India is betting big on data centres but is this the best path to success with artificial intelligence?

For hardware, we should indigenize AI chip design, fabrication, assembly and testing; and for software, we should create our own AI models. This is the only way to guard against being at the mercy of foreign suppliers.

That we form a vast market for AI services—estimated to be the world’s third largest—should not impress us. Nor the fact that we generate diverse troves of data that AI businesses hunger for. Several cutting-edge AI developers have offered us their services either cheaply or free precisely to acquire a massive user base that would give them untold giga-loads to refine their models further.

As social media has shown, we are among the planet’s most prolific generators of data. Not only would it be hard to create enough capacity to store and process what we ourselves spout, there is little to be gained from being a storage utility provider to the rest of the world.

Consider the environmental cost. Data centres consume huge quantities of electricity. While mega projects claim they would run on renewable energy, a boom in overall power demand is likely to make it that much harder to decarbonize India’s economy. Granted, mere data storage is not too power-intensive. But today’s AI race requires a lot more to be done with data—from training models to inference operations—that guzzles electricity.

High-end processing is presumably what the latest data projects are aiming for. These would require advanced processors of the sort made by Nvidia (its Rubin family for example), apart from high-bandwidth memory chips. Yet, as various scenarios of India’s AI evolution plotted by Economic Survey 2025-26 show, we may face the same hurdle in each case: a chip shortage.

Given such constraints, can we make headway with frugal AI? The survey’s chapter on AI recommends a focus on small language models and specialized apps rather than large foundation models. Sarvam, a local model for Indian language text and voice applications, is a good example.

Open-source models from the US and China could also be adapted for local needs. But none of this will suffice for purposes of national security. Even if an open-source model has open weights that can be freely tweaked, its developer might still be able to hold users ransom. For AI sovereignty to be more than a buzzword, we must indigenously develop the key models we cannot do without.

India is one of the few countries in the world with potential talent for innovative AI: a large supply of young people who can be trained in advanced linear algebra, calculus and probability. Another group could work on developing the chips that perform parallel data processing for AI use. China’s success with fewer resources than the US shows that determination can make a difference.

It might be useful to think beyond foreign partnerships and subsidies to attract global majors, and instead deploy that money to fund local research and startups that hold genuine promise. We clearly need to catch up on AI. But for that, our infatuation with data-centre proliferation must yield to inspiration in the realm of original thinking.