【Background】
In recent years, demand for large-scale computing resources capable of advanced computational processing has been growing rapidly, driven by corporate development of proprietary LLMs (large language models) and the rapid evolution of AI and robotics in industrial sectors. To meet requirements such as the expansion of AI models, real-time inference, and support for diverse generation formats, cloud infrastructure must continuously evolve as a "dynamic technology foundation," with regular updates being key to technological innovation.
GMO Internet received certification from the Ministry of Economy, Trade and Industry on April 15, 2024, for its supply security plan for "cloud programs," a specified critical material under the Economic Security Promotion Act (*1). Since November 22 of the same year, the company has been providing NVIDIA Hopper GPUs at industry-leading speeds in Japan (*2). In response to market conditions and technical requirements, we reviewed our additional investment plan announced on May 14, 2025, and decided to deploy the latest generation of the Blackwell architecture, "HGX B300."
This enables "GMO GPU Cloud" to support the high computational performance, scalability, and low-latency communication required for future AI development and operations, evolving as an even more advanced infrastructure platform.
(*1) Ministry of Economy, Trade and Industry: Ensuring Stable Supply of Specified Critical Materials under the Economic Security Promotion Act https://www.meti.go.jp/policy/economy/economic_security/cloud/index.html
(*2) Press Release: "GMO Internet Group Launches 'GMO GPU Cloud' Ranked in TOP500 Supercomputer List" https://internet.gmo/en/news/article/27/
【About GMO GPU Cloud (NVIDIA B300)】
1. GMO GPU Cloud Overview
| Service Name | GMO GPU Cloud Bare Metal Plan |
|---|---|
| Launch Date | December 16, 2025 |
| GPU Configuration | NVIDIA Blackwell Ultra GPU |
| GPU Memory | 2.3 TB |
| CPU Configuration | Intel Xeon 6767P x2 |
| Main Memory | 3 TB |
| Local Storage | 3.84TB NVMe Gen5 SSD x8 |
| Network Storage | NFS File Storage and Object Storage (Coming Soon) |
| Global/Local Storage Network (Shared) | NVIDIA BlueField-3 DPU 200Gbps x2 |
| Interconnect | NVIDIA ConnectX-8 SuperNIC x8 合計 6,400Gbps |
| Use Cases | - High-speed training and fine-tuning of large language models - Accelerated AI reasoning and inference processing ・- Training computer vision models with large-scale datasets- |
| Pricing | Custom quotation per company |
2. NVIDIA Blackwell Ultra GPU (GPU Memory) Features
Specifications (vs. NVIDIA H200)
・Memory capacity: 288GB (204% of H200)
・Newly supports FP4 computation precision
・FP8 Tensor Core performance improved up to 2.25x
3. NVIDIA HGX B300 Features
Enhanced Efficiency for Large-Scale Training
• Training performance improved up to 4x (vs. previous generation H200)
• Key performance metrics including GPU memory capacity and bandwidth significantly enhanced
• Improved performance per watt
Strengthened Inference Capabilities
• Inference performance improved 11x (vs. previous generation H200)
• FP4 support reduces VRAM usage while improving throughput
• Largest GPU memory capacity to date
4. Network Configuration
NVIDIA HGX B300 AI infrastructure is provided with ultra-high-speed communication with up to 6 400Gbps NVIDIA ConnectX-8 SuperNICs and NVIDIA Spectrum-X Ethernet switches that connect and scale multiple GPUs. Its low-latency design accelerates data transfer during large-scale AI training and delivers stable high performance even in distributed learning environments combining multiple servers.
(Left) NVIDIA ConnectX-8 SuperNIC
(Center) NVIDIA BlueField-3 DPU
(Right) NVIDIA Spectrum-4
5. Comparison between NVIDIA B300 and NVIDIA H200
| NVIDIA HGX B300 | NVIDIA HGX H200 | |
|---|---|---|
| GPU Memory | Total 2.3TB (288GB HBM3e ×8) | Total 1.1TB (141GB HBM3e ×8) |
| FP4 Tensor Cores | 144PFLOPS / 108PFLOPS | - |
| FP8 Tensor Cores | 72PFLOPS | 32PFLOPS |
| FP16 Tensor Cores | 36PFLOPS | 16PFLOPS |
| FP32 | 600TFLOPS | 540TFLOPS |
| Maximum Interconnect Bandwidth | 6,400Gbps | 3,200Gbps |
| NVLINK (Inter-GPU Bandwidth / per GPU) | 5th Generation (1,800GB/s) | 4th Generation (900GB/s) |
【About the Bare Metal Plan】
The "GMO GPU Cloud Bare Metal Plan," launching as Japan's first offering today, provides a fully dedicated environment that maximizes computational performance by accessing GPUs directly without a virtualization layer. It leverages computational capabilities specialized for AI inference processing and is suitable for diverse workloads including high-speed inference of large language models and agent AI development.
