Hardware Comes First; Software Must Refine What We Build

As India pursues ambitious goals in semiconductors, advanced manufacturing, artificial intelligence, and deep-tech innovation, a crucial question remains: can technological independence be achieved without building strong foundations in hardware, materials, and core engineering sciences? In this thought-provoking conversation, Dr. Prateek Gupta, Assistant Professor in the Department of Applied Mechanics at IIT Delhi with Education Post’s Prabhav Anand, offers a candid perspective on the realities of India's research and innovation ecosystem. Drawing on over a decade of experience in high-performance computing, computational fluid dynamics, and computational materials science—including research at Purdue University and ETH Zurich—he argues that simulations and software alone cannot deliver technological leadership. From semiconductor manufacturing and AI to research funding, academic incentives, and indigenous hardware development, Dr. Gupta presents a compelling case for reorienting India's innovation agenda toward building real engineering capability, robust experimental infrastructure, and globally competitive deep-tech ecosystems.

1. India is focusing a lot on becoming self-reliant in areas like semiconductors, advanced manufacturing, and energy technologies. In this journey, how important do you think computational engineering and high-performance computing are for building real technological independence?

Computational engineering is a tool in service of building technology and not the eventual destination. Manufacturing capability, experimental infrastructure, and the physical & empirical knowledge that comes from operating both are the pillars of real technological independence. I say this with some self-awareness. Computational research dominates Indian academia because experimental apparatus is perceived to be expensive and with limited access. I have spent most of my career doing computational work and only recently begun building experimental facilities. In some sense, that is a reasonable response to valid constraints also, not a chosen path.

It is important to highlight that this is a response to systemic constraints and not particularly economic constraints. However, computational research that matters and has an impact, is actually more expensive than experiments in India. A comparison I find clarifying is (based on my 5 year experience), one can build a decent compressible flow experimental facility (nothing large-scale as a real supersonic wind-tunnel) at a cost of 2-5 crores INR. A single compute node with out-of-date 128 processors costs around 20L INR from Indian resellers fitting Intel/AMD chips in so-called Indian compute nodes. In 5 crores, one would probably be able to purchase 15 such nodes with an additional >20L INR for the interconnect.

Rest of the cost would be required to setup environment conditioning for their optimum performance: dust collectors, precision Air conditioners, dehumidifiers. That amounts to around 2000 cores working together. Peanuts for a cluster! You could, at best, solve a boundary layer transition direct numerical simulation of a plate as long as, wait for it, 3 cm! So contrary to the popular belief, I think more bang-for-thebuck still rests with experimentalists in India. Of course, one can argue that we should purchase community computational clusters, and that brings down the cost per user. However, even then, due to annual maintenance charges, these things are expensive. Yet, we see more people (even I did in my initial stages) choosing computational research over experiments. The only way to reduce this gap and also to do computational research with impact is to build the hardware and use computations to solve refinement problems in hardware. If hardware is developed in India, then computational research will also be more accessible and research with more impact can be conducted.

Hardware comes first. After all, planes were flying well before computational fluid dynamics (CFD)! So to summarize, I think the road to technological leadership is through hardware and using software to refine hardware.

2. A major part of your research involves multiscale modeling and computational simulations. How can this kind of research help solve real industrial challenges in India, especially in sectors like semiconductors, energy systems, and advanced materials?

Multiscale simulation is most valuable when running ahead of a hardware program. It is useful in reducing experiments needed, identifying failure modes early, compressing development timelines. It becomes much less valuable when it is the program itself, because the hardware investment has not been made.

In semiconductors, we are doing electronic design automation. But process simulation such as modelling dopant diffusion, thin-film stress, etch profiles, etc is what allows a country to understand and improve a fabrication process. That layer of capability requires simulation tools and experimental process development facilities to validate against. One without the other produces knowledge that cannot be trusted in production.

The same logic applies across energy systems and advanced materials. Simulation validated only against other simulations is a closed system, internally consistent and potentially wrong about the world. There are many proposals upcoming encouraging industry and start-up engagement. There is increased intent to enable ease of research as well. It remains to be seen how it pans out. To summarize, computational research which is not being traced by hardware development is an empty run.

3. India produces a large amount of engineering research every year, but only a small part of it reaches industries or becomes real-world technology. According to you, what is stopping research from moving successfully from laboratories to the market?

