal models. By early twenty twenty-six, twelve consortia (from startups like Sarvam AI, Soket AI, Gnani AI, etc., to IIT Bombay) have been funded to build these models on Indian data. The idea is to ensure India has sovereign, open-source models for its own needs.
- Compute Infrastructure: A national AI Compute Portal provides massive GPU resources (tens of thousands of GPUs) at subsidized rates. Originally aiming for ten thousand GPUS, India has onboarded thirty-eight thousand GPUs for public use. (Extra: Globally, such infrastructure investment is huge - for context, worldwide AI spending is projected at two point five two trillion dollars in twenty twenty-six.)
- Data Platform (AI Kosh): The AI Kosh platform aggregates curated datasets and models. It hosts thousands of datasets and hundreds of models (for example, three hundred plus datasets by early twenty twenty-five, now rapidly expanding) spanning many sectors. The goal is high-quality, India-specific data (Indic languages, local domains) so foundation models learn relevant patterns.
- Governance Framework: IndiaAI launched National AI Governance Guidelines (twenty twenty-five), built around seven core "Sutras" (principles) like Trust, Fairness, Safety, etc. The approach is explicitly risk-based, meaning stronger rules for high-risk AI applications. Alongside this, the new Digital Personal Data Protection Act, twenty twenty-three provides a legal framework for privacy and data use in AI. There are also CERT-In advisories and a new AI Safety Institute forming to handle threats and ethics.
- AI for National Priorities: The paper highlights focusing AI on India's challenges - for example, healthcare, agriculture, education, local language technology, etc. (Extra: For instance, India has launched "Bharat-VISTA," a multilingual AI tool to advise farmers using local data.) The idea is that homegrown foundation models and applications help solve on-the-ground problems in an Indian context.
Throughout, the message is that India is building a balanced AI strategy: invest in advanced tech (foundational models, compute, data) while ensuring ethical guidelines, privacy laws, and alignment with India's socio-economic diversity. External data shows this aligns with global trends - India now ranks number three in global AI competitiveness behind only the United States and China - and the world is pouring unprecedented money into AI (global AI spend approximately two point five trillion dollars in twenty twenty-six). In summary, the document outlines how India plans to catch up with big tech in generative AI by growing its own models and ecosystem, while also keeping them safe and locally relevant.