Halving cost-per-learner on a live-cohort LMS at 4,300 concurrent

Vidyapath Learning runs live, cohort-based upskilling programmes — data engineering, product management, cloud — for working professionals. 40,000 learners on the roll, up to 1,200 concurrent in a single evening class, and a unit economics problem that was quietly getting worse every month.

They were paying interactive prices for a passive audience

Every live class ran on a general-purpose video conferencing product, with all 1,200 attendees joined as full participants. But in a lecture, twelve people speak and 1,188 watch. They were paying real-time, bidirectional WebRTC prices — the most expensive way to move video that exists — to deliver a broadcast. The per-learner infrastructure cost was rising faster than the price they could charge.

Everything downstream of the class was manual

• Attendance was reconciled by hand from a meeting export, then keyed into a spreadsheet, then into the certification tracker. For a cohort model where attendance drives certification eligibility, this was both the most important data in the business and the least trustworthy.
• The question bank had leaked. Their assessment tool served the full paper to the client on exam start. A learner with the browser devtools open could read every question — and the answers, in one memorable case — before answering the first one. Screenshots of the bank were circulating in cohort WhatsApp groups.
• Proctoring cost more than it was worth. Their vendor recorded the full session and billed for human review of all of it. They were paying to have a person watch 40 minutes of video of somebody doing nothing suspicious, thousands of times a month.

And it all happened at once

Traffic is not diurnal — it is a spike. Working professionals attend class between 8pm and 10pm IST. The platform was idle for twenty hours a day and screaming for four, and it was provisioned for the screaming.

Migrated cohort by cohort over a term, with the old and new class experiences running side by side for three weeks so learners could switch back if the new one was worse. After the second week, nobody switched back.

Results

• Infrastructure cost per learner fell 58%. The overwhelming majority is the WebRTC-to-HLS split. This was not an optimisation; it was correcting a category error about what a lecture is.
• Peak concurrency reached 4,300 without incident — more than triple the load that used to break them — during a free masterclass they ran precisely because they finally believed the platform could take it.
• Proctoring cost per exam dropped 71%. Same integrity outcomes, a fraction of the human review hours, because the humans now only look where looking is warranted.
• Zero question-bank leaks since launch, across nine assessment cycles. The bank cannot leak from a client that is never sent it.

What it did to the business

Attendance data went from a spreadsheet somebody maintained to a stream the product reads, and the certification pipeline stopped being a monthly manual exercise. More importantly, the unit economics inverted: adding learners to a cohort now costs them very little, so their marginal cohort is profitable in a way it flatly was not before.

They have since doubled cohort sizes on their two largest programmes without adding infrastructure spend, which is the outcome that turns an engineering project into a business one.