India’s Arrakis problem: The price of what low-skilled workers know

It is heartening to see Indian workers helping generate the data that will drive the next generation of robots and artificial intelligence. Every month, a fresh batch of images does the rounds on LinkedIn and X: Indian factory employees with cameras strapped to their heads, their every movement recorded to train AI systems. The message? Tasks once thought resistant to automation are suddenly within AI’s reach. But these images of Indian workers wearing cameras should provoke something more serious than technological excitement or dystopian alarm. They should provoke a strategic question: Who owns data? Who builds models? Who controls these machines, and where does the resulting value accumulate?

The instinctive reaction, of course, is enthusiasm, and these images are easy to celebrate. Yes, AI training is happening in India, and Indian workers are part of it. Gig, factory, and service workers are capturing ordinary physical tasks from a first-person perspective, and thousands of hours of Indian movement are becoming training material for the next generation of AI. Participation is not however synonymous with power. The important question to ask is not whether artificial intelligence is being trained in India. It’s what India is contributing to the process, on what terms, and who will ultimately own the productive assets created from it.

Also read | Your boss may soon know what you’re asking Claude to do at work


Several reports show that filming factory workers is fast becoming normal across Indian cities. In Gurugram and Nagpur, for example, workers are being recorded as they sew, adjust loom tension, and clear machine jams, or are simply under surveillance as they work. Similar data collection is spreading across the service economy. Human Archive, a Silicon Valley-based physical AI data startup, says it has deployed more than 1,000 headsets through partners in India and pays a base rate of $1 an hour for egocentric data collection, acknowledging that its Indian network helps keep compensation lower. Put simply, India’s cheap labour is becoming cheap machine-training data, too.

The ambitions extend far beyond individual experiments and individual startups. EgoLab, an Indian startup which has been collecting first-person footage from Indian factories, was acquired in 2026 by US-registered Build Artificial Intelligence Inc. EgoLab's own material, reviewed by scroll had described their ambition to collect $100 million in footage from 100 million factory workers across major Indian states and sectors, including textiles, automobiles, chemicals, electronics, FMCG, and steel. BuildAI's published current roadmap similarly includes an ego-centric 100-million-dataset target for October 2027.

These are not outcomes, but these ambitions reveal the intended scale of how India's data is being converted into a dataset that can be aggregated, transferred, and used globally. A worker wearing a camera records two outputs. The immediate output is the product of labour and information about how that labour is performed. In the future AI economy, the work is just not generating output; it is also recording the work, which can generate capital in the future.

Also read | Zoho’s Sridhar Vembu sees a painful rebalancing of jobs in the AI age

This is very relevant for a country where 88% of the workers are employed in occupations classified as low competency. This classification may make sense within the conventional labour market, but the same task has become extraordinarily valuable from the perspective of physical AI. Large language models could learn from vast stores of digitized text, images, and code; Physical AI there is no equivalent archive for much of what humans know about operating in the real world. Robotic Machines still struggle with actions people perform almost automatically: adjusting grips and pressure, manipulating unfamiliar objects, coordinating hands and tools, recovering from mistakes, and adapting when physical, emotional, and other conditions change. Therefore, what is routine knowledge for a human worker can become frontier knowledge for a machine. India's workforce is not only a source of inexpensive labour; it is also a potentially enormous reservoir for the demonstration that physical AI still needs to learn how the world works. However, a country that built its IT boom on labour-cost arbitrage, in practice, introduces a striking inversion: cheaper labour once reduced the incentive to automate. Now, if human demonstration can also be acquired cheaply, low wages may also help reduce the cost of teaching automation itself.

In other words, India’s cheap labour no longer simply competes with automation. It could also cut the global cost of training robots and large language models. The point to contemplate is not merely whether machines will displace workers. It’s whether the knowledge workers bring becomes a strategic resource whose value is realised elsewhere. History isn’t reassuring here and we know that economies that supplied essential raw materials often had little control over the industries built around them, and their competitiveness suffered. That’s why Frank Herbert’s Dune offers such a valuable metaphor.

In Dune, Arrakis, a barren desert planet, possesses spice, the indispensable resource on which an advanced civilisation relies for interstellar travel. However, owning this resource does not automatically mean dominance over the system built around it. Instead, others coordinate its extraction, trade, and conversion into power. India faces a version of that dilemma in the emerging physical AI economy. Its spice is the vast stock of embodied knowledge held by its workers - the accumulated judgement, dexterity, and experience in how tasks are performed. Indian companies can collect and annotate worker data but risk missing out on valuable layers if India remains just a raw knowledge supplier, as key datasets, models, robotics, software, platforms, and machines are mainly owned elsewhere. India is repeating an old development pattern in a new technological form.

India’s previous outsourcing boom was built on supplying labour to the West. Physical AI raises the prospect that India will now also supply the knowledge embedded in that labour, captured as data and used to train robots, machines, and large language models worldwide. The strategic challenge for India, then, is to make sure it doesn’t remain merely the source of that input. That means guarding its data privacy and sovereignty, while also developing its own models, robotic capabilities, and intellectual property built on that data.

India doesn’t need to stop the flow of spice. It just needs to make sure it doesn’t become merely Arrakis.


(Amit Kapoor is chair & Sheen Zutshi isresearch manager atInstitute for Competitiveness.).