IndiaAI Mission accelerates sovereign foundation models to build domestic computing power

India has begun rolling out the operational phase of its ₹10,372 crore IndiaAI Mission, deploying subsidised graphics processing units and commissioning indigenous multimodal models designed specifically for Indian languages.

The push marks a structural change in how Delhi handles artificial intelligence. Rather than relying entirely on proprietary systems built in North America or China, the Ministry of Electronics and Information Technology is funding sovereign foundation models that can reason across text, speech, and administrative data in 22 scheduled languages.

At the centre of this effort sits BharatGen, an initiative spearheaded by researchers at the Indian Institute of Technology Bombay under the National Mission on Interdisciplinary Cyber-Physical Systems. BharatGen is building generative foundation models that process Hindi, Tamil, Telugu, and other regional tongues natively, bypassing the tokenisation penalties that drive up inference costs when standard Western large language models handle non-Latin scripts.

What the domestic foundation models deliver

Unlike traditional natural language processing tools that merely translate queries into English before processing them, the new generation of Indian foundation models trains directly on curated domestic cultural and bureaucratic corpuses. This architecture cuts latency, improves factual retrieval on local governance frameworks, and keeps institutional data within Indian sovereign boundaries.

Alongside academic labs, home-grown commercial labs such as Sarvam AI and Krutrim are deploying domain-specific architectures aimed at low-cost inference on enterprise and public-sector workflows. Sarvam's open-source release of Sarvam 2B, an efficient small language model optimised for Indian vernaculars, demonstrated that smaller, well-curated architectures can match or exceed the performance of larger global models on local-language reasoning benchmarks.

Solving the national compute bottleneck

State-backed foundation models cannot scale without silicon. To resolve the compute shortage, the Centre's tender for 10,000 graphics processing units is designed to pool high-end accelerators from global chipmakers and domestic data centres, offering them at subsidised rates to registered domestic startups and universities.

The target is straightforward: ensure that Indian researchers do not hit a computational ceiling when scaling training runs from billions of parameters to frontier-class reasoning models. By anchoring the physical clusters within national borders, the programme protects critical strategic sectors, from agriculture yield forecasting to defence logistics, from foreign cloud service dependencies.

What happens next

The IndiaAI Mission tender for computing capacity is moving into its vendor onboarding phase, with allocations expected to begin reaching selected labs in coming quarters. Meanwhile, BharatGen developers are preparing initial evaluation weights for government testing across health diagnostics, agricultural advisory desks, and court document summarisation.

International interest is also rising. Bilateral technology partnerships with the United States and the European Union increasingly frame India not just as a consumer of frontier systems, but as a primary co-developer of open, safe foundational tools for the Global South.