Every data center has them: forgotten instances, orphaned storage volumes, and background services that keep running long after anyone needed them. Industry watchers call these "zombie workloads," and a growing cohort of FinOps engineers, cost optimization specialists, and inventory managers now spend their days hunting them down before they show up on the cloud bill.
According to Roger Strukhoff, chief research officer at IDCA, zombie workloads typically trace back to teams or individuals who stop using an application but never delete it. They also proliferate after corporate reorganizations or acquisitions, when no one is explicitly responsible for cleaning up unused instances. IDCA research suggests roughly 13% of US cloud usage comes from these leftover workloads, though FinOps tool vendors such as Broadcom and AWS put overall cloud waste even higher, commonly in the 25% to 30% range, once orphaned volumes and forgotten compute instances are factored in.
Old Problem, New Tools
The instinct to kill idle processes dates back to Unix, and serverless computing tried to formalize the practice with scale-to-zero configurations that shut down idle services automatically. But the approach has tradeoffs: mistakenly killing a live service can trigger cold starts and unwanted latency, a real risk for applications with spiky traffic.
To reduce false kills, vendors including Google, Flexera, Datadog, and IBM now offer tools that correlate billing data with performance metrics to flag resources that are being charged for but never used. These platforms continuously index assets, detect broken parent-child service relationships, and apply automated policies to tear down lingering processes.
Microservices Complicated the Hunt
Eric Newcomer, an analyst at Intellyx, notes that the old operating model was simple: pull the plug and wait for somebody to complain. Cloud-native microservices changed that. Applications built from hundreds of interdependent headless services can keep running in the background even after the parent application stops, quietly consuming resources while doing nothing useful.
AI Raises the Stakes
The generative and agentic AI boom has made the zombie problem more expensive. GPU-based hardware costs far more than the commodity CPUs that powered earlier cloud-native architectures, so inefficiencies that used to be rounding errors now show up as real money and real electricity consumption. Long-running training jobs, pipeline crashes mid-execution, and orchestration scripts that fail to clean up after themselves are creating new categories of AI-specific zombies.
Kubernetes, under the Linux Foundation and Cloud Native Computing Foundation, is evolving to better support AI workloads and is spreading to newer AI-focused cloud providers, but that work is still in early stages. For hosting providers and enterprise IT teams managing GPU capacity, the message is clear: the tools that worked for finding cloud zombies won’t fully solve the problem in the GPU era, and new observability and automation approaches are needed to keep AI infrastructure costs under control.