Where the boundary between rented and owned belongs
Most streaming services rent everything because it all arrived in the same bundle. The line worth drawing is not between cloud and metal — it is between the tiers that spike and the tiers that bill by the gigabyte.
The hybrid pattern gets described as a compromise between two options. It is not. It is the result of asking one question per tier — does elasticity earn its price here? — and following the answer even when it splits the estate.
Two workloads wearing one bill
Control traffic and media traffic look like the same system from the outside and behave like opposites underneath.
| Control traffic | Media traffic | |
|---|---|---|
| Volume | Tiny | Enormous |
| Shape | Spiky — sign-in storms, release-day browsing | Steady, proportional to viewers |
| Latency | Matters | Matters, but caching absorbs it |
| Differentiation from paying more | Real | None |
| Right home | Rented | Owned |
Public cloud prices spiky, low-volume, latency-sensitive work well. That is the control tier exactly. It prices steady, enormous, undifferentiated work badly — which is the media tier exactly.
What sits where
- AWS — catalogue and search, authentication and entitlements, billing, the transactional database. Small payloads, real elasticity, a bill that stays in the noise.
- Your metal — GPU encoding, the S3-compatible origin, and the delivery edge. Capital already spent instead of a meter that runs.
- Both — one Terraform configuration and one pipeline.
That last line is the reason this is viable at all. Two estates operated separately is two operating models, and that is why most teams never try it. Two estates declared in the same repository and shipped by the same merge request is one workflow with more hardware in it.
The failure mode we designed for
If AWS is unreachable, sign-in and catalogue browsing degrade. Playback of already-authorised sessions continues from your edge, because the token the edge checks was issued before the outage and has a deliberate lifetime.
When it is the wrong answer
If your traffic is genuinely global and genuinely unpredictable, owned capacity in one region is a physics problem you cannot software your way out of. If you have no rack and no intention of having one, the cloud-native pattern with the cost controls designed in from the start is the honest recommendation, and we will say so.
Is this the bill you are paying?
If the pattern in this write-up describes your setup, the next step is reading your actual numbers against it. Send twelve months of billing and your concurrency curve.