10 Powerful Insights: Why SoftBank AI Computing Platform CHIE-4 Became Japan’s No.1 AI Supercomputer

Overview of SoftBank AI Computing Platform CHIE-4

SoftBank AI computing platform CHIE-4 is SoftBank Corp.’s flagship AI supercomputer, built on NVIDIA DGX SuperPOD technology with the latest DGX B200 systems powered by Blackwell GPUs. In November 2025, SoftBank announced that CHIE-4 ranked No.1 in Japan and No.5 worldwide in the AI-oriented HPL-MxP benchmark at the SC2025 supercomputing conference. It also placed No.3 in Japan on the TOP500 list and No.2 in Japan in the HPCG benchmark, confirming that CHIE-4 is not only strong in AI but also highly capable as a general-purpose supercomputer. ソフトバンク+1

From internal platform to national AI asset

SoftBank originally built its AI computing platform as an internal resource for developing large-scale AI models and improving telecom operations. Over time, the platform expanded in scale and ambition. It evolved from:

a corporate research infrastructure,
to a top-level AI supercomputer in Japan,
and now to a shared platform that SoftBank plans to offer to enterprises and research institutions as a core part of Japan’s AI social infrastructure.

Because of its scale and location inside Japan, SoftBank AI computing platform CHIE-4 now plays a role that goes far beyond one company. It supports domestic large language models (LLMs), sovereign AI initiatives, and high-end research that previously required access to foreign hyperscale clouds.

Naming, role, and relationship with NVIDIA DGX SuperPOD

SoftBank uses the internal name “CHIE-4” for the DGX SuperPOD cluster built with DGX B200 systems. In the corporate press release on the rankings, SoftBank explicitly identifies this AI computing platform as CHIE-4 and describes it as being composed of more than 4,000 NVIDIA Blackwell GPUs in DGX B200 systems. ソフトバンク

NVIDIA DGX SuperPOD is a turnkey AI supercomputer solution that integrates:

DGX servers (such as DGX B200)
a high-performance InfiniBand network
parallel storage
and a complete software stack optimized for AI workloads

By combining this reference architecture with SoftBank’s own network, data-center design, and operations expertise, CHIE-4 functions as an AI factory capable of training very large LLMs and serving massive inference workloads.

Benchmark Results: How CHIE-4 Reached Japan’s No.1 AI Performance

HPL-MxP domestic No.1 and global Top-5

At SC2025, CHIE-4 achieved Japan’s top score and a global fifth-place ranking in HPL-MxP, a benchmark designed to measure AI-oriented mixed-precision performance. ソフトバンク+1

Key points about this result:

HPL-MxP focuses on low-precision arithmetic, such as FP16 and FP8, which are widely used in deep learning.
A high HPL-MxP score indicates that the system can effectively sustain massive matrix operations similar to those in LLM training.
CHIE-4’s ranking shows that its hardware, network, and software stack are well-tuned for modern AI workloads, not just theoretical peak FLOPS.

In other words, SoftBank AI computing platform CHIE-4 is not only large on paper; it delivers real, usable AI performance on a global scale.

TOP500 and HPCG: Balanced performance for AI and HPC

In addition to HPL-MxP, CHIE-4 also posted strong results on two classic supercomputing benchmarks: TOP500 (Linpack) and HPCG. According to SoftBank’s press release, CHIE-4 ranked: ソフトバンク+1

No.3 in Japan and No.17 worldwide on the TOP500 list (double-precision Linpack)
No.2 in Japan and No.6 worldwide on the HPCG benchmark

These results show that CHIE-4 has:

high raw floating-point performance for traditional HPC tasks, and
strong performance on memory- and communication-intensive workloads closer to real scientific applications.

This balance is important. It means CHIE-4 can serve both AI and scientific computing communities: from climate modeling and fluid dynamics to training LLMs and multimodal models. For universities and public research centers, this dual capability makes SoftBank AI computing platform CHIE-4 an attractive shared resource.

Hardware Architecture: DGX B200, Blackwell GPUs, and Network Fabric

Over 4,000 Blackwell GPUs and 510 DGX B200 systems

In July 2025, SoftBank announced that it had deployed a DGX SuperPOD with more than 4,000 NVIDIA Blackwell GPUs using DGX B200 systems, making its platform the world’s largest DGX SuperPOD based on Blackwell at that time.

