RAYSolute's 2026 market report on Indian edtech, titled "The Great Correction & The Phygital Renaissance," lays out the shift plainly: the direct-to-consumer edtech market has corrected from a 22 billion dollar valuation bubble, the era that produced BYJU'S before its collapse, to a sustainable 3 billion dollar market. The operators standing after the correction are "phygital," combining physical centres with digital content, not pure-online plays. PhysicsWallah is the clearest example. It IPO'd this year at roughly a 5.2 billion dollar valuation, runs over 300 offline centres across more than 150 cities, and is opening new ones deliberately in smaller towns like Muzaffarpur, Dhanbad and Ujjain so students do not have to migrate to Kota or Delhi. It layers AI tools, including a Smart Doubt Engine, on top of that physical footprint rather than betting on AI alone. Source.
I want to be fair to what this actually says before I make it about me. This is not a story about AI losing to classrooms. It is a story about what a physical centre sells that a pure digital product has to earn some other way: the parent who can walk in and see the place, the sense that someone is watching if their kid falls behind, proof the business will still be there next year. PhysicsWallah bought that trust with buildings. That is a real cost most AI-only products never pay, and it shows up in why families choose it.
InfiNotes is the counter-case I have actually lived, not a rebuttal to this. It stayed fully digital, no centres, no franchise fees, and still crossed 10,000 registered users, a 12.9% free-to-paid conversion and profitability in five weeks on zero marketing. So digital-only clearly can work in this market. But the honest read is that it worked because the product had to substitute for the trust a centre gives away for free: being transparent about what it gets wrong, being cheap enough that a bad bet does not hurt, earning renewal instead of assuming it. The mistake founders make reading this report is concluding AI needs a building to be believed. It does not. It needs to stop assuming a good model is trust enough on its own.