This is a structural problem. Many like to blame academia, others like to blame industry. Both deserve some slack (or are maybe equally to blame, if blaming alone solves a problem!). The structural problem is a mismatch of timescales and incentives. Academic careers are evaluated on publications, which peak in value at Technology Readiness Level 3 (proof of concept). Industry needs TRL 7 or higher. The gap between them is expensive, slow, unglamorous work. We still like to gauge our researchers and their prestige as per “how many international journal publications or citations have been garnered” not as per “how did this research help Indian industry”. No one wants to recognize that doing something which has been done worldwide (take semiconductor manufacturing for example), in a place where it hasn’t been done is also innovation. Homegrown solutions are sustainable and must be given proportional value. We like this idea in diet, but not in technology.

There is another more fundamental issue that rarely gets named. A significant fraction of Indian academic research, particularly the AI and ML work that now dominates publication counts, has not seen actual translation. A paper that benchmarks a model on foreign hardware, using a foreign framework, contributing to a foreign model ecosystem, is structurally disconnected India. The volume of such work is large. Its contribution to genuine technological independence is negligible.

We need products that work physically, engines, avionics, fabrication tools etc. Premier Indian institutions file patents at impressive rates, but we still fly aircrafts with engines and computers manufactured entirely abroad. The gap between IP creation and physical product is enormous and largely unaddressed. There are many steps being taken in the right direction (such as ANRF MAHA programs); what they need is longer grant durations, larger team sizes, IP protection, and success metrics that include deployment, not publications or patent filings alone.

4. Your work covers areas like compressible flows, nonlinear systems, and advanced computational methods. Do you feel India is investing enough in core engineering research, or are we becoming too focused on short-term technologies and trends?

India spends roughly 0.6–0.7% of GDP on R&D, flat for over a decade, while China crossed 2.5% and South Korea approaches 5%. The absolute amount has grown, but as a fraction of a growing economy, we have stood still.

Within that constrained envelope, the inclination toward AI and machine learning has become, I would argue, actively damaging. The dominant mode is not AI research, it is AI consumption. Students fine-tune existing models, run benchmark comparisons, and build applications on frameworks and hardware stacks built entirely outside India. This produces publications and well-paying industry roles, but it builds no sovereign capability. Mind you (the reader), these are industry roles for foreign industries. A graduate who has spent five years optimising on foreign GPU infrastructure is, at the end of it, a skilled user of technology India does not make and has no near-term path to making. That is not an engineer in any meaningful sense of the word.

What gets crowded out is the work that actually matters for hardware independence: fluid mechanics, materials science, thermodynamics, solid mechanics, the physical understanding that underpins every system India wants to build or manufacture. These fields take decades to build. When a strategic industrial window opens and the foundational knowledge base is not there, the opportunity passes and rebuilding it afterwards costs far more. We have seen this before. Picking immediate returns is not culture, by-the-way, it is basic arithmetic. Talented researchers make this choice because every one wants to be recognized. The system that rewards recognition on poor development is at fault.

5. Artificial Intelligence is now entering almost every field of engineering and scientific research. How do you see AI changing computational engineering and simulations in the coming years? Can AI speed up innovation, or are there risks in depending too much on it?

As the name suggests, it is Artificial intelligence. Most models are not inherently intelligent, they mimic intelligence – to the best of my knowledge. The bitter truth is, most tasks in our everyday life can be navigated through that mimicry. Personally, I have been using tools for coding for 6 months now, and for an avid coder who likes to learn by doing, and who does not have access to a large competent team (or actually even a single like-minded collaborator in computational sciences) who wants to devote time in actual scientific coding (I do not call “import numpy as np” in python as coding), the modern AI tools have been great for me. I got to learn good coding practices, efficient profiling of scientific codes, and so much more. However, unfortunately for me, that is where it stops, at least so far. These tools are great coding partners. Companies acing them deliberately made these tools to ace at coding, because that is the basic rule of civilization right? Become good at reproduction first. So a code generating good code is the best thing to do. Thinking that centuries of scientific knowledge gained over a time with unprecedented amounts of energy, can be done in a very short time scale by accelerating this reproduction, at the moment sounds like a pipe-dream to me. Think about the cumulative energy required to generate all scientific knowledge since – say – Newton. Do we really think that we can do that in a data center without damaging the environment? I would be very much interested in looking at such a calculation from a basic thermodynamics point of view. And if the numbers add up, why not? Ranting aside, AI can speed up development of tools. Current state-of-the-art requires an expert in the loop. Without an expert, these tools are dangerous because they sound like that friend at the bar who has no idea what they are talking about, but surely sounds confident!