Key hardware facts:

The AI computing platform now exceeds 10,000 GPUs in total when earlier generations are included.
Combined performance reaches about 13.7 exaflops in low-precision metrics, a scale suitable for training multi-trillion-parameter models.
The Blackwell generation introduces improved energy efficiency and support for ultra-low-precision formats like FP4, which are vital for large-scale generative AI.

DGX B200 systems integrate Grace CPUs with Blackwell GPUs via high-bandwidth connections, forming nodes where memory and compute are tightly coupled. When hundreds of these nodes are connected through InfiniBand, they form the CHIE-4 cluster.

NVLink and Quantum-2 InfiniBand as the backbone of the AI factory

At this scale, interconnect design is as critical as GPU count. A DGX SuperPOD such as CHIE-4 uses:

NVLink and NVSwitch for high-bandwidth, low-latency communication among GPUs within and across DGX servers
NVIDIA Quantum-2 InfiniBand switches to create a multi-tier network inside the data center

SoftBank and NVIDIA have also described the use of digital twins and detailed cabling design to manage thousands of network connections in their AI-RAN and AI data-center collaboration. ソフトバンク

For users of the SoftBank AI computing platform CHIE-4, this architecture means:

they can scale training jobs across hundreds or even thousands of GPUs with relatively high efficiency, and
they can run many different AI and HPC workloads concurrently without severe network congestion.

Evolution Path: From Ampere and Hopper to CHIE-4

First phase: 2,000 Ampere GPUs and 0.7 exaflops

SoftBank began operating its AI computing platform in September 2023 with approximately 2,000 NVIDIA Ampere GPUs, delivering about 0.7 exaflops of AI performance.

This first phase laid the foundation for:

early Japanese LLM training experiments,
internal AI services for telecom operations, and
SoftBank’s broader vision of AI-driven infrastructure.

Second phase: 4,000 Hopper GPUs and 4.7 exaflops

By late 2024, SoftBank had installed around 4,000 NVIDIA Hopper GPUs, bringing the total to roughly 6,000 GPUs and boosting performance to about 4.7 exaflops.

At this stage, SoftBank publicly positioned its AI computing platform as one of Japan’s top-level AI infrastructures, and signaled that it intended to open the platform to external companies and research institutions as a service.

Third phase: 10,000+ GPUs and 13.7 exaflops-class CHIE-4

The Blackwell upgrade pushed SoftBank AI computing platform CHIE-4 over the 10,000-GPU mark and up to 13.7 exaflops of mixed-precision compute. ソフトバンク

This phased evolution shows a clear strategy:

Start with a large but manageable platform (Ampere).
Upgrade to cutting-edge GPUs (Hopper) and open to external users.
Move to Blackwell at extreme scale, achieving national No.1 AI performance and global top-tier status.

Data Center and Infrastructure Design: Storage, Cooling, and Power

Data lake, parallel file systems, and high-throughput I/O

Training homegrown LLMs and multimodal models requires not only fast GPUs but also a high-throughput data pipeline. Public information about SoftBank’s AI-RAN and AI data-center vision highlights:

the idea of a “Core Brain” data center with a large data lake,
distributed “Regional Brains” for edge and regional processing, and
collaboration with AI-RAN equipment deployed at base stations. ソフトバンク

While detailed storage architecture for CHIE-4 is not fully disclosed, large DGX SuperPOD installations typically combine:

a parallel file system (such as Lustre or Spectrum Scale) for training data and checkpoints, and
object storage for long-term data lake and log retention.

Such a design allows the SoftBank AI computing platform CHIE-4 to sustain high bandwidth for many concurrent training jobs.

High-density racks, liquid cooling, and energy efficiency

Ampere, Hopper, and Blackwell GPUs are all high-power devices. Operating thousands of them requires:

power delivery in the hundreds of megawatts over time, and
advanced cooling, often including liquid cooling.

SoftBank’s plan to build a large AI data center in Tomakomai, Hokkaido—with a site of up to 700,000 square meters and power capacity exceeding 300 MW—shows how seriously it approaches the power and cooling side of AI. ソフトバンクグループ株式会社

By placing major AI data centers in cooler regions and using renewable energy where possible, SoftBank aims to reduce both costs and environmental impact while scaling up SoftBank AI computing platform CHIE-4 and future systems.

