Kazuki Sato
NTT DOCOMO BUSINESS, Inc.
Actions
Kazuki Sato is a software engineer at NTT DOCOMO BUSINESS, Inc., specializing in cloud-native distributed systems and GPU infrastructure. He serves as Lead Architect for GoAT (GPU over APN Testbed), developing nationwide distributed AI platforms leveraging IOWN APN. He also holds a Maritime Officer license — making him one of the rare engineers who handles both physical and digital containers, at least in theory.
Beyond the DC Walls: Building a Nationwide GPU Fabric with KubeVirt and Wide-Area L2
Cloud-native AI infrastructure faces a "physical wall" of power and land limits. To solve this, NTT and SKT built a "Nationwide GPU Fabric", merging a 1,000km+ high-speed L2 network (IOWN APN) with Kubernetes and KubeVirt to operate distributed DCs (Sapporo to Fukuoka) as a single cluster.
We deep-dive into the architecture, performance, and future of GPU orchestration beyond a single DC.
- Distributed vs. Traditional DCs: Achievable workloads and trade-offs across a 1,000km+ fabric from Sapporo to Fukuoka.
- Fabric-Aware Topology Alignment: We demonstrate how our solution maps long-haul IOWN APN constraints into KubeVirt VMs, maintaining high-performance GPU peer communication (NCCL) across 1,000km as a single, virtualized cluster.
- DRA Feature: We share KubeVirt’s static limits and the potential of DRA for dynamic orchestration.
Attendees walk away with a concrete blueprint to shatter the "walls" of a single DC — a scalable solution for global urban AI infrastructure challenges.
How CNCF Projects United Two Countries' Engineers to Solve AI's Power-and-Land Problem
As AI's appetite for compute keeps growing, a single urban data center hits a hard physical wall — not racks, but power and land. No amount of expansion at one site solves it. It's a challenge the whole world shares, and one that hits especially hard in Japan, where land is scarce.
So instead of expanding one site, NTT DOCOMO BUSINESS distributed GPUs across Japan. Over a wide-area L2 network built on IOWN APN (All-Photonics Network) technology, we operate eight data centers — more than 2,000 km from Sapporo to Fukuoka — as a single Kubernetes cluster. But a fast network alone isn't enough. The real challenge is making that distributed capacity usable as one: multiple tenants in secure isolation, existing workloads running as-is, with the feel of a single site. We made that real with cloud native technologies — above all, KubeVirt.
What matters most is who built it: engineers from Japan and Korea, brought together by open source. Because Kubernetes and KubeVirt gave us a shared language, NTT DOCOMO BUSINESS and SK Telecom could design as equals across a national border. Our takeaway: a challenge the whole world shares can be addressed on the common ground of CNCF, where companies and countries co-create a solution together. This one was built by engineers in Japan and Korea — and from here in Japan, we want to take this distributed architecture to a world facing the same wall.
KubeCon + CloudNativeCon Japan 2026 Sessionize Event
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top