India's AI opportunity lies beyond the race for bigger models

A child learns to recognise a chair from a handful of examples; today's most advanced AI systems require enormous computing power to achieve comparable generalisation. India should compete in the AI race, but also ask the deeper question: what does the 20-watt human brain know that our machines do not?

A child can learn what a chair is from a handful of examples. Give a child a few chairs of different shapes, sizes and colours, and soon the child can recognise a chair they have never seen before. They can distinguish a chair from a table, infer that a stool belongs to the same conceptual family, and apply the idea to objects in unfamiliar settings. Human intelligence generalises from remarkably little data.

Now consider what it takes to reproduce even a fraction of this general-purpose intelligence artificially. OpenAI's latest frontier model, GPT-6 Astra, was developed through the company's largest training run, involving more than 100,000 GPUs at its Stargate facility, according to reporting on its launch. OpenAI describes Astra as its most capable model yet, with state-of-the-art performance across reasoning, computer use, coding, science and professional work.

The contrast is extraordinary. A human brain, weighing roughly 1.3 kilograms, operates on approximately 20 watts of metabolic power. It learns continuously, generalises from sparse experience and adapts to situations it has never encountered. We can now build artificial systems of extraordinary capability, but achieving that capability requires industrial-scale computing infrastructure.

This is not merely a technological curiosity. It may be one of India's most important strategic opportunities. The global AI race is increasingly becoming a race of scale. The largest companies are accumulating GPUs, data, capital and engineering talent at a rate that few countries or institutions can match. India should absolutely participate in this race. We need computing infrastructure, foundational models, chips, data centres and world-class AI researchers. But there is little strategic value in deciding that our ambition is to become the second-best version of something that someone else has already built. There is no enduring national advantage in being second at the same game.

India needs to investigate a different question: why does biological intelligence achieve so much generalisation, abstraction and adaptability with so little energy and so little experience?

We still do not have a satisfactory scientific answer. We understand many components of the brain and have made enormous advances in neuroscience, but we do not possess a comprehensive theory of how intelligence emerges from them. We do not fully understand how the brain forms abstractions, transfers knowledge across domains, learns continuously without catastrophic loss of previous knowledge, or constructs useful models of the world from relatively sparse experience.

These are precisely the questions that could produce disproportionate advances in AI. A better understanding of biological intelligence could lead to fundamentally different approaches to learning and computation. A new principle for generalisation could matter more than another generation of larger models. A radically more efficient architecture could matter more than another expansion of computing infrastructure. The largest opportunity may therefore lie not in scaling today's paradigm, but in discovering what today's paradigm is missing.

This is a difficult research programme, and that is precisely why universities should pursue it. Markets and companies naturally favour research with identifiable commercial returns. Fundamental questions about cognition, abstraction and consciousness may take decades to resolve and may produce no immediate product. Yet some of the most consequential technologies in modern history emerged from research whose eventual applications were impossible to predict at the time.

Modern neuroscience has developed increasingly powerful instruments for observing the brain from the outside. Perhaps the next advances will require us to combine these measurements with much more systematic investigation of subjective experience from within. It requires recognising that a complete science of cognition may need methods capable of connecting first-person experience with third-person measurement. This could become a distinctive Indian research programme: neuroscience, cognitive science, AI, mathematics, philosophy and contemplative practice brought together around one fundamental question: what are the computational principles of intelligence?

The objective should not be to build an Indian copy of GPT-6 Astra. By the time we reproduce today's frontier, the frontier will have moved again. The objective should be to understand something that the frontier itself does not yet understand.

India should therefore make two bets. We should invest aggressively in the infrastructure required to compete in today's AI economy. Alongside it, we should create long-horizon research programmes willing to pursue questions where the probability, timing and form of the payoff are unknown.

The second bet may turn out to be far more important. The 20-watt brain is telling us something profound. Intelligence does not necessarily require enormous amounts of computation. We simply do not yet understand how nature achieves it.

India's opportunity is to study that gap. If we spend the next decade trying to catch the leaders, we may become very good at following them. If we spend part of that decade discovering a fundamentally different theory of intelligence, we may have a chance to lead.

Modern neuroscience has developed increasingly powerful instruments for observing the brain from the outside. Perhaps the next advances will require us to combine these measurements with much more systematic investigation of subjective experience from within. It requires recognising that a complete science of cognition may need methods capable of connecting first-person experience with third-person measurement