Serverless was a big mistake... says Amazon

Amazon Prime Video’s recent cost-cutting move is landing as a direct challenge to the serverless microservices playbook: switching from distributed serverless components to a traditional monolith architecture reportedly cut Amazon Web Services spending by 90%. The core message isn’t that serverless is “fake,” but that the economics of distributed systems can erase the promised efficiency gains—especially when orchestration overhead and data movement dominate the workload.

Prime Video’s pipeline needed to analyze audio and video for issues like freeze and corruption. Instead of one unified service, it relied on multiple serverless functions—described as Step Functions coordinating Lambda-like tasks—to handle each stage: an entry point triggers file conversion, conversion turns audio/video streams into frames for detection, machine-learning detectors analyze the frames, and a final function aggregates results and stores them in an S3 bucket. That design created a repeating cost pattern: every handoff between functions required serializing/deserializing data and network communication. Because the system had to run repeatedly for every second of a video stream, the overhead compounded quickly.

The architecture also ran into practical limits. Prime Video hit bottlenecks tied to AWS account limits while trying to orchestrate the workflow at the required frequency. On top of that, the team temporarily uploaded intermediate files to S3, and the transcript frames bucket access as another major cost driver. In short, the distributed design increased both coordination overhead and storage/transfer expenses—turning what should have been “scalable” into an expensive bottleneck.

The fix was a bold re-architecture: consolidate the components into a single container, effectively turning the workflow into a monolith. With everything running in one place, the system can only scale vertically—bigger servers rather than independently scaling each component horizontally. That sounds like a disadvantage, but the transcript argues the trade-off paid off: removing inter-service communication reduced network usage and eliminated much of the serialization overhead. The result was the reported 90% reduction in AWS costs, which for a large-scale product translates into millions of dollars saved.

Still, the takeaway comes with a warning. Netflix famously moved from a monolith to hundreds of microservices after a major monolith failure in 2008, prioritizing independent scaling and fault tolerance. The transcript uses that contrast to land on a broader principle: cloud architecture doesn’t have universal winners—only trade-offs. For small teams, serverless can mean faster deployments and lower operational risk (especially when relying on free tiers and avoiding infrastructure management). For large, high-throughput pipelines, distributed overhead can outweigh the benefits, making a monolith—or at least a more consolidated approach—financially smarter. The central insight: serverless and microservices can be efficient in theory, but real workloads often reveal hidden costs in orchestration and data movement.