After nearly four years at OpenAI, Shyamal Anadkat, former researcher with the AI firm, returned to India earlier this year to build a stealth venture focused on AI for scientific discovery. In an interview with Poulomi Chatterjee, he discusses India’s AI opportunity, the talent returning from global labs and lessons from OpenAI. Excerpts:
Q. What made you return to India?
I grew up in India and moved to the US after 18. But over the past few years, I felt increasingly connected to the ecosystem here. The startup landscape has gained momentum, and I had worked with companies such as Portkey and Composia before joining OpenAI. Sarvam’s growth in sovereign AI has also been exciting. I felt talent was beginning to rally together. The ecosystem is still fragmented, but there is a lot of enthusiasm around the next generation of startups.
Q. What can you share about your new venture?
We are in stealth, but the broader goal is to accelerate scientific discovery. AI can generate hypotheses, including new molecules and materials, but verifying them in the physical world remains difficult. Scientific verification is scarce. It involves deciding what to test, what simulations to run, how to conduct experiments and how to measure and learn from the results. We want to close that loop. We are working on infrastructure and plan to build a series of scientific models that can help R&D teams become more AI-native.
Q. How difficult is it to find the right talent?
These are interesting times because a lot of global talent is returning. In recent months, people at labs such as OpenAI and Anthropic who are considering coming back to India have reached out. But finding the right talent remains hard. Raw technical ability and people who learn quickly are easier to find. What is much harder is finding people who have trained, evaluated and operated frontier systems at scale. India has relatively high AI skill penetration, at about three times the global average. Risk appetite has also increased significantly, which I found lacking earlier. The developer community is huge and growing, so my experience has been largely positive.
Q. What will it take for India to build for the world?
Until now, Indian technology exports have largely meant increasing headcount. That needs to change. We need to compound knowledge. AI is a major enabler of that. We should be able to repeatedly build and ship difficult, complex systems that are useful globally. Compute is another critical requirement. India’s available capacity is now closer to 50,000 GPUs, but these need to be made available to researchers for sustained work.
The IndiaAI Mission has done good work, but more is needed. The government should ensure Indian firms and researchers are prioritised over US companies that often pre-book GPU clusters for years. We also need to move beyond pilots and build globally trusted products. A frontier model needs compute, capital and talent, with the first two closely linked. Talent is the hardest. I am exploring ways to bring researchers back as advisers or on part-time projects.
Q. What lessons did you take from OpenAI?
People underestimate the distance between a great AI model and a reliable product. At OpenAI, my job was to shorten that distance — defining what it meant to perform real-world tasks well, measuring where systems failed and using those learnings to improve them. That is also a core principle in science. OpenAI showed me the value of small, high-agency teams that can do outsized work when compute, research and product are aligned. Lean teams with access to significant compute can be extremely effective.