Key Workloads: LLMs, Multimodal AI, and AI-RAN

Homegrown Japanese LLMs such as Sarashina

SoftBank subsidiary SB Intuitions is developing Japanese-focused LLMs, including the Sarashina family, as part of a broader “Large Telecom Model (LTM)” strategy for the telecom sector. ソフトバンク+1

These models are trained and served inside SoftBank’s domestic data centers, leveraging:

proprietary network and operations data,
Japanese text corpora, and
CHIE-4-class compute power.

Because SoftBank AI computing platform CHIE-4 is located within Japan and operated by a domestic carrier, it enables end-to-end processing of sensitive data without leaving the country, an important requirement for sovereign AI.

Enterprise generative AI, RAG, and multimodal services

SoftBank positions its AI computing platform as a base for many enterprise-level use cases, such as:

custom LLMs fine-tuned on internal documents,
RAG (retrieval-augmented generation) chatbots connected to knowledge bases,
multimodal models that combine text, images, video, and sensor data.

Companies can choose between:

using SoftBank’s homegrown models (like Sarashina),
leveraging OpenAI models via SoftBank’s partnership, or
training their own models directly on the SoftBank AI computing platform CHIE-4.

This flexibility helps organizations move from simple pilots to large-scale, production-grade AI services.

Telecom optimization and AI-RAN integration

SoftBank is also exploring AI-RAN, where AI accelerators near base stations help optimize wireless networks. In its research topics, SoftBank explains that AI-RAN will collaborate with large AI data centers—its “Brain Data Center”—for tasks such as training massive LLMs that support network intelligence. ソフトバンク+1

In this ecosystem:

AI-RAN nodes handle real-time inference close to users.
SoftBank AI computing platform CHIE-4 and successor systems handle heavy offline training and global optimization.

This shows how CHIE-4 is not just a general-purpose AI cluster, but also a backbone for next-generation telecom infrastructure.

Role in Japan’s National AI Strategy

Data sovereignty and complementarity with global hyperscalers

Japan is investing heavily in AI infrastructure, with policy discussions emphasizing:

secure domestic compute,
support for generative AI, and
reduced dependence on foreign cloud regions for critical data. ソフトバンクグループ株式会社

SoftBank AI computing platform CHIE-4 fits this strategy by:

providing large-scale compute physically located in Japan,
enabling training and inference within domestic legal jurisdiction, and
complementing public cloud services from global providers like AWS, Azure, and Google Cloud.

Organizations can keep high-sensitivity data and workloads on CHIE-4 while using foreign clouds for less sensitive or globally distributed tasks.

Public subsidies, AI data centers, and ABCI 3.0

SoftBank has been selected for large subsidies from Japan’s Ministry of Economy, Trade and Industry (METI) to expand its AI computing infrastructure, indicating that the government views platforms like CHIE-4 as part of national digital strategy.

At the same time, public AI systems such as ABCI 3.0—an open AI infrastructure with over 6,000 H200 GPUs—are also expanding. arXiv

Together, ABCI 3.0 and the SoftBank AI computing platform CHIE-4 provide diverse options for Japanese academia and industry, mixing public and private resources.

Service Model: CHIE-4 as Infrastructure-as-a-Service (IaaS)

Access models for enterprises and academia

SoftBank has indicated that it plans to offer its AI computing platform, including CHIE-4, as an IaaS-style service for companies and research institutions. In practice, this could include:

bare-metal or virtualized GPU clusters,
Kubernetes or container-based environments for MLOps,
managed services for LLM training and inference.

Because the platform is operated by a telecom carrier, it can be integrated with:

closed networks and VPNs,
5G/6G infrastructure, and
on-premise or edge environments.

For many users, this makes SoftBank AI computing platform CHIE-4 a middle ground between hyperscale public clouds and costly on-premise clusters.

Security, compliance, and sovereign-cloud alignment

SoftBank is also involved in sovereign cloud initiatives and works with partners like Oracle to provide compliant cloud environments for government and highly regulated industries. ソフトバンクグループ株式会社+1

CHIE-4 can be seen as a high-performance AI extension of these efforts, offering:

clear data-location guarantees (Japan),
integration with telecom-grade security controls, and
the ability to build end-to-end systems that satisfy both AI performance and regulatory requirements.

