As companies move from experimenting with AI tools to using them in their daily operations, the fundamental question is no longer which model to choose, but where the processing will actually take place. 日本HP’s experience in Japan offers a practical framework for weighing cloud services against on-premises systems, while clearly warning against treating on-premises operation as an end in itself.
This perspective came from a presentation by 裕史勝谷, Director of AI and Data Science Market Development at 日本HP, during the HP Partner Communication 2026 conference held in Tokyo on July 16, 2026, as well as from an individual interview conducted with him by ITmedia. 勝谷 says that many of the companies contacting him have already begun using AI, but are now looking for ways to expand its benefits across their operations.
From General Experimentation to Enterprise Operations
勝谷 suggests that companies progress from using chat and summarization, to connecting models with internal documents, and then to automating tasks such as preparing quotations and proposals through AI agents. In his presentation, he divided the roles of what he called “AI employees” into three categories: an assistant responsible for meeting minutes and document preparation, a specialized employee supporting sales or design work, and an analytical role that helps present data and support decisions.
This concept is not limited to software. 日本HP presented a combination of AI PCs and training, workstations for testing large language models locally, and operating environments based on multiple graphics processing units for corporate departments. The idea is to align the infrastructure with the stage of use, rather than selling a device or service separately from the process in which it will be used.
Data Sensitivity Is the Starting Point
勝谷 uses cloud AI services when searching publicly available information, but says he runs processing locally when the data concerns work that cannot leave the company. However, this does not mean that every task containing internal data should automatically be moved to a local device; companies should test the local environment’s ability to perform the required task and compare it with the cloud’s speed and flexibility.
The experience highlights an important distinction between where an application is used and where inference is performed. According to the article, OpenAI’s Codex application runs on a computer for some aspects of automation and programming, but the inference itself takes place in the cloud. Therefore, IT teams must verify where the data is sent and where processing is performed, rather than relying solely on the application interface being present on a computer inside the company.
日本HP is also testing an open-source AI agent, OpenClaw, in a local environment, but 勝谷 noted that processing occasionally stopped in his environment, and that he still chooses Codex in some cases. This experience supports a practical position: access to fully local operation is not guaranteed merely because the hardware or model is available.
Device Selection Changes with the Scale of Use
After determining the need for local operation, the scale of use must be established. 勝谷 uses the HP ZGX Nano G1n AI Station to conduct initial tests and verify whether customer requirements can be implemented, describing it as HP’s version of the NVIDIA DGX Spark. Its small size also allows it to be taken to a location such as a production line for testing there.
This type of device is suitable for research and development, use by a single person, or an initial trial with a limited budget. Running a task for several people within a single department may require a larger workstation or a shared environment. A model’s success in an individual test does not prove that its performance will be comfortable when used by an entire group. The number of users, acceptable waiting time, and processing frequency are factors that must be determined before selecting the hardware.
The decision also includes where to place the devices. They may initially be located close to users, then moved to a server room as their number increases. With agents that repeatedly plan, perform inference, execute tools, and review results, it is not enough to look only at the graphics processing unit; main memory, SSD storage, and the central processing unit may all become bottlenecks.
Cost Is Not the Cloud Price Versus the Device Price
勝谷 believes that the cloud may be the right choice when the focus is on short-term cost, while an on-premises environment gives some companies greater continuity, particularly when they face problems with the enterprise network’s proxy server or changes in prices and usage limits. Some customers may choose to share a workstation equipped with a graphics processing unit among several users to reduce token costs.
However, an accurate comparison must include electricity, maintenance, management, and the initial investment in an on-premises environment, in addition to contract terms and the scale of cloud usage. It is also necessary to ask whether both sides can perform the same processing with the same quality and repeatability, and what effect a service or network outage would have on the work.
The Human Factor May Be the Bottleneck
Even after the model and device are provided, the process may stop at a human approval required to execute an action involving permissions. 勝谷 acknowledges that his own workload may make the person a bottleneck. Therefore, introducing agents requires defining the approval stages and who is responsible for them, not merely connecting the agent to the tools.
The practical conclusion from 日本HP’s experience is that selecting an AI operating environment should begin with the data, followed by the task, the number of users, cost, continuity, and the capabilities of the operations team. The selected model and the reason for choosing it should also be documented so that knowledge can be transferred when those responsible change. These guidelines remain based on the experience of an official at 日本HP and are not a general standard proving that the cloud or on-premises operation is better in every case.