📰 Key Takeaways

Pocket FM (an Indian audio storytelling platform) has doubled its annualized revenue to $500M over the past year, with AI now powering 93% of its total content library and 99% of new content production — driving a major boost in economic efficiency. CEO Rohan Nayak says AI has cut content production costs by roughly 80x: 100 hours of content that used to take a year to produce can now be made in a single day. The company still runs on a “humans handle creative ideation, AI handles production at scale” model, led by Vasu Sharma — a former Meta and Tesla scientist who now heads AI — who’s using years of production data and listener engagement signals to train the company’s own creative writing and text-to-speech models in-house. The platform’s more than 550,000 creators now produce about 2.5 million hours of AI-assisted content a year, up from just around 100,000 hours in its entire library two years ago — the total catalog of audio series has now surpassed 770,000 titles. That surge in content volume has also pushed 12-month revenue retention from 44% two years ago up to 76%. Annualized revenue climbed from about $250M a year ago to $430M this past April, and now to $500M (calculated as monthly revenue x 12, not contracted recurring revenue). Growth has been driven by expansion into the UK, Germany, and France, plus the launch of user-generated content in the US. The platform now has 96 titles that have crossed $1M in revenue, 13 of which have topped $10M. Pocket FM got its start primarily in the Indian market and now has more than 250 million listeners across over 20 countries — the US market grew about 70% over the past year and has become its largest market, accounting for roughly 70% of annualized revenue.

💬 JudyAI Lab Take

Pocket FM pushing its content library to 93% AI-generated, scaling annual output to 2.5 million hours, and doubling annualized revenue to $500M in a single year — that’s a growth rate worth paying attention to for anyone watching AI.

What’s instructive here for AI builders is that the team didn’t hand creativity entirely over to AI — they kept a “humans do creative ideation, AI does production at scale” division of labor, and trained their own purpose-built models using years of accumulated production data and listener engagement signals, rather than just bolting on a generic off-the-shelf tool. The ~80x drop in production cost, and content that used to take a year now taking a day, shows that the real payoff comes from turning proprietary data into a training asset — not just swapping in a more powerful model.

Worth thinking about: what accumulated user interaction data do you already have sitting around that could train a model purpose-built for your own use case?

📅 Source Info

🔗 Further Reading