Jeff Bezos came out of retirement as CEO to build Project Prometheus, an “artificial general engineer” for designing and manufacturing physical products. It has raised roughly $18 billion at a $41 billion valuation and shipped nothing. Physical Intelligence is worth $5.6 billion without a commercial product. Skild AI sits at $14 billion. Funding for world models grew roughly fivefold in 2025.
All are chasing the same enormous prize. Unplanned downtime alone drains $1.4 trillion a year from the world’s 500 largest manufacturers, 11 per cent of their revenue. An idle automotive line burns $2.3 million an hour, or $600 a second. Eventually autonomous manufacturing will be enormous. But their approach of building world models is wrong.
History has been unkind to new technology in manufacturing, but this time it’s different
It has humbled smarter money before. During the Industry 4.0 wave, manufacturers ran eight pilots on average and scaled only 30 per cent. The rest died in what McKinsey called pilot purgatory. Preventative maintenance ran on tabular data in one corner of the plant, defect detection ran on camera feeds in another, and neither compounded into anything resembling intelligence. At best, Industry 4.0 could aggregate and visualise data, but not necessarily act on it. Stalwarts like PTC, Rockwell Automation, etc. have learnt their lesson the hard way.
Two things have genuinely changed since then. Sensors and edge compute are now affordable and available, and AI can close the loop from detection to decision to action. Massive pent-up demand is awaiting this useful technology – case in point: Tata Steel’s Kalanginagar plant runs 260 AI algorithms for real-time decision-making; JSW Steel pioneered an at-scale predictive maintenance platform across ~3,000 assets; Nucor partnered with C3.ai to optimise scrap and scheduling. Promising signals include plant-level digitisation budgets and leadership, active startup partnership programs, AI adoption as part of plant management performance appraisals, more active IT–OT partnerships, and a general hunger to use technology. It’s clear that the time for change is now.
World models will stall at the factory gate
General-purpose models work in the familiar 80 per cent of situations, and a small shift in conditions trips them. The gap is measurable: robots that hit 95 per cent accuracy in labs drop to around 60 per cent in real-world conditions. There is no internet-scale corpus of physical manufacturing data to close that gap. The data must be earned through deployment, machine by machine, slowly and expensively. Edge cases dominate, and edge cases are exactly what simulation cannot invent.
In manufacturing, anything less than 100 per cent is disqualifying. A chatbot hallucination costs an apology. A shop-floor hallucination can cost a safety incident, a scrapped batch, or massive production or reputation loss.
Look closely at the flagship deployments, and the gap between headline and reality widens. Figure AI’s celebrated BMW pilot totalled about 1,250 runtime hours and 90,000 parts handled before the fleet was retired back to headquarters. Fortune reported that for months the much-publicised “fleet” was a single robot practising pick-and-place during off-hours. Skild, the most commercial of the cohort, now argues that real-world deployment is what improves its models and has put robots on Foxconn production lines to get that data. The debate about whether deployment data wins is over. The race is about who holds the largest and most varied deployment surface.
The moat is data volume and entropy; this is where India shines
The playbook starts with a wedge, a niche deployment that pays for itself while generating real operational data. Predictive maintenance is the classic example. Each deployment feeds a flywheel: more deployments produce more data, which produces better vertical models and new use cases, which win more deployments. What compounds are vertical IQ 🡪 the accumulated knowledge of real failure modes, operator behaviour, schedule-power-quality swings, and Jugaad-modified machines that no lab can replicate.
No country has more entropy than India. Around 250,000 registered factories and 79 million registered MSMEs, with MSMEs alone producing over a third of manufacturing output. Manufacturing GVA is compounding at 11 per cent; the sector is projected to grow from $1.7 trillion toward $2.5 trillion by 2031, and the National Manufacturing Mission targets 25 per cent of GDP by 2035.
American factories are standardised and instrumented. Chinese data is closed. India’s messiness, the voltage swings and mixed-vintage machines and extreme SKU variety, is precisely the edge-case-dense data that world models lack. A model that survives an Indian shop floor generalises globally. The reverse is not true.
The playbook is already successfully running
At Avataar Ventures, we have seen the richness of Vertical IQ earned through niche deployment 🡪 flywheel 🡪 vertical dominance across our portfolio. Infinite Uptime does prescriptive maintenance for heavy industries, with data from over 900 plants and 100,000 sensors. It started in predictive maintenance and has expanded into energy savings and process optimisation, on its way to manufacturing automation. Chef Robotics builds AI robots for food manufacturing and has collected data from over 100 million servings across 80 robots. Deployment-first companies are out-collecting the headline names by orders of magnitude.
The last industrial revolution asked India to supply labour. This one asks it to supply something rarer: ground truth. The stealth labs can always raise multi-billion-dollar funding rounds to run more simulations. What they cannot raise is a decade of factory data. India is already generating it.
