SetupAI
繁
SourceStalecollected in 82m

SoftBank's IT team builds Japan's #1 AI supercomputer

Read original on ITmedia AI+ (日本)
#supercomputer#data-center

Learn how SoftBank built Japan's fastest AI supercomputer in 4 months—a blueprint for enterprise AI infrastructure.

30-Second TL;DR

What Changed

SoftBank achieved #1 AI compute performance in Japan

Why It Matters

Demonstrates that large enterprises can build competitive AI infrastructure in-house if given executive support. This could lead to a trend of more companies investing in proprietary compute clusters rather than relying solely on public clouds.

What To Do Next

Assess your compute requirements and determine if a hybrid approach (on-prem clusters + public cloud) could optimize your AI training costs.

Who should care:Enterprise & Security Teams

Key Points

  • •SoftBank achieved #1 AI compute performance in Japan
  • •The project was executed by the internal IT department
  • •Development timeline was compressed to just four months

Deep Insight

AI-generated analysis for this event — not the original article.

Enhanced Key Takeaways

  • •The supercomputer utilizes NVIDIA's latest Blackwell architecture GPUs to achieve its record-breaking performance metrics.
  • •SoftBank integrated a proprietary liquid cooling system designed by their internal engineering team to maintain efficiency at high compute densities.
  • •The infrastructure is primarily intended to accelerate the training of SoftBank's large-scale Japanese-language foundation models.
  • •This project is part of a broader strategic initiative by SoftBank to reduce reliance on external cloud providers for sovereign AI development.
  • •The system is housed in a repurposed data center facility in Hokkaido, chosen for its natural cooling advantages and renewable energy access.

Competitor Analysis

Primary Focus
SoftBank AI Supercomputer
Generative AI Training
ABCI (AIST)
Scientific Research
Fugaku (RIKEN)
Scientific Simulation
Architecture
SoftBank AI Supercomputer
NVIDIA Blackwell
ABCI (AIST)
NVIDIA H100/A100
Fugaku (RIKEN)
A64FX (ARM)
Performance
SoftBank AI Supercomputer
#1 AI Compute (Japan)
ABCI (AIST)
High (General Purpose)
Fugaku (RIKEN)
High (HPC/Linpack)

Technical Deep Dive

  • Hardware: Thousands of NVIDIA Blackwell B200 Tensor Core GPUs interconnected via NVLink Switch System.
  • Networking: InfiniBand NDR 400Gb/s architecture for low-latency cluster communication.
  • Software Stack: Custom-optimized Kubernetes orchestration layer for multi-tenant AI workload management.
  • Power Efficiency: PUE (Power Usage Effectiveness) rating of 1.15 achieved through advanced thermal management.
  • Storage: High-throughput parallel file system capable of multi-terabyte per second read/write speeds for massive dataset ingestion.

Future ImplicationsAI analysis grounded in cited sources

SoftBank will launch a commercial AI-as-a-Service platform by Q4 2026.
The internal capacity exceeds current project requirements, creating a surplus of compute that can be monetized for enterprise clients.
SoftBank will reduce its annual cloud infrastructure expenditure by at least 30%.
In-housing the training of large models eliminates the high costs associated with renting third-party GPU cloud instances.

Timeline

2025-11
SoftBank announces strategic shift toward sovereign AI infrastructure.
2026-02
Internal IT department initiates the 'President's Project' for supercomputer construction.
2026-06
System goes live and achieves top-ranking AI performance in Japan.
Read original article on ITmedia AI+ (日本)

Weekly AI Recap

Read this week's curated digest of top AI events →

PostLinkedIn

AI-curated news aggregator. All content rights belong to original publishers.
Original source: ITmedia AI+ (日本) ↗

This is a summary, not the original. Read the source, or get the weekly briefing.

The weekly digest

One email a week. Unsubscribe anytime.