Artificial intelligence is forcing venture capitalists to rethink how they invest.

For decades, investors backed startups by judging founding teams, products, and markets. But as AI advances at breakneck speed, many now believe they need a much deeper understanding of the technology itself — or risk missing breakthroughs that can create entirely new categories overnight.

Pratyush Choudhury, co-founder of Activate AI, India’s first VC fund dedicated exclusively to AI, is among those embracing that shift.

To stay ahead, he has built a global network of researchers and developers, and regularly experiments with frontier AI models, spending as much time understanding the technology as evaluating the companies building it.

Activate AI, the $75 million fund Choudhury launched with co-founder Aakrit Vaish in December, participated in the funding round that catapulted Sarvam, one of India’s best-known AI startups, to unicorn status with a valuation surpassing $1 billion. Vaish founded one of India’s first conversational AI chatbot firms, Haptik, and was an adviser to the national IndiaAI mission. Choudhury comes armed with experience at Amazon Web Services — the infrastructural backbone of AI startups — and Together Fund, which backs AI-native firms.

Before launching the fund, Choudhury backed vibe-coding platform Emergent, which recently became a unicorn, and Rocket AI, which is reportedly in talks to raise around $50 million.

Choudhury spoke to Rest of World about how AI is changing venture investing and why understanding the technology is a competitive advantage.

The conversation has been edited for length and clarity.

How do you decide which AI startups to back?

Unless you fundamentally understand what the technology can and cannot do, I don’t think you can build a truly great AI company.”

It is almost impossible to find investment alpha in AI without understanding the technology. The days when a business-only founder could build a great AI company without deep technical fluency are probably behind us, at least for the next few years.

Unless you fundamentally understand what the technology can and cannot do, what is driving the gains in capability, and what might reduce its current limitations, I don’t think you can build a truly great AI company. For me, that means looking for founders who can see around the corner technically, and who are building companies around durable customer problems — not temporary capability gaps that the next model release could erase.

How do you keep up with such a rapidly changing field?

I read research papers. I rely heavily on X to discover which papers are worth reading, and I try to engage publicly and privately with the researchers and contributors behind them. 

I also speak regularly with researchers and applied-AI builders working at different frontiers across the U.S., Europe, and China. I try to understand the second-order effects they are seeing: How might a foundation model from China affect a legal-tech startup in the U.S.? How could a semiconductor breakthrough in India help a company in London?

I implement a lot of the technology myself. These days, I consume between 300 and 500 million tokens a day just to get my work done. That daily use gives me strong, firsthand intuition for what AI can and cannot do.

How much do you invest in firsthand experiments with AI?

I spend anywhere from a few hundred to a few thousand dollars a day. I am grateful to the friends and companies that have given me heavily subsidized access. Otherwise, I would probably burn through our entire fund just learning about AI.

A few thousand dollars a day is steep. Why is it so expensive?

I use frontier models for almost all of my work, from brainstorming event formats and guest lists to researching startups to implementing research papers on GitHub, and they are the most expensive. OpenAI’s Codex and Anthropic’s Claude together account for roughly 90% to 95% of my usage.

I also use Gemini and Grok extensively, have occasionally used Manus, and we use Granola heavily across the organization. I also use the paid version of Cursor.

How far along is India’s sovereign-AI ambition?

The two biggest constraints on India’s sovereign-AI ambitions are compute and data. With better access to both, Indian talent could produce far more high-quality research.

Should India build frontier AI? Should we invest in it? Absolutely.

The AI ecosystem, which will effectively become the broader technology ecosystem in a few years, is currently heavily dependent on an ally whose priorities can change from quarter to quarter. It is extremely risky not to have a credible alternative to fall back on.

We need to reduce our dependence on, and net imports of, foundational technologies like these.