Practical Performance Confirmed for Remote Distributed AI Infrastructure Between Tokyo and Fukuoka Utilizing 'IOWN APN'

GMO Internet, Inc. (Headquarters: Shibuya-ku, Tokyo; President and CEO: Masatoshi Kumagai; hereinafter "GMO Internet"), NTT East Corporation (Headquarters: Shinjuku-ku, Tokyo; President: Naoki Shibuya; hereinafter "NTT East"), NTT West Corporation (Headquarters: Osaka-shi, Osaka; President: Ryota Kitamura; hereinafter "NTT West"), and QTnet, Inc. (Headquarters: Fukuoka-shi, Fukuoka; President: Yoshio Ogura; hereinafter "QTnet") have completed a technical demonstration of a remote distributed AI infrastructure between Tokyo and Fukuoka utilizing the "APN (All-Photonics Network)" of the "IOWN (Innovative Optical and Wireless Network)."

In this demonstration, conducted from November 2025 to February 2026, a dedicated IOWN APN line was established between Tokyo (storage) and Fukuoka (GPU) to measure and evaluate AI workload performance on an AI development platform connecting GPUs from "GMO GPU Cloud" with high-capacity storage. The results confirmed that for Large Language Model (LLM) training, performance degradation was limited to only about 0.5% compared to a local environment, indicating that the impact is extremely minimal. For image classification tasks involving data loading, it was confirmed that processing at a practical level is possible even in a remote environment through methods such as learning data optimization. This demonstrates that practical AI development in a remote distributed environment is achievable through design tailored to workload characteristics.

Prior to this demonstration, the four companies conducted a preliminary test (Phase 1) in July 2025 using a simulated remote environment assuming a distance of approximately 1,000 km between Tokyo and Fukuoka, and have published the details in a technical report.

Press Release: https://www.ntt-west.co.jp/news/2510/251002a.html Technical Report: https://www.ntt-west.co.jp/news/2510/251002a_1.html

Building on the results of this demonstration, the four companies will continue to advance efforts toward the practical implementation of remote distributed AI infrastructure tailored to customer needs.

[Background and Purpose] With the recent spread of generative AI and Large Language Models (LLMs), the demand for AI development infrastructure is expanding rapidly. Conventionally, it has been considered essential for GPUs and high-capacity storage to be physically adjacent. However, to address data center space constraints and the need to manage data within one's own facilities, there is a demand for a distributed AI development platform that transcends geographical limitations. The four companies have been examining the technical feasibility of connecting remote GPUs and storage using the high-speed, high-capacity, and low-latency characteristics of IOWN APN.

[Overview and Results of Preliminary Demonstration (Phase 1)] In July 2025, a delay adjustment device, "OTN Anywhere," was installed in a Fukuoka data center, and two test tasks—image recognition (ResNet) and language learning (Llama2 70B)—were executed using GMO GPU Cloud. Under simulated delay conditions equivalent to the distance between Tokyo and Fukuoka (15 milliseconds), the decline in ResNet benchmark scores was confirmed to be approximately 12%. This was judged to be within a commercially viable range, leading to the current demonstration.

[Overview and Results of This Demonstration (Phase 2)] In this demonstration, the second headquarters of GMO Internet Group (Shibuya-ku, Tokyo) and the data center of QTnet (Fukuoka-shi, Fukuoka) were connected via IOWN APN (100GbE) as the actual inter-site network. A GPU server, "NVIDIA HGX H100," was placed on the Fukuoka side, and high-speed storage, "DDN AI400X2," was placed on the Shibuya side to measure AI training performance when using remote storage.

- Demonstration Period: November 2025 – February 2026 - Connection Section: Shibuya-ku, Tokyo (GMO Internet) – Fukuoka-shi, Fukuoka (QTnet) - Demonstration Content: Measurement of training time for image classification tasks (ResNet) and Large Language Model processing tasks (Llama2 70B)

[Results of the Demonstration] The results of the demonstration confirmed that even in a remote distributed environment via IOWN APN, performance comparable to a local environment (connection within the same data center) can be achieved.

Large Language Model (Llama2 70B) Training Task - Local Environment: 24.87 minutes - Remote Environment (via IOWN APN): 24.99 minutes - It was demonstrated that for LLM training, which is primarily computational, the impact of latency is extremely limited (a difference of approximately 0.5%).

Image Classification (ResNet) Task - Local Environment: 13.72 minutes - Remote Environment (via IOWN APN): 14.38 minutes - It was confirmed that even for tasks involving data loading, processing at a practical level is possible in a remote environment by performing appropriate data formatting.

*The results of this verification have not been officially verified or approved by the MLCommons Association.

For details, please refer to the following attachment: Details and Results of GPU-Storage Connection Performance Test in Remote Distributed AI Infrastructure Utilizing 'IOWN APN' URL: https://www.ntt-west.co.jp/news/2603/260330b_1.html

[Transformation Brought About by This Demonstration] The success of this demonstration marks a major turning point in solving the challenge of "separation of computing resources and data" caused by physical distance. Conventionally, it has been common to transfer and replicate data required for AI training to a cloud provider's data center. However, the model demonstrated here—where "data is not moved, and computing resources access the data remotely"—presents a new option for fields with strict data sovereignty and security requirements. This is expected to enable reductions in data transfer time and costs, the elimination of redundant management, and an expansion of options for computing resources by combining on-premises and cloud environments. In particular, this model, which allows the use of domestic cloud GPU resources while keeping data under the management of one's own facilities and organization, is expected to contribute significantly to the realization of "sovereign clouds" in fields with strict internal controls and cross-border data regulations, such as finance, healthcare, defense, and government.

[Expected Use Cases] With the practical application of the technology confirmed in this demonstration, the following uses are expected. Please note that in actual application, performance may vary depending on individual conditions such as the distance between the GPU and storage and the network configuration, so the feasibility must be examined for each use case.

- AI training while maintaining large-scale or confidential data: Executing AI model training on remote cloud GPUs without storing data managed by the company externally. - Hybrid use with existing on-premises environments: Building a flexible AI development environment by procuring insufficient GPU resources from the cloud while utilizing the company's existing storage and GPU resources. - BCP support through regional distributed placement: Building a high-availability environment that ensures AI processing continuity even in the event of disasters or failures by geographically distributing computing resources and storage.

This demonstration shows the path for IOWN APN to evolve not just as a communication line, but as social infrastructure supporting AI and cloud platforms. The four companies will continue to promote the spread of IOWN APN (NTT East and NTT West's "All-Photonics Connect powered by IOWN") and collaborate with cloud service providers such as "GMO GPU Cloud" and regional data centers like QTnet, aiming for the social implementation of IOWN APN as the backbone of AI infrastructure.

*The information and demonstration results described in this press release are current as of the date of the announcement. This content was obtained under a specific verification environment and does not guarantee equivalent performance or results in any environment.