Subaru Cuts AI Container Pull Times by 60x with Cloud Native Infrastructure
Key Takeaways
- ▸60x improvement in container pull times (3 hours → 3 minutes) through Kubernetes and Envoy Gateway optimization
- ▸Adopted GitOps with Argo CD and Helmfile to standardize deployments and improve reproducibility across AI development environments
- ▸Automated ML pipelines with Argo Workflows, significantly reducing development bottlenecks and accelerating ADAS innovation
Summary
Subaru has dramatically improved its AI development infrastructure by reducing pull times for 30+ GB AI container images from three hours to just three minutes—a 60x improvement. The automotive manufacturer achieved this by optimizing its Kubernetes networking architecture using Envoy Gateway, Gateway API, and MetalLB, winning the CNCF End User Case Study Contest at KubeCon + CloudNativeCon Japan 2026.
The optimization was driven by bottlenecks in Subaru's on-premises GPU environment, where large AI container images routinely delayed the start of workloads. To address operational friction, Subaru adopted cloud native technologies including GitOps with Argo CD and Helmfile for standardized deployments, and implemented Argo Workflows for end-to-end ML pipeline automation.
These improvements directly accelerate development of Subaru's next-generation EyeSight advanced driver assistance systems (ADAS), enabling engineers to focus on model accuracy rather than infrastructure management. By combining multiple CNCF projects, Subaru created a scalable platform that improves developer productivity, infrastructure efficiency, and ML reproducibility while demonstrating measurable business impact from open source cloud native technologies.
- Won CNCF End User Case Study Contest, demonstrating how cloud native technologies solve real-world AI infrastructure challenges at scale