Ecosystem: SB Intuitions, SB OAI Japan, and Startups

Collaboration with OpenAI and sovereign AI

SoftBank and OpenAI have announced a joint venture called SB OAI Japan to provide OpenAI-based solutions tailored for the Japanese market. ソフトバンクグループ株式会社+1

In this ecosystem:

CHIE-4 supplies domestic compute power.
SB Intuitions builds homegrown Japanese LLMs like Sarashina.
SB OAI Japan delivers applications and services based on OpenAI models.

This layered approach allows customers to choose between overseas models, local models, or hybrids, while still leveraging the SoftBank AI computing platform CHIE-4 when they need private training or large-scale inference.

Impact on universities and startup innovation

By opening CHIE-4 to external users, SoftBank can support:

universities and public research institutes seeking access to large GPU clusters,
startups that want to experiment at scale without building their own infrastructure, and
joint research projects that combine telecom, robotics, and AI.

Such collaboration could accelerate Japan’s overall AI innovation capacity, much like how national labs and cloud programs support AI ecosystems in the US and Europe.

International Comparison: Fugaku, ABCI 3.0, and Global AI Clusters

CPU-based Fugaku vs GPU-centric CHIE-4

Japan’s famous Fugaku supercomputer, developed by RIKEN and Fujitsu, is based on ARM CPUs and excels at traditional HPC tasks such as simulations and scientific modeling. hpcwire.com+1

SoftBank AI computing platform CHIE-4, by contrast, is:

GPU-centric,
optimized for deep learning and generative AI, and
tightly integrated with telecom and AI-RAN strategies.

Rather than competing, Fugaku and CHIE-4 represent two complementary pillars:

Fugaku for physics-based simulations,
CHIE-4 for data-driven AI and LLMs.

Positioning against overseas hyperscale AI clusters

Globally, companies like Microsoft, Google, Meta, and OpenAI operate AI clusters with tens of thousands of GPUs. CHIE-4 is smaller than the very largest of these, but:

it is the largest DGX B200 SuperPOD with Blackwell GPUs as of mid-2025, and
it ranks in the top tier of AI supercomputers worldwide. hpcwire.com

Most importantly for Japanese users, SoftBank AI computing platform CHIE-4 offers world-class AI performance inside Japan, reducing latency and ensuring domestic data handling.

Business Impact on SoftBank Group and Its AI Strategy

Beyond Carrier: AI data centers as a growth engine

SoftBank Group describes AI data centers as one of its four priority areas, along with AI chips, AI robots, and energy to power them. ソフトバンクグループ株式会社

For SoftBank Corp., the AI computing platform including CHIE-4:

strengthens its position not only as a telecom operator but also as an AI infrastructure provider,
supports internal business units (mobile, broadband, fintech, e-commerce), and
creates new revenue streams from external AI customers.

In this sense, SoftBank AI computing platform CHIE-4 is a strategic asset at the heart of its “Beyond Carrier” growth story.

Synergy with AI chips, robots, and energy investments

SoftBank Group invests in:

AI chip companies and architectures,
robotics and automation, and
large-scale energy and data-center projects.

CHIE-4 is the place where these technologies come together in practice. It provides:

a testbed for new AI hardware and software,
a platform to deploy robotics and AI services at scale, and
a demand driver for renewable energy and advanced cooling technologies.

Technical Challenges and Risks

GPU supply, geopolitics, and cost structure

Building and operating a platform like CHIE-4 requires:

long-term access to cutting-edge GPUs,
large capital expenditures on data centers and energy,
and careful management of geopolitical and export-control risks. hpcwire.com

SoftBank mitigates these challenges by:

partnering closely with NVIDIA, Oracle, and OpenAI,
seeking government support and subsidies, and
diversifying its investments in AI chips and infrastructure.

Operations, orchestration, and human resources

Operating more than 10,000 GPUs is not just a hardware problem. It demands:

sophisticated job schedulers and Kubernetes-based orchestration,
monitoring, fault detection, and automated recovery,
and teams of experts in HPC, networking, and AI engineering.

NVIDIA’s documentation on AI-RAN and SoftBank’s research topics highlight the role of unified orchestrators that manage RAN workloads and AI workloads on shared infrastructure. ソフトバンク+1

For customers using the SoftBank AI computing platform CHIE-4, this operational maturity translates into higher reliability and easier onboarding, but it remains an area of continuous improvement.

