India in the AI Race: Infrastructure, Ambition, and What's Actually Happening

In March 2024, the Indian government approved the IndiaAI Mission with a budget of ₹10,300 crore — roughly $1.24 billion. The stated goals were broad: build compute infrastructure, fund foundational AI models, develop domestic talent, establish AI governance frameworks, and position India as a global AI power rather than a consumer of AI products built elsewhere.

A year later, those goals are easier to evaluate. Some have moved faster than expected. Others reveal just how steep the starting position is.

The Compute Story

The headline figure from the IndiaAI Mission is 38,000 GPUs, made available at a subsidized rate of ₹65 per hour to startups and researchers. For context, that's roughly $0.78 per GPU-hour — well below what AWS or Azure charges for comparable compute. The access policy explicitly targets organizations that can't afford hyperscaler rates, which describes most Indian AI startups and nearly all academic researchers.

The scale matters. India's GPU deficit has been one of the most cited structural disadvantages in its AI development story. High-quality AI training requires massive parallel compute. Without domestic access, Indian researchers and companies either rented expensive hyperscaler time or, more often, operated at smaller scale than global competitors.

The 38,000 GPU figure doesn't close the gap with the US, China, or even the Gulf states that have been making aggressive sovereign AI investments. But it's not meant to. The intention is to create a viable domestic compute layer that makes serious AI research and development possible without full dependence on foreign infrastructure.

The private sector is moving faster. Reliance Industries is building a 1-gigawatt data center in Gujarat — expandable to 2 gigawatts — powered by NVIDIA Blackwell processors. The estimated investment is $20–30 billion. L&T has partnered with NVIDIA on what they're calling a "gigawatt-scale AI factory." These aren't research projects. They're infrastructure bets that assume AI compute demand in India will grow at a rate that justifies gigawatt-scale capacity.

BharatGen and the Foundational Model Question

The more contested question in India's AI ambitions is the foundational model question: does India need to build its own large language models, or should it build on top of models from OpenAI, Anthropic, Google, and Meta?

The IndiaAI Mission's answer leans toward yes — India needs its own. BharatGen AI was launched as India's first government-funded multimodal LLM, trained to support 22 Indian languages and designed to reflect Indian cultural context. The project runs through IIT Bombay.

The argument for domestic foundational models isn't purely economic. It's about representation. LLMs trained predominantly on English-language internet data don't perform equally across languages or cultural contexts. A model optimized for English reasoning about English-world problems is structurally disadvantaged when applied to the linguistic diversity of India — where there are 22 constitutionally recognized languages, hundreds of dialects, and large populations whose primary language isn't well-represented in global training datasets.

BharatGen is a partial answer. It's a research project, not a product, and it won't compete with GPT-4 or Claude on standard English benchmarks. What it might do — and this is the actual ambition — is enable better AI performance on Indic-language tasks and serve as a foundation for domain-specific applications in healthcare, agriculture, and education built specifically for Indian contexts.

NASSCOM's projection puts this in economic terms: IndiaAI could generate $500 billion in economic value and create 750,000 jobs. Those are optimistic figures with wide confidence intervals, but they reflect a serious assessment of what AI-enabled productivity gains could mean for an economy of India's size.

The Startup Layer

India's generative AI startup landscape grew 3.7x between 2023 and H1 2025, reaching over 890 startups. The application layer — companies building AI products rather than foundational infrastructure — grew 4x to cross 740 startups.

The concentration is in Bengaluru, with Mumbai and Hyderabad as secondary hubs. The sectors getting the most attention: enterprise software and SaaS (where India's existing software services talent pool is an advantage), edtech, healthtech, and fintech. These are categories where India has a large domestic market, established regulatory familiarity, and a competitive export story.

What's notably less developed: hardware, semiconductor design, and the physical infrastructure layer. India has ambitions in chip manufacturing — the government approved a $10 billion semiconductor scheme — but TSMC and Samsung don't have Indian facilities, and the timelines for meaningful domestic chip production are long.

The Talent Equation

India produces roughly 1.5 million engineering graduates annually. The quality is uneven, but the top tier is world-class, and competition for that talent is intense. A significant fraction still emigrates — NASSCOM's data consistently shows that Indian-origin engineers are overrepresented in AI research teams at US and European labs. Keeping and attracting that talent domestically is an unresolved challenge.

The IndiaAI Mission has invested in Centers of Excellence in healthcare, agriculture, and sustainable cities — research nodes intended to build frontier AI expertise in India rather than export it. The AI education budget in the 2025 Union Budget included ₹500 crore for a new CoE focused specifically on AI in education. Whether these investments create retention is an empirical question that won't have a clean answer for several years.

What "Winning" Would Actually Mean

The framing of AI as a race implies there's a finish line. There isn't. What there is: a period of competitive advantage for countries and companies that build serious AI capability early, and a much harder path for those that arrive late as dependent consumers of technology built elsewhere.

India's position is genuinely more complex than most single-country narratives allow. It has structural advantages — scale of talent, a massive domestic deployment surface, established software exports, and a government willing to spend on AI infrastructure. It has structural disadvantages — compute scarcity relative to the US and China, late entry to the foundational model race, and a regulatory environment that's still forming.

The $1.24 billion IndiaAI Mission is not going to build the next OpenAI. It might build the infrastructure for Indian companies to build on, and the research capacity to develop AI applications that serve Indian-specific needs well. That's a more modest goal — and one that's more likely achievable.

The more important question isn't where India ranks on a global AI leaderboard. It's whether Indian businesses and institutions can access the AI capabilities they need to be competitive, and whether the benefits of AI productivity gains distribute broadly across an economy where a huge proportion of economic activity still happens in informal and rural sectors that no current AI product is designed for.

That question is harder and more important than the compute count.