Last week, on a 600-acre site near Visakhapatnam, Google broke ground on what is being dubbed as India’s largest AI hub. The infrastructure partners are AdaniConneX and Airtel’s Nxtra. The silicon will be designed in California and fabricated in Taiwan. The model weights, when the data centre lights up, will be Google’s. India’s contribution is the land, the power, the construction, and the operations. It is a triumph of sorts. Alas, we have done something like this before.
Thirty years ago, the IT-services boom began with a related bargain. India offered low-cost engineering labour, fluent in English, to American firms that needed someone to write and maintain the systems they had designed. The arrangement worked. Infosys, Wipro and TCS came of age. A middle class emerged. A reputation was earned. Indian services climbed the value chain, and the global capability centres in Bengaluru and Hyderabad now do genuine research and product work for the world’s largest firms. We were not wrong to celebrate this. But the task was incomplete.
What we did not do was use the proceeds to build the next layer. We did not fund research universities of global rank. We did not raise national R&D intensity from below 1 per cent of GDP, where it has remained for decades. We did not develop deep-tech venture capital, manufacturing depth, or patient state-aligned industrial finance. Korea built Samsung. Taiwan, smaller than Karnataka, built TSMC, which now sits at the chokepoint of the global AI economy. India built outsourcing. We learned magnificently to deploy systems that other people had designed. We postponed the work of designing them.
The bill on that postponement is now arriving, in the form of artificial intelligence. The leading edge of AI demands capital, research depth, and concentrated talent at a scale India has not built. A frontier model costs hundreds of millions of dollars to train, draws on decades of accumulated research, and depends on a talent pipeline that runs through a few dozen universities in the United States, China, the United Kingdom, and Canada. India has none of this.
In the past fortnight, two voices have read this moment differently. Ruchir Sharma, at the Express Adda, warned that global investors now see India as the “anti-AI play”. Nandan Nilekani, in a recent essay with Ravi Venkatesan, argued that India’s edge lies in diffusing AI across its small firms and informal economy. Sharma is reading the high-rent layers and seeing India’s absence. Nilekani is reading the diffusion layer and treating it as India’s destination. Both stop short of the harder question. Neither asks how a country with talent and ambition becomes a producer of intelligence rather than a tenant of it.
What we are doing in Vizag, and in Hyderabad and Mumbai and Pune, risks making the same trade we made in the nineties, with stronger lock-in this time. The data centres are real and useful. They will be optimised, for the next decade, to serve foreign models efficiently rather than to train Indian ones at scale. Every Indian-language inference running on a foreign model from a server in Vizag is a small recurring rent paid to Mountain View, on infrastructure built by Adani. The arrangement is not new. The technology is.
The stakes go beyond productivity. As AI diffuses into military systems and strategic decision-making, this gap will bear on India’s strategic autonomy in a world where security partnerships cannot be assumed. India does not need to beat OpenAI at frontier reasoning today, but it needs to learn to compete: Alternative model architectures and approaches at meaningful scale, sovereign compute, and foreign deals structured around expanding Indian capability rather than only power and land. The Indian state knows how to create a gigawatt of power for a foreign data centre. It needs to learn how to create a research department, a laboratory, a generation of scientists. Concrete it understands. Cognition it does not.
The Economic Survey of 2025-26 noted that only 2 per cent of the world’s AI training-data startups are based in India, against 40 per cent in the United States and 21 per cent in the European Union. The state knows the gap. The announcements still say far more about infrastructure than about capability.
The deeper danger is that India’s public conversation is too quick to convert constraint into strategy. Because we cannot yet win at the frontier, we tell ourselves the frontier does not matter. Because we are good at diffusion, we tell ourselves diffusion is enough. Some of this is an honest assessment of what India can do. Some of it is an alibi for what we will not do.
We must not forget, however, that even diffusion rests on uneven foundations. UPI worked because Indian banking had spent 30 years digitising, Aadhaar because the state had been identifying citizens for 60. Where those rails were missing, in health, in school education, in agricultural extension, the same stack thinking has produced architecture without adoption. AI diffusion may run into the same problem.
This past week in Visakhapatnam, we poured the foundations of buildings that will host someone else’s intelligence. The buildings will be ours. The intelligence will not. Whether that is enough this time is the question we should ask before the concrete sets.
Lamba is assistant professor, Cornell University and Rajan is professor, University of Chicago and former governor, Reserve Bank of India
