Midokura Case Study | Hedgehog

The Challenge

Midokura, a Tokyo-based Sony Group company, builds sovereign AI infrastructure for enterprises and public institutions in Japan and beyond.

Its AI Factory as a Service gives customers the frictionless developer experience of a public cloud — running entirely on private, on-premises hardware they control.

That sovereignty promise is why customers choose Midokura: organizations with proprietary datasets increasingly can't — or won't — move them into hyperscale clouds. Delivering on it meant building a multi-tenant platform where enterprise clients run isolated training and inference workloads on shared bare-metal GPU clusters, with low-level infrastructure hidden behind a simple UI and API.

The application layer was manageable. The network was the bottleneck: manual fabric configuration, complex tenant isolation on bare metal, and synchronized AI traffic patterns that overwhelm conventional networks and stall GPUs when packet loss or congestion occurs.

The Solution

Midokura partnered with Hedgehog to automate its physical network fabric and remove the burden of manual switch configuration.

Using Kubernetes-native YAML and CRDs, the team now declares network intent once, and Hedgehog translates it into BGP underlay, VXLAN segmentation, and tenant isolation across the fabric.

This fit Midokura’s GitOps workflow and gave the platform a consistent, version-controlled way to replicate environments. Hedgehog also configured lossless RoCEv2 networking out of the box, keeping GPU traffic moving efficiently while avoiding vendor lock-in through SONiC and multi-vendor white-box hardware.

Critically for a sovereign platform, tenant isolation is enforced in the fabric itself — in switch ASICs and SmartNICs, not host software — so every tenant boundary is auditable at the wire level.

Data Sovereignty

Why this matters for sovereign AI

Japan's enterprises and public institutions are investing heavily in sovereign AI — infrastructure where proprietary data never leaves their control. Multi-tenancy is the hard part: sovereignty claims are only as strong as tenant isolation.

By enforcing isolation in the network hardware itself, Midokura offers its customers separation they can verify, not just trust.

At-a-Glance

COMPANY

Midokura, a Sony Group company

HEADQUARTERS

Tokyo, Japan

INDUSTRY

AI Infrastructure

SCALE

Multi-GPU bare-metal clusters (8+ GPUs/server), spine-leaf fabric

HARDWARE

White-box switches with Broadcom ASICs, multi-NIC / SmartNIC GPU servers

PRODUCTS USED

Fabric, VPC isolation, RoCEv2 QoS automation

WEBSITE

www.midokura.com

CUSTOMER VIDEO

Midokura discusses how they use Hedgehog to manage their network fabric.

Dam Dimitri
Founder & CTO, Midokura

HOW MIDOKURA BUILT IT

Read Midokura's own engineering deep-dive on automating their switch fabric with Hedgehog and SONiC.

Engineering an Automated Switch Fabric With Hedgehog and Sonic

Vendor Lock-in Eliminated

Since Hedgehog runs on SONiC and works across white-box hardware, Midokura can mix vendors instead of being tied to a proprietary switch stack.

GPU's Fed, Not Stalled

Hedgehog configures lossless RoCEv2 fabric behavior for GPUDirect out of the box, keeping expensive GPUs training instead of waiting on retransmits.

Reduced Time to GPU Value

Hedgehog enabled Midokura to reduce deployment time from months to hours by automating the fabric and removing manual switch-by-switch work.