Subaru won the CNCF End User Case Study Contest at KubeCon + CloudNativeCon Japan 2026 after transforming the AI infrastructure behind its next-generation EyeSight development platform. The automaker combined Kubernetes with multiple CNCF projects to slash container image pull times from roughly three hours to three minutes, directly addressing a bottleneck that had slowed machine learning cycles for its advanced driver-assistance system.
Developing AI models for EyeSight required continuous data processing, training, validation, and inference. As workloads grew, Subaru's on-premises GPU environment became a choke point. AI container images regularly exceeded 30 GB, and downloading them occupied hours before any work could begin. Deployment relied on manually executed scripts rather than declarative configuration, and complex ML pipelines lacked a unified orchestration framework. Reproducibility and efficiency suffered.
Building a cloud native AI platform
Subaru's engineering team built a Kubernetes-based platform that brought together Envoy Gateway, Gateway API, and MetalLB to handle networking for massive container images. The result was a 60x improvement in pull times, from approximately three hours to just three minutes. Alongside that, the team adopted a GitOps workflow using Argo CD and Helmfile, standardization that built on existing Helm-based deployments. Argo Workflows then automated the end-to-end machine learning pipeline, removing manual steps and improving operational consistency.
"We wanted to build an AI development platform that removed operational friction and enabled our engineers to focus on improving model accuracy instead of managing infrastructure," said Ryoji Kobayashi, DevOps engineer in Subaru's ADAS development department. "By embracing cloud native technologies and GitOps practices, we've significantly reduced development bottlenecks while creating a more reproducible, scalable platform for machine learning."
What the CNCF prize recognizes
Chris Aniszczyk, CTO of the Cloud Native Computing Foundation, said the award highlights how open source tools can solve real-world infrastructure challenges. "Subaru demonstrates how combining Kubernetes with projects like Argo, Envoy, Helm, Harbor and MetalLB can solve real world infrastructure challenges while accelerating AI innovation," Aniszczyk said. "It's an excellent example of open source cloud native technologies delivering measurable business impact."
The award was announced during a keynote session at KubeCon + CloudNativeCon Japan, where Kobayashi presented the use case in detail. Subaru plans to continue expanding its Kubernetes and CNCF-based infrastructure as AI workloads grow more complex, aiming to further improve developer productivity and platform capabilities.
Why this matters for IT and development professionals
For teams running large-scale AI workloads on-premises, Subaru's results offer a concrete reference point. A 60x reduction in image pull times is not a marginal gain-it's the difference between a workflow that stalls daily and one that keeps engineers coding. The stack-Envoy Gateway, Gateway API, MetalLB, Argo CD, Helmfile, and Argo Workflows-is entirely open source and available to any organization willing to invest in cloud native practices. This case study, recognized by the CNCF, shows that applying these tools to AI infrastructure can deliver measurable speed and reproducibility without proprietary lock-in. IT professionals exploring AI for IT & Development will find a practical pattern for bridging the gap between machine learning requirements and operational reality.
Your membership also unlocks:
