Costing video delivery before you pick a provider

When the learning platform needed to serve lesson video, the obvious question was where to put the files. That turned out to be the less important half of the question.

For a teaching product, stored volume grows slowly and predictably: a term’s lessons, recorded once. Egress grows with every student who watches, every time they rewind, on every device. The bill tracks viewing, not library size — so a comparison based on storage price per gigabyte is comparing the number that barely moves.

Model the pattern, not the catalogue

Before looking at any provider’s pricing page, we wrote down the usage shape:

• —Library size, and how much it grows per term
• —Average lesson length and bitrate
• —Watch-throughs per enrolled student per lesson, including rewatching before an exam
• —Peak concurrency, which is not the average — it is the night before an assessment
• —How much is served to returning viewers, and is therefore cacheable

That last one matters more than anything on a pricing page. In a cohort-based product, a lesson released on Monday is watched by most of the cohort within days. That is a workload a CDN cache is unusually good at, and the cache hit ratio moves the bill more than the choice of origin does.

The three shapes of pricing

Providers price egress in roughly three ways, and the shape matters more than the rate:

1. 01Metered egress. You pay per gigabyte served. Simple, and it scales linearly with success — which is the problem.
2. 02Zero-egress storage. Storage costs more per gigabyte; egress to the internet is free or near it. For a read-heavy library this inverts the usual arithmetic.
3. 03Bundled delivery. A platform bills per minute delivered or per stream, folding in encoding and player. Convenient, and the unit rate stops looking convenient at volume.

The crossover between the first two depends entirely on your read-to-stored ratio. Below some ratio, metered egress is cheaper. Above it, zero-egress wins and keeps winning. Compute your own ratio; the number where it crosses is not a general fact.

Do not quote a rate you have not checked this month

Any specific price in a post like this is wrong within a year. What survives is the method:

1. 01Write down your viewing model first, in your own numbers.
2. 02Compute annual cost for each provider under that model, including request charges and any minimum storage duration.
3. 03Run it again at ten times the volume. A provider that is cheapest today and worst at scale is a migration you have scheduled without noticing.
4. 04Include the cost of leaving: egress to move the library out is itself a bill.

What we chose, and why the reason matters more

We put lesson video in object storage behind signed, expiring URLs, served through a CDN, with the origin on a zero-egress tier. The decisive factors were the cacheable cohort pattern and the read-to-stored ratio, not the headline storage rate.

The decision is written up as an architecture decision record with the model attached. When the numbers change — and they will — the next person does not have to reconstruct the reasoning. They can re-run the model and see whether the conclusion still holds.