The Japan Atomic Energy Agency (JAEA) is among the partners in the project through which Preferred Networks (PFN) and Tenstorrent will develop artificial intelligence processors based on integrated photonics and electronics. NEDO, Japan’s New Energy and Industrial Technology Development Organization, announced the selection on October 20, 2026.
The program runs for five years and falls under NEDO’s initiative on post-5G communication infrastructure and advanced semiconductor manufacturing technologies. The goal is to build a computing platform, called xPU, that links multiple specialized chiplets through photonics and advanced packaging. PFN will lead the development of the AI chiplets, featuring three-dimensionally stacked memory and very high bandwidth. Tenstorrent will handle the CPU chiplets built on an open architecture, with confidential computing capabilities.
JAEA will work as a subcontractor alongside the University of Tokyo, Kobe University, Toyo University, the Japan Agency for Marine-Earth Science and Technology and Tenstorrent Japan. This group of research institutions will focus on domain-specific applications for scientific AI and physical AI. The press release does not specify the tasks assigned to JAEA. The involvement of the agency and the marine science institute suggests that workloads typical of nuclear and geophysical research will serve as a testbed for the new processors.
The backdrop is the surge in demand for computing power to train and run AI models. PFN and Tenstorrent are aiming for system-level integration that delivers both performance and efficiency, in training as well as inference. The energy consumption of data centers is one of the main constraints on this expansion, and integrated photonics is meant to reduce it by moving data with light rather than with electrical signals alone.
For the nuclear sector, the project is of indirect interest. Reactor simulations, materials analysis and fusion studies all require very powerful machines, and more efficient hardware lowers the costs and energy use of these activities. After the research phase, the two companies intend to bring the results to market for applications such as AI agents.
If the chiplets perform as expected, JAEA could count on a platform designed with its own simulation codes in mind, developed in Japan and drawing on domestic expertise.