Best Practices for Using SoftBank AI Computing Platform CHIE-4

Designing LLM and generative AI workloads for CHIE-4

To fully exploit CHIE-4, users should:

design models with data, tensor, and pipeline parallelism so they can scale across hundreds of GPUs,
adopt mixed-precision training (FP16 or FP8) to maximize throughput,
carefully plan checkpointing and I/O patterns to avoid storage bottlenecks, and
use distributed training frameworks that support DGX SuperPOD environments.

By aligning workload design with the architecture of the SoftBank AI computing platform CHIE-4, researchers and companies can reduce training time and cost.

From PoC to production: a practical roadmap

A realistic path for enterprises might look like:

PoC on small clusters or public cloud GPUs, validating business value.
Scaling experiments on CHIE-4 for larger models or datasets.
Integration with internal systems—identity, logging, monitoring, and data governance.
Production deployment with SLAs, rollback strategies, and regular evaluation of bias and safety.

Throughout this journey, organizations must treat SoftBank AI computing platform CHIE-4 not just as raw compute, but as part of a larger MLOps and governance framework.

FAQ: Common Questions About CHIE-4

Q1. What exactly is SoftBank AI computing platform CHIE-4?

It is SoftBank Corp.’s large-scale GPU supercomputer built with NVIDIA DGX SuperPOD and DGX B200 systems using Blackwell GPUs. The system is nicknamed CHIE-4 and has achieved Japan’s No.1 AI performance ranking in HPL-MxP at SC2025. ソフトバンク

Q2. Can external companies or universities use CHIE-4?

SoftBank has stated that it intends to offer its AI computing platform as a service to enterprises and research institutions, effectively turning CHIE-4 into an IaaS-style resource. Detailed terms depend on future commercial offerings, but the direction is clear: CHIE-4 is moving from internal system to shared national infrastructure.

Q3. How is CHIE-4 different from global cloud GPU services?

Global clouds provide massive scale and global reach, but many regions are outside Japan. CHIE-4 offers:

domestic data location,
tight integration with SoftBank’s networks, and
direct access to homegrown LLMs and sovereign AI initiatives.

For workloads with strict data-sovereignty needs, SoftBank AI computing platform CHIE-4 can be a strong alternative or complement to public clouds.

Q4. How does CHIE-4 compare with Fugaku or ABCI 3.0?

Fugaku is a CPU-based supercomputer optimized for traditional HPC simulations. ABCI 3.0 is a public AI infrastructure with H200 GPUs. CHIE-4 is a private-sector AI supercomputer using Blackwell GPUs and tightly integrated with telecom and enterprise services. Together, they form a diverse ecosystem of Japanese high-performance computing resources. arXiv+1

Q5. What kinds of AI models are being trained on CHIE-4?

SoftBank and SB Intuitions use the platform to develop Japanese-language LLMs such as Sarashina, as well as telecom-specific models like the Large Telecom Model (LTM). These models power internal tools and may also be offered as services to external customers. ソフトバンク+1

Q6. Is CHIE-4 environmentally sustainable?

Operating CHIE-4 consumes substantial power, but SoftBank is working on large AI data centers in locations like Tomakomai, Hokkaido, which are designed to use efficient cooling and renewable energy where possible. The company also positions energy infrastructure as one of its strategic investment areas to support AI data centers sustainably. ソフトバンクグループ株式会社

Conclusion: Strategic Meaning of SoftBank AI Computing Platform CHIE-4

SoftBank AI computing platform CHIE-4 is more than an internal project or a benchmark-winning machine. It represents:

Japan’s No.1 AI performance system, with world-class rankings on HPL-MxP, TOP500, and HPCG. ソフトバンク+1
a multi-generation evolution from Ampere to Hopper to Blackwell, culminating in more than 10,000 GPUs and 13.7 exaflops-class compute.
a strategic hub for sovereign AI, telecom innovation, and enterprise generative AI in Japan. ソフトバンク+1

For researchers and enterprises, understanding how to leverage the SoftBank AI computing platform CHIE-4—technically, economically, and strategically—will be crucial in the coming years. As global AI competition intensifies, CHIE-4 offers Japan a powerful, domestically controlled engine for innovation.