Making Large-Scale Live Streaming Smarter: Netflix Rolls Out VBR Across All Live Events

Since January 26, 2026, Netflix has been encoding all Live events with VBR (Variable Bitrate) instead of CBR. In the Live pipeline, it uses AWS Elemental MediaLive's QVBR setting, aiming for more efficient encoding and lower network load.

With the existing CBR, once you set the resolution and nominal bitrate per stream, the actual transmission volume stayed roughly constant, making capacity planning easy. But since compression difficulty varies greatly depending on scene complexity in actual video, this led to problems: unnecessarily high bit usage for simple scenes, or insufficient bits for complex scenes.

VBR doesn't use a fixed bitrate but instead adjusts the bitrate to hit a target quality. For simple scenes, it significantly lowers the bitrate to deliver the same quality with less data, while for more complex scenes, it uses more bits to prevent quality degradation.

The effects were clear.

• On average, transmitted bytes decreased by about 15%
• Peak per-minute traffic decreased by about 10%
• Rebuffers per hour decreased by about 5%

However, since VBR can cause the bitrate to fluctuate significantly depending on the scene, stability became a key challenge. If a low bitrate persisted for a long time, the server would appear idle, so if additional sessions were added based only on current traffic as before, the bitrate could spike during subsequent scene transitions, potentially overloading the link or NIC.

To prevent this, Netflix changed how servers determine whether they can accept additional sessions: instead of relying on current traffic, they now reserve capacity based on each stream's nominal bitrate. This way, the server isn't fooled by VBR's temporary low-bitrate periods, and it can maintain a manageable level even when the bitrate rises again.

On the quality side, problems also emerged when simply reusing the existing CBR ladder. VMAF offline analysis and initial A/B testing showed an overall drop of about 1 point lower VMAF, with the gap being especially large in lower bitrate ranges.

The solution was to compare CBR and VBR per rung, and only slightly raise the nominal bitrate for ranges where VBR fell more than 1 VMAF point below CBR, closing the gap. Most high-bitrate ranges were left unchanged, resulting in a VBR ladder that matched the CBR ladder's quality while still achieving lower overall traffic.

Going forward, two directions are being pushed further.

• Improving the device ABR algorithm so it better reflects the actual segment size for the next few seconds
• Applying a discount that reflects actual VBR behavior to the currently conservative nominal-bitrate-based capacity reservation, to secure additional headroom

The key point is not a simple configuration change, but a redesign of the operational model that preserves VBR's efficiency while maintaining the predictability and stability required for large-scale Live operations.