For 45 years after Independence, India grew at an anaemic 3% — the infamous “Hindu rate of growth”. Triggered by a balance-of-payments crisis, India found the gumption to liberalise. Within a decade, GDP growth surged. The lesson: crisis breeds reform, and reform breeds exponential outcomes. Artificial Intelligence (AI) provides the same transformational leverage that liberalisation provided in 1991. The question is not whether India can afford to invest — it is whether India can afford not to. After 12 years of Prime Minister Narendra Modi’s leadership, India has reached another WhatsApp moment in reform — dare I say, to take it from a Hindu rate of growth to the Bharat rate of growth — 8% and beyond for the next decade.The case for making tokens freeIndia has since proven that it can leapfrog entire generations of infrastructure. Aadhaar enrolled 1.38 billion people in a biometric identity system, the largest on earth. UPI now processes 250 billion annual transactions worth $3.4 trillion, handling 50% of the world’s real-time digital payments. Reliance Jio, launched in 2016, added 100 million subscribers in just five months and made mobile data essentially free. How India can replicate its Aadhaar, data availability deluge and UPI miracles in the age of AI — by embracing open models, diversifying compute hardware, and making tokens as free as data.India spends just about 0.65% of its GDP on research and development (R&D), well below China (2.4%), the United States (3.5%), South Korea (4.9%), and Israel (5.4%). The excuse of the socialist Congress governments of 1947-1991 was: “we can’t afford anything more”, which does not and should not apply now. The central question is this: Has the time come to make AI tokens free? If so, how do we do it? Can we begin by making tokens free for the country’s top 100 national R&D institutions and 5,000 schools, so that AI becomes a teammate for scientists and students alike?India spends approximately $49 billion each year subsidising calories, chemicals, and carbon. The question is whether it is now ready to subsidise cognition. The answer is that it already can — at a fraction of what it spends today. The entire AI token subsidy for the country’s top 100 universities and national R&D institutions, along with 5,000 high schools, would cost about $2 billion annually — roughly 0.06% of GDP. The AI investment is around one-fourteenth of India’s food subsidy, one-tenth of its fertilizer subsidy, and less than the amount paid to compensate oil marketing companies for LPG under-recoveries in a single quarter. This is not a moonshot budget. It is a rounding error in India’s existing welfare expenditure — but one with the potential to generate transformative, compounding returns.India does not need more money; it needs new priorities, new frameworks, and some reallocation. First, AI tokens can be funded without cutting a single rupee from existing beneficiaries by merely freezing subsidy growth for a year. (The fertilizer subsidy alone has been compounding at 11% CAGR over the past decade.) Second, India should forge a public-private partnership with hyperscalers such as AWS, Google, and Microsoft, exchanging data centre land, power subsidies, and data sovereignty assurances for free inference capacity. India’s 1.4-billion-user market gives it negotiating leverage that few countries can match, though such techno-commercial negotiations, shaped by geopolitics and uncertainty, will require considerable skill. Third, if necessary, paid enterprise tiers can cross-subsidise free access for schools and research institutions.Free data worked and free tokens will too. The government did not directly subsidise Jio. Instead, it created the right regulatory environment through spectrum auctions, net neutrality, infrastructure sharing, and a controversial long pilot.The result: data prices fell from about $3 per GB to $0.10 in under three years. The same playbook can apply to AI tokens. Create the right conditions, and the market will deliver abundance. India’s role is not to write the cheque, but to make it unnecessary.Hosting modelsIndia must build the competence to host and operate large language models, not merely consume them through APIs from San Francisco or Beijing. With Sarvam, it has already shown that frontier models can be trained on Indian soil. The next step is to host models such as Qwen, DeepSeek, Kimi, Llama, and Sarvam on sovereign infrastructure, even if some open-source models originate from China. The models may be free, but the intelligence to host them at scale is not. AI infrastructure should be treated as a strategic national capability, much like what the government has already done for space and nuclear programmes. India should pursue a hybrid strategy that combines sovereign and open-source models to avoid dependence on any single ecosystem. Open-source models offer four key advantages: sovereignty by reducing reliance on foreign APIs that can be restricted overnight; lower costs by eliminating per-token licensing fees, thereby enabling free education and research uses; customisation for Indic languages and cultural contexts; and transparency through auditable model weights for government and defence applications.Hosting LLMs at national scale is not “just deploy a container”. India must develop deep expertise in availability (multi-region redundancy, 99.99% uptime SLAs for critical government services; latency (<200ms token latency across Tier-2/3 cities); efficiency (like batch scheduling to maximise tokens-per-watt); security (data residency compliance, prompt injection defence, audit trails).India cannot afford to be locked into a single-vendor monopoly for its AI compute future. NVIDIA’s dominance comes at extraordinary cost — both financial and strategic. NVIDIA controls 80%+ of AI training hardware but creates deep vendor lock-in that extends far beyond hardware. For India to build sovereign AI infrastructure for 1.4 billion people, the math simply does not work at NVIDIA pricing.India spent relatively little to build UPI; it cannot afford to spend $50 billion on NVIDIA GPUs. The alternative hardware ecosystem is not just more economical — it is strategically essential. India should adopt a 40:30:30 hardware mix rather than rely on a single vendor. Around 40% should be built on AWS Trainium and AMD for cost-effective inference workloads available in India. Another 30% should use Google TPUs for research, model training, and academic partnerships. The remaining 30% should rely on NVIDIA for specialised training, legacy compatibility, and a domestic silicon hedge.AI token policy, implementation timelineIndia should announce a National AI Token Policy and implement it over the next 24 months. It should begin by signing public-private partnership agreements with AWS, Google, and Microsoft to establish a multi-vendor sovereign compute framework. The first pilot should provide unlimited research tokens to the top 20 IITs and the IISc. As hyperscaler inference capacity comes online, India should open an API sandbox for 500 startups, expand access to 100 universities, and launch an AI literacy pilot across 500 high schools in 10 States. The programme should culminate in the publication of the country’s first sovereign Indic AI model benchmarks.Thereafter India needs to scale the ecosystem through cross subsidies as mentioned above and deploy fine-tuned models into health care, agriculture, judiciary and education. It must follow this up with full deployment at 5,000 high schools and all 22 languages. With this India’s token consumption will enter the global top five. In two years, India-trained models will become competitive on international benchmarks and 10,000+ AI-native startups will bloom, rocketing GDP growth.In sum, India’s AI leapfrog is a sine qua non for national transformation. India needs to replicate its DPI (digital public infrastructure) miracles in the age of AI by embracing open models, diversifying compute hardware, and making tokens as free as data. India has all the necessary conditions — favourable policies, stable macroeconomy, prodigious talent and technology institutions — to become a global leader in AI applications. The sufficient condition is India’s decisive, visionary leadership. It will take three moves. One decade. And a transformed (AI) nation.Srivatsa Krishna is an IAS officer. The views expressed are personal