In 2015 a single thread on a laptop beat 128 cores at graph processing. A decade on, companies still pay seven figures a year to parallelise their own overhead. The one-week measurement that would catch it is missing from almost every design review.
Your Distributed System Is Slower Than a Laptop
Lloyd Moore’s deep dive into the economic and performance reality of distributed systems in 2026. The article argues that the industry’s default-to-distributed reflex—formed when hardware constraints required it—persists even though modern servers have 192 cores, 24TB memory, and SSDs reading 14GB/s. He shows how most companies pay millions in coordination overhead to parallelize their own infrastructure, not their workload.
Moore walks through a concrete example: a SaaS company’s .4M distributed event pipeline (Kafka, Flink, platform engineers) processing 2 billion events a day—which is a single-machine workload with a hundredfold capacity margin. A competent two-server setup would cost 7K/year and be faster.
The core argument: distribution is a cost to be justified by measurement, not a default justified by convention. Before approving any distributed architecture, someone should build the single-machine baseline and prove the distributed system beats it by enough margin to justify the operating costs. Most won’t.
A must-read for anyone designing systems at scale or defending architectural choices in design reviews.