A Bangladeshi EdTech platform cut video delivery costs 55% in 30 days

A video-heavy business in Bangladesh runs into the same wall sooner or later. The library grows, the audience grows, and the delivery bill grows faster than both. One of the country's larger online learning platforms hit that wall last year. Its monthly video bill had roughly doubled over two quarters while enrollment had not kept the same pace. Finance wanted an answer before the next renewal. The company asked us to find where the money was going and whether Vidinfra could bring it down. Thirty days after the switch, video delivery cost 55% less.

The platform asked us to keep its name out of this, so the story stays anonymous and the numbers stay real.

The tiered-pricing trap

The platform had grown up on a top video host we'll call XYZ. XYZ is priced in tiers. Each plan sets a ceiling on storage, on bandwidth, and on how many videos you can keep. Cross any one of those ceilings and you do not pay for the slice you went over. You get moved to the next plan up, and that plan costs roughly double the last one.

That math works against a company that adds videos every week. The platform crossed a storage ceiling and the bill doubled. A few months later an exam season pushed bandwidth past the next line and it doubled again. Neither jump matched how much more the platform was actually using. It could go 5% over a limit and pay 100% more. Across about a year, XYZ had moved it up the ladder several times, and each rung was twice the height of the one below.

So the platform left XYZ to get off the ladder and built a setup it ran itself. That is where the storage vendor, the transcoding company, and the global CDN came in. The tier jumps stopped, but a new bill took their place.

Where the money was going

The platform hosts about 14,000 lecture videos and streams to 35,000 to 45,000 students on a normal exam-season evening. Almost all of them watch from inside Bangladesh, on mobile data, on mid-range Android phones. The self-run stack was cheaper than XYZ's top tier, but it leaked in its own way. Every gigabyte a student watched left the origin, crossed a paid egress meter on the CDN, and came back priced for worldwide reach the platform did not need.

Three line items drove most of the cost. Egress fees were the largest. Per-minute transcoding charges came next. The third was plain waste: the CDN pushed one high-bitrate rendition to phones that could not use it, so students on 3G spent data and the platform spent bandwidth on pixels the screen threw away.

What changed

We put hosting, transcoding, and delivery on Vidinfra as one pipeline. A video uploads once and transcodes into an adaptive ladder, so a phone on a weak connection gets a 360p stream and a laptop on wifi gets 1080p from the same source. Delivery runs from edges close to the audience instead of a general-purpose global network, and there are no egress fees between the storage and the CDN. That removed the meter that had been running on every view.

For students, nothing on the surface changed. The player, the URLs, and the course pages stayed the same. The work sat behind them.

The 30-day rollout

Week one was migration. We copied the back catalogue across and re-transcoded it into the adaptive ladder while the old system kept serving live traffic. In week two, new uploads started going straight to Vidinfra, and both delivery paths ran side by side so we could compare playback on real sessions. By week three the platform pointed most of its traffic at Vidinfra and watched the dashboards through a full exam weekend, the heaviest load it sees all year. Week four was cleanup, and the old CDN and the separate transcoding contract were switched off.

The numbers after 30 days

Video delivery cost fell 55% against the previous month's bill, for the same library and a slightly larger audience. Most of the drop came from the egress fees going away and from adaptive streaming cutting the bytes sent to mobile viewers by close to a third.

Playback got better at the same time. Median start time went from about 3.4 seconds to 1.2. The rebuffering rate, the share of sessions that stalled at least once, fell from 6.1% to 1.8%. Tickets mentioning buffering or video that would not load thinned out over the month, which for a platform judged on whether a lecture plays during revision week counts for as much as the invoice.

The lower cost held after the rollout, so this is the running rate now, not a one-month dip. The platform puts the difference into making more videos instead of moving them.