Fabs take three years to build. The people who run them take fifteen
For four years, our semiconductor conversation was mostly paperwork. Policy frameworks, incentive structures, the occasional foundation stone with a photograph attached. This month it finally turned into a conversation about output. Three of the approved units are running commercial production. Twelve projects across six states have been cleared, carrying investment commitments of more than Rs 1.64 lakh crore. And a second phase has been approved with an outlay of Rs 1,27,500 crore. Anyone who has watched Indian industrial policy for a while knows how rarely a programme gets far enough for the argument to move on to what happens next. So why did this one move? Three things, as far as I can tell. The scheme stayed narrow at the start, going after fabrication and packaging instead of trying to fix the whole value chain in one go.
The process had some give in it, so a weak application could be sent back and improved rather than simply turned down. And every single project went up to the Union Cabinet, which meant nobody could quietly let it sit. A scheme that had six years to receive applications had committed nearly all its money in about three. Other missions could learn from that combination, though I suspect most will take the outlay and skip the discipline. Phase two is a harder brief, because it is so much wider: design IP, equipment, materials, chemicals and gases, more fabs, advanced packaging, research, skilling. You can arrange money for all of it.
You can arrange land. The two things that will actually decide how this ends are not for sale. Start with what self-reliance is supposed to mean, because we keep getting this one wrong. There is a pull in our public conversation towards reading strategic autonomy as making everything ourselves. Nobody in this industry does that. The Netherlands sits on lithography. Japan sits on chemicals and materials. Taiwan sits on advanced fabrication. Every one of them is dependent on the others, and none of them seems embarrassed about it. What you want is to be hard to switch off, which is a different thing from being alone. In practice that means buying equipment, partnering on process technology, staying open to people and capital, and building real capability only in the places where being cut off would actually hurt.
The same applies to AI, where the pressure points are compute, model weights and data rather than machines and gases. We lose time whenever we treat openness and autonomy as a choice between two camps. The second problem is course, not just engineering. One is that using AI should be taught horizontally. It is not a computer science elective. A law student, a nursing student, someone doing agriculture or architecture, each of them should finish their course knowing how these tools behave inside their own subject and, just as usefully, where they fall apart. people, and this is where we look worst. The design numbers are genuinely good. Something like 315 universities are now working with industry-standard EDA tools, about 68,000 students have been through chip design training, and over a hundred startups have access to the same software the global firms use. But a fab does not run on designers.
It runs on technicians, clean-room operators, process engineers, and maintenance staff. A lakh of them, roughly, working in conditions where one speck of dust ruins an entire batch. Our system is very good at producing degrees and quite bad at producing calibrated hands. Unless the ITI and polytechnic network gets rebuilt around real equipment and real factory partnerships, we will end up having bought the hardware and rented the competence to run it. Which brings me to students, and here I would ask for two things from every The other matters more and gets said less. When the machines are competent, what you are worth is your ability to notice when one of them is wrong. That takes knowing your subject well enough to feel that an answer is off, and knowing the slower route to the correct one so you can check.
A wrong answer delivered fluently and with confidence is going to cost people their jobs and occasionally their patients over the next ten years. Domain depth is the only thing standing in the way. That has awkward implications for how we examine people. As long as exams reward reproduction, students will outsource the reproduction, and who could blame them. Vivas, lab work, projects you have to defend, problems with real constraints attached: all of it is harder to set and much harder to grade, which is exactly why we avoid it. The programmes that take a student design all the way through to a working chip have this right. The learning is in carrying an idea to something that either works or embarrassingly does not. We have shown we can build the factories. Whether we can staff them with people who understand their field well enough to argue with their own instruments is the open question, and it will not be settled in Dwarka or Dholera. It gets settled in classrooms, and on that front we are already running late.
*The author is a tech and social entrepreneur, public policy commentator and columnist