Video is not “big files”. It is millions of small, identical, time-critical requests concentrated into the same seconds — a workload that stresses a completely different part of a CDN than a webpage does. Evaluate on these criteria, in this order.
Segment behavior beats headline throughput
A live stream at a 4-second segment length means every viewer requests the same object within the same 4-second window. The property that survives this is request collapsing (also called request coalescing): the edge sends one origin fetch per object and holds the crowd on it. Ask each candidate directly how many origin requests result from 10,000 concurrent viewers hitting a new segment on a cold POP — the right answer is a number close to the number of POPs, not close to 10,000. Providers without robust collapsing pass thundering herds to your origin or packager, which is how streams die at kickoff.
Throughput where your audience is, at p95
Video quality is decided by sustained per-viewer throughput at the 95th percentile, per region — not by peak Gbps in a datasheet. A viewer needs roughly 1.5× the bitrate of the rendition you want them to hold: about 8 Mbps sustained for 1080p at 5 Mbps. Benchmark by downloading real segments from real viewer networks at your traffic peak and reading the p95, per city, exactly as in our first-benchmark guide — but score throughput rather than TTFB.
The cache hierarchy question
For catalogs (VoD) the differentiator is mid-tier caching: an intermediate cache layer between edge POPs and your origin. Without it, every edge that misses goes to origin independently, and long-tail catalogs produce brutal origin load. With it, one regional fetch feeds many edges. Ask whether the tier is included or a paid add-on, and whether it is on by default for your traffic profile — the answer moves both performance and your origin bill.
Media pricing traps
Per-GB rates dominate media bills, so the definition of a gigabyte matters more here than anywhere: confirm whether mid-tier fill and origin shield traffic bill separately. Per-request fees look negligible until you multiply by segment count — a 2-second segment length doubles requests versus 4-second for identical bytes. And regional pricing: audiences in Oceania, LATAM, or India often bill at 2–4× the North American rate, so weight quotes by your actual regional mix, not the blended headline.
Features that are table stakes
Tokenized URLs / signed cookies for content protection, TLS at no per-request premium, native support for range requests and low-latency HLS/DASH, real-time log delivery (5-minute-old logs are useless mid-event), and instant purge measured in seconds. Any candidate missing one of these is not cheaper — it is incomplete, and you will buy the gap elsewhere.