The 2026 World Cup shows unprecedented numbers. At 3 am London time, 9.1 million British fans stayed up to watch England beat Mexico, the largest UK audience ever recorded in that timeslot, while BBC iPlayer logged 48 million requests for its biggest day in history. From the other side of the pond, FOX drew 30 million viewers for USA–Belgium, at the time the most-watched soccer telecast in American history. Netflix's Tyson fight peaked at 65 million concurrent streams. Cricket on JioHotstar just hit 72.5 million. But the average global viewership for a FIFA match is 175 million. The 2022 final averaged 571 million global viewers. So, the question is: was this FIFA World Cup the largest event ever streamed? The answer depends entirely on what is measured, whether that's concurrents, average minute audience, requests, or reach. But it mostly reveals about where live video infrastructure is headed, and why the hardest engineering problems in our industry now happen at 3 am.

Audiences are engaging with video across more platforms, formats, and moments than any production team can manually serve — vertical content for social, highlights for engagement, subtitles for accessibility. Until recently, applying AI to live video at broadcast latency was impractical: the computational demands of modern ML models conflicted with the strict real-time constraints of live encoding pipelines. Learn how AWS Elemental has removed that constraint, what becomes possible when content intelligence operates in parallel with live encoding rather than as a post-production step, and how broadcasters are using structured video metadata to drive engagement, unlock new revenue streams, and reach audiences across every screen in real time.

Validating anti-piracy mechanisms under realistic live conditions is challenging due to the complexity of simulating malicious subscriber pools at scale. Traditional piracy detection relies on evaluating raw CDN logs, introducing data egress costs and delaying mitigation until after the event. 

This session demonstrates an edge-native approach to identifying token abuse in real time. We will show how to design a simulation framework to create pools of valid and pirate users. Then, utilizing edge-level token parsing and time-limited window tracking, we will flag bad actors with high confidence. You'll learn how to design the simulation framework and corresponding orchestration layer with real-time alerts, protecting stream monetization and avoiding overloading the operations center.

As sports streaming environments become increasingly saturated, static editorial curation is no longer enough to capture and sustain fan attention. To drive meaningful user retention and deeper daily engagement, the CBS Sports app successfully integrated a highly responsive, AI-powered personalized recommendation engine alongside its traditional editorial coverage.

This session breaks down the technical and architectural realities of designing and launching the new CBS Sports "For You Page" feature. We will explore how to blend algorithmic machine learning models with human-curated editorial content in real time, using embedding models to represent content as vectors. Finally, we will cover the infrastructure required to generate and process these AI embeddings against hyper-local user behavior signals, and the ultimate impact this personalized shift has had on customer retention and long-term app engagement.

Delivering a seamless playback experience requires a caching footprint that serves different workloads and access patterns using the best-matching operational model. While traditional demand-filled caches remain a popular pattern, Netflix increasingly relies on versioned datasets, snapshot-published with periodic refreshes, to power the streaming control plane. 

This session breaks down four distinct versioned dataset architectures deployed at scale. We will examine how they vary across dimensions of size, read latency, staleness tolerance, and cost, ranging from small in-memory datasets written in near-real-time to distributed petabyte-scale datasets published from offline jobs. Attendees will learn how matching the right caching architecture to specific media workloads (from asset manifests to user state) sidesteps classic failure modes of demand-filled caches such as stampedes and cold-start misses while meeting strict read-performance KPIs.

Building AI features for creators means chasing a moving target where quality and compute are in constant contention.
Training and deploying models is expensive and higher quality usually means bigger models; so how do you iterate quickly and build highly aesthetic, low compute models?

Taking it a step further, how do you build creation features that are aligned with what creators want to make and also what users want to see, and do so on demand? In this talk, we'll peek into how Meta is thinking about this, how we navigate these tradeoffs, how we train aesthetic models and adapt to virality, and where we're going as AI demand grows.

For years, the streaming industry has been trapped in a toxic relationship with legacy, segment-based protocols. We've twisted HLS and DASH into every low-latency shape imaginable, only to hit the inherent architectural limits. It's time to move on.

Media over QUIC (MOQ) promises to change the game, offering ultra-low latency without sacrificing the scalability of modern CDN architectures. But what happens when you actually stop looking at the IETF drafts and start writing code?

In this talk, we'll dive deep into the reality of building a real-time streaming pipeline using MOQ and MOQtail (our open-source implementation). We'll skip the introductory fluff and get straight into the engineering trenches.

Supporting a massive breadth of legacy, modern, and even non-traditional devices while delivering resource-heavy features has made the client application layer a critical engineering bottleneck. Traditional implementations routinely fail under the constraints of embedded hardware, rendering pipelines, and concurrent stream demands.

Moderated by Steph Lone (Global Leader, Solutions Architecture M&E and Sports at AWS), this panel explores the client-side architecture required to deploy premium playback experiences at scale. David Van der Voort (Senior Manager, Connected TV, International Engineering, Paramount+) breaks down the technical realities of building a converged, multi-tenant application with Paramount Lite+, while Jonathan Colwell (Director of Engineering, FOX) details the resource management and stream synchronization strategies required to power dynamic Multiview for the World Cup. Finally, Josh Lamb (Product Director, Red Bull Media House) discusses how to squeeze maximum performance out of diverse playback devices and pushing streaming standards to their limits to unlock cutting-edge features.

