[Through BCL India] An Inquiry — Part I

You can download the full report here:

https://bclindia.in/wp-content/uploads/2026/03/Report-on-AI.pdf

Excerpts:

On the chip subsidisation program:

Of the total ~INR 10,372 crore allocated to the mission, “India AI Compute Capacity” is allocated ~INR 4,600 crore, roughly 44% of the total. What began with a target of 10,000+ GPUs has now almost quadrupled, to 38,000+ GPUs, (with an additional 20,000 GPUs committed in the AI summit, bringing it up to ~60,000 GPUs ) being made available at ‘one-third’ the global cost, and ‘unlike many countries where big tech controls GPU access’.

While this is a sensible enough proposition, one must recognize that subsidising procurement to chips that are controlled, in their entirety, by a global supply chain (which in turn is subject to geopolitical volatility) is a transitory fix to a complicated problem.

A problem that exists, and will intensify at the interaction of inordinate concentration of market power, mercurial government policies on export control, and an unprecedented, uncertain upsurge in demand. India does not make chips of its own and while Budget 2026 has indicated further government intent with the India Semiconductor Mission 2.0 being given an a further outlay of Rs. 40,000 crore, any possibility of a sovereign supply chain is distant. To that extent, the declaration that India’s compute infrastructure is not controlled by ‘big tech’ is… well, ironic.

On the case for small language models

For the boundaries within which India’s AI story has to take origin and evolve — scarce compute, fledgling research ecosystems and almost prohibitive diversity — carving out compact models for precise problem statements that don’t need data center acreage seems infinitely more plausible, grounded in reality and solution-oriented than any attempt to build a frontier model with a wide sweep. And given their ability to run on private servers (or on edge), these models are inherently architected to be secure, allowing safe deployment critical sectors like healthcare and defence. All put together, maybe India’s AI story scales new heights by starting small.

On the case for regional language models

When one is asked to name mankind’s three greatest inventions, language is probably not at the top of that list — but maybe it should be. If one acknowledges that the act of expression is intrinsic to being human, the systematic erosion of language that has already been caused by the overwhelming proliferation of English should be cause for concern. And this erosion will only be exacerbated when technology that is purported to change our lives as we know it is predominantly trained in English. Expression emerges from deep cultural context — and even for technology that is designed to be as patronizing as AI, in the absence of such context, global (in its truest sense) penetration will fail.

After all, like Turkish artist Refik Anadol said at the WEF Meeting in Davos, “How on Earth can we create an AI that doesn’t know the whole of humanity?”

On India’s positioning in the data race:

Data has been the currency of the global economy for a while now — but its importance in an age where generation of statistically probable text has been made out to be singularly, critically important cannot be understated. While research and development in generative artificial intelligence is already hurtling towards the use of synthetic data (artificially generated information that mimics real-world data) as a workaround for data gaps (The Epoch AI research team projects, with an 80% confidence interval, that the current stock of training data will be fully utilized between 2026 and 2032) and privacy laws alike, use of synthetic data may come with significant downsides, including, not insignificantly but perhaps disturbingly, autophagy —‘a phenomenon (and a future?) where generative AI systems may increasingly consume their own outputs without discernment, raising concerns about model performance, reliability, and ethical implications.’ And that may mean opportunity for India — after all, for a country as large and diverse as ours, ‘becoming the data capital’ of the world may well be the lowest hanging fruit.