The term "generative AI" has become something we hear in our daily lives. Generative AI, including ChatGPT and Gemini, is being utilized in various situations such as writing and summarizing text, information retrieval, image generation, and program creation. In companies, expectations for generative AI are rising for purposes like improving operational efficiency, utilizing information, and creating new services. On the other hand, because generative AI has become more accessible, many people may want to clarify the basics: "What exactly is generative AI?" "What is different from previous AI?" "How can it be specifically utilized in companies?" "What kind of environment is necessary to fully leverage generative AI?" This time, as the use of generative AI in businesses is progressing, we will revisit and organize our understanding of generative AI, its potential, and the computing environment that supports it.

While the use of generative AI is expanding, manufacturing industries and research institutions are facing challenges such as: - Inability to upload confidential data and design information to the cloud - Desired AI models not fitting into GPU memory - Desire to reduce cloud usage costs and waiting times In particular, large-scale language models (LLMs) and vision-language models (VLMs), which process images and text together, are becoming larger. It is not uncommon for general 48GB or 96GB class GPUs to be unable to load the entire model. A new option to address these challenges is the "MSI XpertStation WS300," equipped with 748GB of coherent memory. This article will clearly explain the features of the WS300 and what types of AI development it is suitable for.

One of the current demands for AI is "physical AI," and I believe there are increasing opportunities to hear related terms such as "Omniverse" and "Isaac Sim." This time, I would like to unravel the current role and position of "Omniverse."

We exhibited at the "Robotics and Mechatronics Conference" held at the Hakata International Exhibition Hall & Conference Center from June 30 to July 1, 2026. Thank you very much to everyone who took the time to visit our booth despite your busy schedules! Today, we will provide a report on the atmosphere of the booth and the exhibits.

In the research and development (R&D) department of the manufacturing industry, the demand for computational resources has reached unprecedented levels. In addition to traditional fluid analysis (CFD), structural analysis, and materials informatics, there has been a surge in the need to build "secure local LLMs (RAG)" trained on internal technical standards and confidential data. However, attempting to execute these advanced processes on-premises requires SXM-based ultra-high-end servers, leading to practical barriers such as "data center power and cooling facilities (like water cooling) not keeping up" and "installation costs being too high." Currently, the "NVIDIA H200 NVL," which is limited to a TDP of up to 600W (air-cooled/PCIe-based) and is easier to implement in existing server room environments, is attracting significant attention from R&D departments. This article will explain the overwhelming advantages gained by bundling four units together using "4-Way NVLink," rather than viewing the H200 NVL merely as an expansion card, as well as the "unexpected pitfalls" that field engineers are likely to encounter during implementation.