The world of ad-supported digital media is evolving at a breathtaking pace. With streaming audiences growing by the millions and advertiser expectations rising just as quickly, ad infrastructure can’t just be fast—it has to be smart.

Ad servers can scale over 1 -2 million transactions per second and serve 30 -50 million concurrent users around the globe.

What makes that scale possible? It isn’t just powerful infrastructure—it’s intelligence. By applying AI and ML on AWS, we’ve built systems that adapt, optimize, and respond in real time.

Today, AI and ML aren’t just accelerators. They’re the foundation of what’s next in ad tech.

The Complexity Behind Every Ad Impression

Delivering an ad during a streaming session—whether it’s on Prime Video, Twitch, Freevee, or any other publisher we support—might seem straightforward. But behind the scenes, it involves a dependency chain of a series of ultra-low-latency and high-throughput systems. Each impression relies on real-time coordination between the player, the stitcher, the ad server, the bidders, the creative services, and the policy engine for real-time bid flow.

All of this happens in under one second. And within that, the ad server typically has a latency budget of 20 milliseconds to respond. In that brief moment, the ad server evaluates and matches deals and targeting, ranks bids, applies rules & policy, manages frequency, creates pods for ad breaks, and sends it back to the player for stitching in the stream.

It’s a high-speed, high-stakes orchestration—one that increasingly runs on intelligent systems designed to learn and adapt in real time.

Predictive Infrastructure: Preparing for What’s Next

In ad-supported streaming, demand is unpredictable—it surges. For example, a Thursday Night Football kickoff or the start of a major esports final can send millions of viewers online within seconds, pushing our infrastructure to its limits. In those moments, there’s no time to scramble—manual scaling is not an option.

Such scenarios warrant the capability to accurately forecast supply inventory using machine learning techniques and tools like AWS forecast. The models typically use supply behaviour patterns, historical trends, ad server constraints, targeting features, inventory attributes being key inputs.

Optimizing Ad Delivery with Machine Learning

One of the first things I realized in this space is that not all ad requests deserve equal treatment. Some are tied to high-impact campaigns—guaranteed deals, premium placements, or launch moments where performance really matters. Others are important but can afford flexibility. Finding a way to reflect that nuance in how our systems respond became a clear priority.

We introduced reinforcement learning to help our ad server make smarter, value-aware decisions. Now, instead of assigning standard execution paths with fixed latency budgets, our models dynamically shift resources, prioritizing high-yield traffic while routing lower-stakes requests through more cost-efficient paths which are driven by ‘advertiser or publisher’ desired outcomes as well as influenced by yield based algorithms..

That shift didn’t just improve our efficiency—it aligned technical execution with business intent. And we’re not alone in that thinking: a 2024 Statista and MVFP study found that 85% of media leaders expect AI to drive revenue, with most viewing it as a strategic advantage rather than a disruption.

AI-Powered Resilience: Incident Detection and Resolution

In ad tech, the incidents that teach you the most aren’t always the ones that make headlines—they’re the quiet ones. A slight rise in latency, a subtle dip in fill rate—easy to miss, but costly if ignored. And at a global scale, even the smallest signals matter.

To stay ahead of that curve, we rely on anomaly detection tools like AWS Lookout for Metrics to keep a constant pulse on critical KPIs: bid response times, fill rates, delivery metrics, and more. These models act as our early warning system, surfacing subtle deviations before users ever notice.

What’s made the biggest difference is pairing those alerts with a diagnostic engine trained on real incidents. Now, what used to take hours of combing through logs and dashboards is narrowed down to minutes, dramatically cutting our mean-time-to-resolution (MTTR) and giving our teams the confidence to move faster when it matters most.

Supporting Engineers with Intelligent Tooling

AI and ML don’t just help the systems—they empower the people behind them. At our scale, even a few minutes of ambiguity during an incident can ripple across millions of users and billions of ad transactions. That’s why we’ve invested in intelligent tools that turn complexity into clarity—arming our engineers with real-time insights, guided automation, and faster decision-making.

Here's a quick snapshot:

These tools reduce cognitive load and give engineers the clarity they need to act decisively, turning incident response from reactive firefighting into informed, data-backed problem-solving.

Learning from Live Events: Scaling with Precision

Monetizing live events has been one of the most demanding and rewarding challenges we’ve taken on. In moments like sports broadcasts or entertainment premieres, timing isn’t just important—it’s everything.

In the past, ads for live streams were baked into the feed, which left little room for real-time targeting or yield optimization. That didn’t sit well with us. So, we built systems using server-side dynamic ad insertion (DAI) to enable ads to be sold and served live, as the action unfolded.

To make it work, we engineered a cellular ad server architecture that could scale horizontally, handling up to 1.2 million transactions per second, without missing a beat. AI-driven traffic shaping helped balance the load across these cells, even under extreme pressure.

What we gained was game-changing: flexible monetization, sharper geo-targeting, support for split creatives, and smarter segmentation—all in real time.

Ad Selection That Learns and Improves

Legacy ad selection used to feel rigid, driven by static rules and hard-coded waterfalls that couldn’t keep up with real-time nuance. We knew that had to change. Today, we’re rolling out contextual bandit models that make smarter, moment-by-moment decisions.

These models weigh live signals like viewer segments, past engagement, pacing goals, and even ad fatigue to choose the most effective creative in real time. They adapt continuously—learning what works, trying new variations, and optimizing for performance without sacrificing experience.

It’s a shift from rule-based to learning-based strategy—one that mirrors a broader move toward true personalization in marketing. As Forbes notes on impactful first impressions, relevance from the first interaction is essential, and machine learning makes it scalable.

Pioneering the Next Chapter in Ad Intelligence

“AI is one of the most profound things we're working on as humanity. It's more profound than fire or electricity.”

— Sundar Pichai, CEO of Google and Alphabet

That may sound bold, but in advertising, we’re already seeing it unfold. After decades of optimizing for speed and scale, the ad industry now faces a more complex challenge: building systems that can make smarter decisions, in real time, at massive scale.

By converging AI, ML, and cloud-native design, we’re building platforms that don’t just support billions of transactions—they learn from them. They predict demand before it spikes. They detect anomalies before they surface. And they empower teams to act with clarity, not instinct.

As privacy frameworks tighten, identity signals evolve, and audience expectations rise, the true competitive edge won’t come from infrastructure alone—it will come from intelligence that adapts with every signal.

At Amazon Ads, we’re not just responding to what’s next. We’re shaping it.

About the Author

Faizan Ahmad is a seasoned technology leader and Sr. Manager of Technical Programs at Amazon Ads Publisher Tech in Seattle. He leads strategic programs across 10+ systems, including Prime Video and Twitch, supporting over $56B in annual ad revenue. With prior roles at Facebook and Capital One, he brings 17+ years of experience driving innovation in cloud-scale systems, AI/ML, and digital monetization.