Another more important point for India is this: the dominant mode of AI research in academia right now is consumption, not creation. PyTorch and TensorFlow run on NVIDIA GPUs, fabbed at TSMC on ASML machines — none of which India makes or has a credible path to making. Enthusiasm about AI research that runs entirely on this stack is, at its core, enthusiasm about becoming better users of foreign infrastructure. That is a legitimate economic activity but it is not technological independence. The AI research that would serve India's long-term interests is faster development of validated and verified scientific tools which aid in materials discovery, process optimization, and hardware design. The emerging photonic and neuromorphic computing paradigms (for example) might disrupt the GPU-dominated AI stack. Building toward those platforms, rather than continuing to optimise on foreign hardware, is where AI research and hardware ambition need to converge.

6. When people talk about India’s semiconductor mission, the focus is mostly on manufacturing plants and investments. But how important are simulation, modeling, and computational research behind the scenes for building a strong semiconductor ecosystem?

The focus on fabs and plants is an absolutely right direction — they are the actual goal. The simulation question is about whether the manufacturing capability, once built, will be genuinely understood and owned, or merely operated on foreign knowledge. Process simulation is what allows a fab to develop new process nodes rather than only running licensed recipes. The1.60 lakh crore INR committed across approved ISM projects is serious hardware investment; the parallel investment in the modelling capability that lets Indian engineers understand what happens inside those fabs is where the gap remains.

But the conversation that is almost entirely missing is raw materials. India imports approximately 90% of its rare earth compounds, overwhelmingly from China. For conventional silicon this is a known vulnerability. For next-generation hardware such as neuromorphic chips, the materials dependency might be even more acute, and domestic processing at electronics grade is essentially nonexistent.

The National Critical Mineral Mission and the rare earth permanent magnet scheme are a beginning. What is needed alongside them is a research program that treats materials processing as a first-class engineering science problem — using simulation to predict processing routes, optimise purification methods, and identify domestic substitutes — directly coupled to the extraction and processing pipeline. Simulation in service of building the materials supply chain, not as a substitute for it.

7. Through your work at IIT Delhi, you have been closely involved with advanced computational research. Looking ahead, which area do you think has the biggest potential for India in deep-tech innovation — advanced materials, energy systems, AI-driven simulations, or something else?

To be honest, I do not consider most of my work to be advanced computational research in the frontier sense. In fluids, I write codes that have been written worldwide, run on bigger machines, for more challenging problems. We like to use these codes to answer fundamental questions about fluid mechanics which remain un-answered. I do face skepticism from fellow researchers: why are you writing codes? What is the point? The point, to me, is twofold. First, to train the next generation of researchers who know how to build new open-source community solvers from scratch. Second, and more personally — solving a fluid mechanics problem through a solver you built yourself provides physical insights that are almost never accessible when the goal is simply to run someone else's code. Alas, I still have to do this on non-indigenous processors. In computational materials science, in collaboration with ETH Zürich, I did work that I think genuinely qualifies as advanced computational research. How that develops and gets accepted remains to be seen.

Only recently, I received a grant to build hardware for experimental testing, and I am quite excited about it. What many of my colleagues overlook is that hardware was my original love. I spent four years of my undergraduate degree in the workshop (alongside classes, of course) designing and manufacturing Formula Student race cars. But because my PhD was not with an experimentalist, so I am not a true experimentalist. In Indian academia, we let experience based labels to define the limits of what a trained engineer can learn (not)!

As far as the potential areas go, the most consequential opportunity is not a single technology domain but a full vertical, raw materials through processing, fabrication, and systems integration in an area where India has genuine assets and a closing strategic window. For instance, in semiconductors, the strategic logic could be the following: competing at advanced silicon nodes below 7 nm requires tens of billions of dollars and decades of accumulated process knowledge. That ship has sailed for new entrants. Photonics and neuromorphic computing are earlier in their development curve. A country that invests now in the full stack of such computing hardware has a realistic path to a leadership position. Research institutions in India have early activity in silicon photonics, but it currently depends entirely on foreign substrates and foreign fabs. Closing that dependency, starting with materials, is the work

8. Today, many students move away from core engineering research because they feel it offers fewer opportunities compared to corporate or software careers. In your opinion, how can India make deep-tech research more exciting, industry-connected, and rewarding for young engineers and researchers?

