A live-streaming platform's largest infrastructure line item is data transfer. What architectural levers reduce it, and which are unavailable?

Why egress dominates

Video delivery is bandwidth. Every viewer-minute is data leaving the network, and unlike compute or storage, per-gigabyte transfer pricing does not improve much with scale — it is the line item that grows fastest with success and resists the usual economies.

The levers that work

1. Encoding efficiency. A more efficient codec delivering equivalent perceptual quality at a lower bitrate reduces every byte delivered. This is the single largest lever, and it costs encoding compute and client compatibility work — an excellent trade when transfer is the dominant cost.

2. Adaptive bitrate tuned deliberately. The default ladder is often more generous than viewers can perceive on their devices. Matching rendition to device class and network reduces delivery without visible quality loss, and the savings are proportional to the whole viewership.

3. Peering and interconnection. Direct connections to large access networks bypass transit and change the per-gigabyte economics substantially. This is infrastructure work with a long lead time and it is where large delivery platforms end up.

4. Caching efficiency at the edge. Every cache hit is a byte not fetched from origin. For live content the cacheable window is short, which makes request collapsing and origin shielding more important than long TTLs — the goal is that a segment is fetched from origin once regardless of how many edges want it.

5. Peer-assisted delivery, where clients share segments. Meaningful savings at large concurrent viewership, with real complexity and privacy considerations.

6. Not delivering what nobody watches. Ending streams to idle clients, avoiding prefetch that is discarded, and not delivering high renditions to backgrounded players.

The levers that are unavailable

Compression beyond the codec. Video is already compressed; generic compression adds nothing.

Caching the uncacheable. Live content by definition has a short reuse window, so the cache-hit strategies that work for on-demand content apply weakly.

Reducing quality below perception thresholds, which trades the product for the bill and is usually the wrong trade — though it is a legitimate degradation control during a capacity event.

The architectural consequence

For a bandwidth-dominated workload, transfer cost is an architectural constraint, not an operational one. It shapes the encoding pipeline, the rendition ladder, the CDN strategy, the peering roadmap, and ultimately whether owning delivery infrastructure is justified — since the cost gradient, rather than the intercept, is what ownership changes.

Teams that treat egress as a bill to be negotiated rather than a design input find that no negotiation matches what an encoding or caching change delivers.