Real-time quality enhancement and metadata parsing do not happen in isolation. This session explores how to scale video quality metrics, enhancements, and metadata processing simultaneously. We will break down how we build a unified data flywheel and use that loop to continuously measure and improve operational performance at scale.

Video is token-hungry by nature, directly driving inference cost across compute, memory, and latency at scale. One minute of video at 30fps could represent 1.8 million words, yet the vast majority is spatially, temporally, or perceptually redundant. We argue that AI video token compression shares a profound connection with conventional video compression engineering. RD optimization, motion-compensated prediction, perceptual salience modeling, and ROI encoding all have direct analogs in efficient video tokenization. Further combined with lightweight domain-specialized pre-analyzers and vertical-specific priors, compression-domain thinking offers a practical path to making large-scale multimodal video inference computationally tractable.

It is a common misconception that WebRTC cannot scale beyond a few hundred (or thousand) concurrent viewers. What is inarguably true is that thanks to its peer-to-peer roots, scaling WebRTC takes special care and effort: because it was never designed with CDN delivery in mind, each WebRTC distribution system is ultimately a customized implementation with its own quirks and tradeoffs.

Learn how Amazon IVS built a WebRTC CDN capable of serving 1MM+ concurrent viewers on a single broadcast with unit economics similar to HLS/DASH. We’ll talk about lessons we learned along the way and core investments we’ve made in improving time-to-video (TTV), quality of service (QoS), and cost. Lastly, we’ll talk about the future, including intentional divergence from the WebRTC standard, first class multicodec and transcoding support, and forays into MoQ (Media over QUIC).

Building a Free Ad-Supported Streaming TV platform is deceptively simple at the start — spin up some channels, stitch together a playlist, serve a stream. The hard problems emerge when you're operating thousands of channels, a library of millions of VOD assets, live events that can't fail, and a platform deployed across multiple geographic regions. This talk traces the full arc of that journey: the early architectural decisions that either buy you flexibility or create ceilings you'll hit painfully later, the inflection points where systems that worked at hundreds of channels collapse at thousands, and the operational realities of keeping linear streams reliable at a scale most engineers never encounter. 

We'll cover the core engineering challenges — playlist assembly and scheduling at scale, ad insertion complexity, content ingestion pipelines that have to absorb millions of assets without falling over, live event integration into a fundamentally pre-recorded medium, and the regional distribution architecture that makes global delivery economically viable. The talk is grounded in real production experience and designed for engineers and architects who are either building a FAST platform or scaling one that's already outgrown its original design.

Traditional live audio cleanup relies on noise suppression, which attenuates non-speech frequencies within a mix. This approach sacrifices detail and never yields a truly clean signal. True dialogue isolation takes a different approach: instead of suppressing the mix, it extracts dialogue as an independent stem. Historically, this has only been viable in post-production, because the context a separation model needs for high-quality output is fundamentally at odds with the latency budget of a live feed. 

This session covers the engineering required to close that gap and run true dialogue isolation inline with a live broadcast at 11ms. We start with the central trade-off in source separation: better separation requires more signal context, and real-time constraints force hard decisions about how much context you can use. We then break down the latency budget across the broadcast signal chain, examining where time actually accumulates, and why broadcast-grade latency is a combined problem of fast model architecture and sufficient compute. Finally, we look at the downstream payoffs across viewer engagement, transcription and captioning, and compliance.

Large-scale video streaming platforms are notoriously difficult to operate — fragmented telemetry, siloed systems, and incident investigations that demand engineers manually correlate data across dozens of services before they can even begin to diagnose a problem. 

This talk makes the case that observability isn't just an operational nicety; it's the prerequisite for applying AI to streaming infrastructure. Drawing on a real-world audit of a major streaming platform's video pipeline, we'll walk through how implementing clean, standardized telemetry with OpenTelemetry creates a unified data foundation that unlocks the next generation of intelligent operations — AI-driven root cause analysis, autonomous remediation of common failures, and continuous optimization of pipeline performance, reliability, and cost. We'll cover the engineering trade-offs of instrumenting at scale, what "good" telemetry actually looks like in practice, and why telemetry quality is the critical gating factor that determines whether AI delivers meaningful operational value or just expensive noise.

Media over QUIC (MoQ) is no longer a theoretical next-gen protocol; it is active in production, proving its worth under some of the most unforgiving use cases. Today, MoQ is being used to remotely fly drones, drive boats, fold t-shirts in real-time robotics labs, and is soon set to pilot autonomous vehicles. But what do these telemetry and control use cases mean for the future of traditional media? We will explore how its unique sub-second transport mechanism solves massive pain points in remote live production (REMI), camera-to-cloud workflows, and interactive fan engagement at scale.

Engineering and product leaders have always been faced with the challenge of balancing content protection, increasing operational awareness, and accelerating the pace of user experience innovation. In this session, we'll explore ways of solving these challenges at scale and what we learned doing so.

We will detail how next-generation content protection is evolving, showcasing how real-time concurrency tracking acts as a critical line of defense against piracy and unauthorized credential sharing. We'll dive into the latest in operational dashboards and using under-the-hood data to enable natural language inquiries for bespoke business intelligence. Finally, we will explore how AI can be used to not only create personalized user experiences but also increase feature velocity.