Key takeaway

• Large-scale live streaming is now a resilience problem, not just a performance problem. At 15 to 20 million concurrent viewers, reliability depends on how the full stack handles segment delivery, origin routing, storage availability, and vendor failure.
• Sardius Media showed what multi-cloud resilience looks like in production. Its architecture writes segments across four S3-compatible storage providers and uses a latency-weighted selector to route each request to the fastest available origin.
• The reported result was a 0.03% global rebuffer ratio. That number came from live events at 15 to 20 million concurrent viewers, making the Sardius Media architecture the strongest concrete example from the panel.
• Qencode CEO Murad Mordukhay pushed the resilience discussion beyond failover. His point was that teams also need infrastructure that can adopt new encoding, AI optimization, and delivery capabilities without waiting on one vendor’s roadmap.
• The broader lesson is that modular, multi-cloud video infrastructure gives streaming teams more room to adapt. It helps teams stay reliable now while keeping the stack flexible enough to improve as codecs, AI workflows, and delivery economics keep changing.

Why this matters now

Live streaming has moved from a feature to a revenue line. A single broadcast can pull 20 million concurrent viewers, run on infrastructure assembled from four or five vendors, and lose money the moment a segment fails to load. The economics of streaming have changed, but the architectural assumptions behind most stacks have not. That gap is what the SMPTE and Backblaze panel on April 29 set out to address.

Qencode CEO Murad Mordukhay joined Jason Shore and Ari Burt of Sardius Media, along with Troy Liljedahl and Dave Simon of Backblaze, for a working session on what production streaming resilience actually looks like at scale. Watch on demand at smpte.org.

What the panel covered

The conversation cut past multi-cloud marketing and into the actual mechanics: how segments get routed, what the failure modes look like, and what teams trade away when they consolidate.

Per-request provider racing.
Sardius Media writes every output segment to four S3-compatible storage providers in parallel. At the CDN origin, each request resolves through a latency-weighted selector that picks the fastest responding origin per region, per request, in real time. The result: a 0.03% global rebuffer ratio measured across events at 15 to 20 million concurrent viewers. For reference, ITU-T P.1203 benchmarks for “excellent” quality of experience sit above 0.5%. Per-request racing is one of the design patterns that makes multi-cloud video infrastructure pay off rather than just add overhead.

Redundancy without four full copies.
Most teams assume resilience means writing every file to every storage provider, four times over. Sardius Media does not. A segment has to exist once, somewhere, and the manifest points the player at whichever origin returns it first. The four providers fill in over the next few seconds. Viewers see no difference. The storage bill drops by roughly 4x compared with full active-active replication, and the architecture stays economical at broadcast scale. In engineering terms this is called eventual consistency. In business terms, it is the difference between paying for four copies of every minute of video you stream and paying for one. This is the economic case that makes multi-cloud video infrastructure work at the scale Sardius Media operates.

Feature-set resiliency as a first-class concern.
Murad’s argument on the panel: infrastructure failover is solved, or at least tractable. Most teams know how to fail over a region or swap a CDN. The harder lock-in problem in 2026 is feature velocity. AV1 adoption, AI-driven encoding optimization, per-title rate control, and content-aware compression are moving in months, not years. A team that picked a video platform two years ago for its uptime numbers is now watching that same platform fall behind on the codec and AI capabilities that move the unit economics. Teams who built their stack so they can swap one component, the encoder, the storage layer, the player, the AI optimization pass, without re-architecting the rest, will ship those gains as they arrive. Teams on monolithic platforms wait for their vendor’s roadmap. Resilience, in that frame, is not just about staying up during a regional outage. It is about staying competitive across a three-year window where the underlying technology keeps shifting under everyone’s feet. That is the real argument for modular, multi-cloud video infrastructure, and it is the gap Qencode is built to fill.

The bigger picture

The thread running through the session: adaptability now beats raw performance as the design goal. The video stacks that will hold up over the next three years are the ones built from interchangeable parts, with clear contracts between layers, no assumption that any single vendor stays the best at any single job. AI is going to keep rewriting the cost and quality curves for video encoding, packaging, and delivery. The teams positioned to capture those gains are the ones whose multi-cloud video infrastructure lets them adopt the next improvement without a six-month migration. That is the case for modular video infrastructure, and it is the case Qencode keeps making in production with customers like Sardius Media.

Watch it

The full session on SMPTE runs about 50 minutes. It is most useful for engineering leads and platform owners rethinking how their video stack is assembled, especially anyone weighing the tradeoffs between a single-vendor video platform and a composable, multi-cloud video infrastructure. The Sardius Media numbers alone are worth the time.

Frequently asked questions

What is multi-cloud video infrastructure?
Multi-cloud video infrastructure is a streaming architecture that writes, stores, and serves video assets across more than one cloud or storage provider, with intelligent routing between them. The goal is to remove any single vendor as a point of failure or a bottleneck on cost, performance, or feature velocity.

How did Sardius Media achieve a 0.03% rebuffer ratio?
Sardius Media writes every output segment to four S3-compatible storage providers in parallel, then uses a latency-weighted selector at the CDN origin to pick the fastest responding origin per request, per region, in real time. The result is a 0.03% global rebuffer ratio across live events at 15 to 20 million concurrent viewers, well above the ITU-T P.1203 threshold for excellent quality of experience.

Why not just replicate every video segment across all four storage providers?
Full active-active replication costs roughly 4x more in storage. Sardius Media uses an eventual consistency model: a segment has to exist once, the manifest points the player at whichever origin returns it first, and the other providers catch up over the following seconds. Viewers see no difference, and the storage bill stays in line with single-provider economics.

What is “feature resiliency” and why is it different from failover?
Failover is about keeping the stream up when infrastructure breaks. Feature resiliency, the term Murad Mordukhay used on the panel, is about keeping the stack able to absorb new technology fast. AV1, AI-driven encoding optimization, per-title rate control, and content-aware compression are moving in months, not years. A modular video stack lets a team adopt the next improvement by swapping one component. A monolithic platform forces the team to wait on a single vendor’s roadmap.

Is multi-cloud video infrastructure cost-effective at scale?
Done as full active-active replication, no. Done with per-request provider racing and eventual consistency, the storage cost stays close to single-provider pricing while delivery resilience improves by roughly an order of magnitude. The Sardius Media architecture covered on the panel is one production example. Qencode’s partner ecosystem is designed to make this model the default rather than a custom build.

Where can I watch the full SMPTE panel? On demand at smpte.org. The session runs about 50 minutes.

Speakers

• Dave Simon, Sr. Director M&E Alliances, Backblaze
• Troy Liljedahl, Sr. Director of Solution Engineering, Backblaze
• Jason Shore, CEO and co-founder, Sardius Media
• Ari Burt, Co-founder, Sardius Media
• Murad Mordukhay, CEO and co-founder, Qencode

Moderated by Joel Welch on behalf of SMPTE.

---