The service also maintains the flexibility inherent to bare metal while incorporating mechanisms to reduce the effort required for initial setup and network construction.
■Key Features:
● Provided with Ubuntu OS pre-configured
● Pre-configured interconnect network (local network also available)
● Server management via management dashboard
● Additional storage available
This enables users to utilize a GPU environment that balances high performance with flexible tuning capabilities while minimizing deployment effort.
Server Configuration Image for NVIDIA B300
Comment from Tadashi Ito, Representative Director, President and CEO
Since the launch of GMO GPU Cloud in November 2024, we have been supporting infrastructure for advanced AI development. We have consistently maintained our commitment to cutting-edge, high-performance technology, stable operational environments, and comprehensive technical support, continuously working to support our customers' AI development initiatives. Being the first in Japan to provide NVIDIA B300 represents a significant step in accelerating Japan-originated AI innovation to the next level. Going forward, we will strengthen our technical collaboration with NVIDIA, enhance our AI development platform to support engineers' endeavors, and robustly promote the creation of globally competitive next-generation AI.
Comment from Masataka Osaki, Japan Country Manager, VP Worldwide Field Operations, NVIDIA
"For developers and researchers tackling AI development daily, access to vast computational resources through high-performance and scalable computing environments is essential. The launch of 'GMO GPU Cloud' equipped with NVIDIA HGX B300, which enables highly efficient training and high-speed inference, represents a significant step in leading Japan's AI infrastructure..."
【Future Development】
GMO Internet will contribute to technological innovation in the rapidly evolving AI and robotics fields through its AI infrastructure strategy centered on "GMO GPU Cloud." By continuing to provide the latest AI computing infrastructure and constructing flexible cloud environments tailored to customer needs, we will serve as an essential domestic AI infrastructure for Japan's AI industry, contributing to AI innovation creation for society and industry.
【Background on GMO GPU Cloud
(URL: https://gpucloud.gmo/)】
(URL: https://gpucloud.gmo/)】
"GMO GPU Cloud," Japan's fastest-class GPU cloud service built on NVIDIA technology, launched services featuring high-performance "H200 GPUs" in November 2024. The "NVIDIA B300" instance of GMO GPU Cloud, which now incorporates the latest NVIDIA Blackwell Ultra architecture, is a GPU designed for building ultra-large-scale AI factories, equipped with 5th generation Tensor Cores supporting "NVFP4" (*3), HBM3e memory (*4) reaching 2.1TB, and high-speed communication via 5th generation NVIDIA NVLink and NVLink Switch. This significantly reduces training time for large language models compared to the previous H200, substantially improving AI development efficiency.
Through this service, GMO Internet provides companies and research institutions working in the generative AI field with optimized infrastructure and flexible, customizable computing environments tailored to customer workloads, contributing to shortened development periods and cost reduction while promoting the development of Japan's AI industry.
(*3) "NVFP4": a 4-bit format purpose-built to deliver exceptional inference latency, throughput, and efficiency—all while maintaining production-grade accuracy.
(*4) "HBM3e": The latest high-bandwidth memory standard with improved bandwidth compared to previous versions, accelerating training and inference processing for large-scale AI models.
About GMO Internet, Inc.
GMO Internet, Inc. launched with a new structure on January 1, 2025, to integrate the strengths of GMO Internet Group's Internet Infrastructure business and Advertising & Media business.
Maximizing the solid revenue foundation of the Internet Infrastructure business and the respective strengths of the Internet Advertising & Media business, under our corporate tagline "Internet for Everyone," we deliver "smiles" and "inspiration" to all stakeholders and challenge ourselves to create value that shapes a new future with AI.