Students who leave for corporate careers are responding rationally to real incentives. A PhD in mechanical sciences (mechanics and associated fields), after five or six years, enters a faculty position at a salary that compares unfavourably with a mid-level software role. That is arithmetic, not culture.

But the AI and ML pull has made things worse in a specific way: it offers students the appearance of cutting-edge technical work while actually making them users of foreign platforms. The path of least resistance is to attach a neural network to any problem, publish on an international benchmark, and move into industry. It is unknown what they end up doing at these industrial jobs, which also, are not offered by Indian industry. This is individually rational and collectively damaging — it produces a large, skilled, substitutable workforce for companies whose critical technology decisions are made elsewhere.

There is another dimension that discourages capable researchers at every career stage, from staying in the system at all: merit is not reliably rewarded, and the gap between what is rewarded and what is meritorious tends to widen as careers progress. Academic positions, grant allocations, journal editorships, award nominations, and conference invitations in India are heavily shaped by who you know, which institutional/regional networks you belong to, and how faithfully you have maintained those affiliations, far more than by the quality of your work. A researcher who has chosen hard, independent problems, the kind that involve building things from scratch, working outside fashionable areas, producing results that lead to hardware development, consistently finds themselves at a structural disadvantage relative to peers who have invested more in network navigation than in the research itself. This is also a self-feeding cycle in many ways.

Making research careers genuinely attractive to the best people requires addressing this alongside the economic problem. Stronger fellowship stipends and structured industrial exposure matter. But none of it will retain researchers who have the ability to be elsewhere if the system continues to consistently reward connection over capability. Transparent, externally evaluated hiring and promotion, with meaningful representation from genuinely independent reviewers outside the immediate institutional network, and with accountability for research outcomes rather than output volumes, would change the incentive landscape considerably.

9. India wants to become a global technology and innovation leader in the next decade. What changes do you think are most important right now in India’s engineering and research ecosystem to make that possible?

The problem is not a shortage of ideas but a shortage of honest diagnosis. Here is my honest diagnosis. We have all the tools, the intellectual power, the energy, and the most of all, availability of a vast and young demographic. We just need to be honest in directing its might.

The first thing to be honest about is what kind of independence we are building. Software and simulation running on foreign hardware, using foreign frameworks, validated against foreign experimental data, is a fragile form of self-reliance. We are producing, at large scale, a sophisticated user base for a technology stack we do not control. The semiconductor manufacturing opportunity of the 1990s is a useful reminder that strategic windows close, and catching up afterwards is disproportionately expensive. The window for photonic integrated circuits and neuromorphic computing (continuing my examples) is open now. That investment needs to start with raw materials.

The second honest assessment: Spending at 0.64–0.66% of GDP has not grown as a fraction of the economy in a decade. Calling a server indigenous when its processors are still foreign and its interconnect is at proof-of-concept stage is the kind of accounting that sounds good and changes nothing. A credible commitment to 2% of GDP in R&D, with a meaningful share going to physical infrastructure, experimental facilities, and genuinely indigenous hardware, would be a structural shift.

The third, and the most resistant to policy intervention, as discussed earlier, the ecosystem needs to become merit-based in practice, not just in principle. Advancement through institutional networks and affiliations rather than demonstrated capability is a self-reinforcing problem. Researchers doing hard, independent, hardware-oriented work face structural hurdles at every stage and eventually leave, while the system continues to reward network proximity over scientific output. Transparently evaluated processes with accountability for research outcomes rather than output volumes are not a luxury. These are prerequisites for everything else. We have to make the grass greener on this side. We must not be shy in recognizing and rewarding useful research which does not subscribe to the hype. Definition of “usefulness” should be anything which helps indigenous hardware development. I am pushing myself to do useful research against all hurdles, but usefulness is also about being open about the problems being faced. For instance, even if I want to, I cannot contribute to useful research in a lot of allied fields due to extreme gate-keeping, which prevents the right exposure.

I believe that only by re-orienting the priorities of young researchers towards product and hardware, we can sustain a computational research program in the long-term. Computations are supposed to help actual development, and not stop at simulations and proof-of-concepts.