The Challenge
As adidas' cloud native platform scaled, onboarding hundreds of application teams across multiple Kubernetes clusters, cloud infrastructure costs scaled with it. Every new team and service added compute, and over time the gap between what was allocated and what was actually used kept growing. The cloud bill was climbing, but the value delivered per dollar spent was not.
Closing that gap at adidas' scale isn't a side project. It means rethinking how resources are allocated, how nodes are sized, and how idle capacity is handled across the entire cluster fleet, without disrupting the teams that depend on it.
Why Giant Swarm
Giant Swarm has managed adidas' cloud native infrastructure since 2017 — not just as a vendor, but as an extension of their platform team, building, scaling, and operating the platform side by side.
"I call our cloud-native platform the field of dreams. The Giant Swarm guys are really amazing and know their stuff in and out. Specifically in the area of containers and Kubernetes and everything around this I never met a more knowledgeable partner."
— Daniel Eichten, VP Enterprise Architecture, adidas
When cloud costs became a priority, adidas didn't need to pull engineers off product work or bring in a new partner starting from scratch. Giant Swarm already knew the infrastructure inside out, and with experience from managing 150+ production clusters across industries, had the Kubernetes expertise to optimize it. The result: optimizations designed around adidas' specific workload patterns and constraints, without application teams having to get involved.
The Solution
A joint team of Giant Swarm and adidas platform engineers tackled the problem through four automated, infrastructure-level measures, built entirely on open source tooling. All four were applied globally with an opt-out model so application teams kept full control.
The four levers:
- Intelligent Node Management: Karpenter selects the right instance types and consolidates workloads onto fewer, right-sized nodes
- Automatic Resource Optimization: Kyverno + VPA policies adjust CPU and memory requests based on actual usage, eliminating over-provisioning that quietly accumulates over time
- Scheduled Scaling: kube-downscaler reduces replicas outside office hours, freeing compute that was previously running idle
- Optimization Blocker Removal: Kyverno policies prevent misconfigured Pod Disruption Budgets from keeping half-empty nodes alive
Every measure runs continuously and adapts as workloads change. Teams that need different behavior opt out with a simple label.
Capabilities used: Kubernetes, Security. Delivery model: Fully managed.
The Results
The team implemented all four measures in 3 months. The impact was immediate:
- Up to 50% reduction in overall cloud costs across development and staging environments
- ~30% savings in CPU and memory alone through automated resource optimization
- 2x pod density per node — same workloads, half the infrastructure
Cloud costs went from growing with every new team to actively decreasing, while the platform continued to scale.
"Giant Swarm leads the way in platform engineering excellence, enabling us to focus on delivering what truly matters."
— Paul Vassu, VP Platform Engineering, adidas
What's Next
The non-production optimizations are now available for production clusters through an opt-in model, giving application teams the choice of which measures to enable and how to configure them.
The compute optimization also revealed the next opportunity: a significant share of cloud costs comes from cross-availability-zone network traffic, and reducing it starts with seeing where the traffic actually flows. Giant Swarm is already building that visibility into the observability platform, so adidas' teams can spot expensive patterns and make informed decisions about service placement. That's the value of a long-term platform partnership — the curated stack evolves with adidas' needs, and each capability builds on what came before.
