Tokens are becoming venture capital’s new fieldwork.
Pratyush Choudhury spends between 300 million and 500 million tokens a day understanding AI technology. The co-founder of Activate AI also spends anywhere from a few hundred to a few thousand dollars daily on AI experiments, turning technical curiosity into part of the investment process.
“Unless you fundamentally understand what the technology can and cannot do, I don’t think you can build a truly great AI company,” Choudhury says. His routine includes OpenAI’s Codex and Anthropic’s Claude, which together account for roughly 90% to 95% of his AI usage.
Choudhury says he reads research papers, uses X to discover which papers are worth reading, and engages with the researchers and contributors behind them. That is a demanding way to screen companies, but AI startups now sell technical possibility as much as software — and possibility is difficult to price from a pitch deck.
Activate AI, co-founded by Choudhury and Aakrit Vaish, is India’s first venture capital fund dedicated exclusively to AI. The $75 million fund launched in December and participated in the funding round that made Sarvam a unicorn with a valuation surpassing $1 billion.
AI funding meets infrastructure reality
Choudhury brings experience from Amazon Web Services and Together Fund, while the wider investment market is testing whether AI spending can produce returns before the invoices arrive. Vinod Khosla, founder of Khosla Ventures, has discussed the funding race, and Marc Palet, a venture capitalist at OMVC, has added to the recent debate around the venture capital industry.
India’s biggest constraints on sovereign-AI ambitions are compute and data. That leaves investors evaluating not only models and founders, but also access to the machinery and information needed to train and operate them. A brilliant demo cannot negotiate with a GPU shortage.
Gavin Baker, a hedge fund manager at Atreides Management, says AI companies are starting to show improved cash flow and that demand remains strong. GPU prices remain high, he says, which points to strong demand for AI computing.
Investors sold off memory and chip stocks in July because of concerns about AI infrastructure spending. Meta’s free cash flow plunged 91% year over year to $784 million in the second quarter of 2026, yet Baker argues that the spending cycle is beginning to answer its return-on-investment questions.
“You’re going to see a lot of acceleration that’s going to answer these ROI questions. You’ve started to see that this quarter,” Baker says. He also says, “I don’t think anyone in ’24 or ’25 thought that the prices of old GPUs would still be going vertical in 2026.”
Cloud providers have not fully monetized their AI infrastructure because existing lower-priced contracts still cover part of that capacity. The market is therefore stuck between enormous demand and business agreements written before AI became the preferred excuse for every capital budget.
Research moves beyond frontier models
Academic AI research has shifted during the last six months from model capabilities toward questions less likely to be addressed by private companies. The change reflects cost as much as academic taste: researchers face high prices when querying models from OpenAI, Anthropic, and Google.
Many AI academics now build specialized models for data analysis, predictions, or simulations of physical systems. The AI2050 program offered funding last week to help university researchers buy GPUs, giving academic teams access to hardware that commercial labs can treat as a line item.
Training frontier models remains expensive, pushing researchers toward smaller, more efficient models and new architectures. Tim Dettmers, a computer scientist at Carnegie Mellon, believes AI models can make human scientists more efficient, but automation will not move at one speed across every field.
Empirical science is slower to automate than mathematics because collecting data is slow. Some AI researchers are concerned that models have already solved research problems in mathematics, raising questions about the future role of human mathematicians.
Google DeepMind’s AlphaFold team, the group behind AI for protein structure prediction, was disbanded last month. Anthropic has also found a technique to probe deeper into Claude’s workings, while its newest flagship product, Claude Science, focuses on AI for science.
Those advances arrive with hazards attached. A flaw makes large language models vulnerable to attacks such as sabotaging aircraft navigation systems, and a startup claims to have broken through a bottleneck holding back LLMs — a claim that still faces skepticism.
The investment story is no longer just about finding the next unicorn. It is about understanding which models deserve expensive infrastructure, which research problems need private capital, and whether the technology can deliver useful work before the token bill becomes the business model.
Based on
- How one VC burns through hundreds of millions of tokens a day to find the next unicorn — restofworld.org
- I’m a VC Who’s Chosen $120 Worth of AI Subscriptions Over an Intern – Business Insider — businessinsider.com
- Vinod Khosla on funding Discovery Loop, state of AI race and AI economics — cnbc.com
- Hedge fund manager Gavin Baker says investors have been too quick to judge Big Tech’s AI spending | Business Insider Africa — africa.businessinsider.com
- AI professors are negotiating the new realities of academic research | MIT Technology Review — technologyreview.com
