Japanese computing and telecom company Fujitsu is aiming to develop an AI inference device that combines its own neural processing units (NPUs) with central processing units (CPUs), a high-stakes bid to forge greater technological independence and national economic security.
According to Japanese media reports, the new device will be produced by Japan’s IC foundry Rapidus at its second fab in Hokkaido using a 1.4-nm process developed in partnership with IBM – by the end of the decade, if everything goes according to plan.
A large part of the development costs is likely to be picked up by Japan’s New Energy and Industrial Technology Development Organization (NEDO), which operates under the auspices of the Ministry of Economy, Trade and Industry.
Significantly, Fujitsu is not trying to compete with the graphics processing units (GPUs) pioneered and dominated by Nvidia, which are used for data-intensive AI training.
NPUs, as explained by technology analyst and entrepreneur Vaclav Vincalek in his Recurrent Patterns blog, “are designed to mimic the human brain’s neural structure and mimic synaptic transmission. What makes them different from the other chips?”
“First, the power consumption is 100x – 1000x less than your traditional GPU. While we are talking about gigawatts of energy for data centers, these chips consume energy in low teens or even single digit watts. As you can imagine that’s a major must for any mobile application.”
“Secondly, these chips can learn without explicit reprogramming and adapt to real-world scenarios. They can be used in real-time sensory data processing or pattern recognition. Anytime you hear the word robot, autonomous system or IoT device, the chances are that it will contain one of these chips.”
Intel, IBM and BrainChip – headquartered in Australia, with engineering done primarily in California – are currently the leading developers of NPUs. Both GPUs and NPUs work together with CPUs, which manage the flow of data and instructions in a computer system.
The Fujitsu-Monaka CPU is an Arm-based processor developed for use in data centers and Japan’s FugakuNEXT supercomputer. Now fabricated by TSMC using a 2-nm process, it will be produced by Rapidus when its first fab comes online in 2027. Both devices, the NPU and the CPU, are designed to maximize energy efficiency and reduce the power consumption of AI data centers.
In March, Fujitsu started manufacturing AI servers at its Kasashima factory in Ishikawa Prefecture to support government-designated critical industries including energy, telecommunications, logistics, healthcare, aerospace and defense. These servers are equipped with Nvidia Blackwell GPUs and Fujitsu-Monaka CPUs.
Production at Kasashima will be facilitated by Fujitsu’s collaboration with California-based IT solutions provider Supermicro, which contributes its development, manufacturing, sales and maintenance expertise. The AI servers will be sold in Europe as well as Japan, a reminder of the strategic ties between the two.
Arm is a British integrated circuit design company owned by Japan’s Softbank Group. Fujitsu-Monaka was developed with the support of NEDO. Rapidus and IBM announced a strategic partnership in December 2022. IBM and Fujitsu are both shareholders in Rapidus.
Earlier this year, it was reported that IBM has developed a new and vastly more efficient heat-modeling technique with electronic design automation (EDA) company Synopsis and the US Defense Advanced Research Projects Agency (DARPA) that will facilitate production at 2-nm and “be a direct requirement for performance at 1.4-nm,” according to Russ Robison, a senior data analysis engineer at IBM Research. As circuit features shrink, heat management becomes a critical issue.
To speed up heat calculations at ever-smaller process nodes, IBM and Ansys, a subsidiary of Synopsis, developed a new machine learning tool with support from DARPA. Ansys, which was acquired by Synopsis in July 2025, specializes in engineering simulation software. IBM can be expected to share this technology with Rapidus.
Last December, Fujitsu and Riken decided to invest and participate in the efforts of Saimemory, a subsidiary of SoftBank Corp, to develop a more efficient alternative to the high-bandwidth memory used with the GPUs designed by Nvidia and its nearest competitor, AMD.
Riken (literally the Institute of Physical and Chemical Research) is Japan’s leading scientific research institute. SoftBank Corp is part of the SoftBank Group. Taiwanese IC foundry Powerchip Semiconductor Manufacturing Corporation (PSMC), Japanese semiconductor packaging manufacturer Shinko Electric and the University of Tokyo also work with Saimemory.
In February, Saimemory and Intel agreed to collaborate on the development of Z-Angle Memory (ZAM), a high-capacity, high-bandwidth and low-power-consumption semiconductor technology.
They aim to develop memory chips with twice or three times the capacity and half the power consumption of the high-bandwidth memory produced by SK Hynix, Samsung Electronics and Micron Technology. The plan is to create prototypes by the end of the fiscal year to March 2028 and commence mass production of chips for use in AI data centers by March 2030.
According to SoftBank, this project “will leverage the next-generation memory foundational technologies and technical expertise validated by Intel’s Next Generation DRAM Bonding (NGDB) initiative that was completed under the Advanced Memory Technology (AMT) program managed by the US Department of Energy and National Nuclear Security Administration through the Sandia National Laboratory, Lawrence Livermore National Laboratory and Los Alamos National Laboratory.”
Japan’s FugakuNEXT supercomputer is also scheduled to be deployed around the end of the decade, in Kobe. Led by the Riken Center for Computational Science, the project will combine Nvidia’s GPUs and Fujitsu’s CPUs, and leverage Nvidia’s global market presence, Fujitsu’s system integration capability and Riken’s software and algorithm technologies.
Nvidia notes that “the contract enables the partners to work side by side in shaping the system’s architecture to address Japan’s most critical research priorities.”
These include:
- Scientific research: Accelerating simulations with surrogate models and physics-informed neural networks;
- Manufacturing: Using AI to learn from simulations to generate efficient and aesthetically pleasing designs faster than ever before; and
- Earth systems modeling: Aiding disaster preparedness and prediction for earthquakes and severe weather.
Riken also collaborates with the Japan Aerospace Exploration Agency (JAXA) on space science and the development of satellite technology, while its advanced materials, quantum computing and other research activities have “dual‑use” civilian and military potential.
Last October, Nvidia CEO Jensen Huang told reporters in Tokyo: “We are creating Japan’s AI infrastructure – from silicon to systems to AI models and software – designed with Japan for Japan’s industries and society.”
Satoshi Matsuoka, director of the Riken Center for Computational Science, said: “Partnering with the American company that makes the world’s best GPUs is a major strategic move. It will also lead to the global adoption of Japanese technology, including Fujitsu’s CPU, which is aiming to be the best in the world.”
The buzzword used by Nvidia, Fujitsu, SoftBank and others is “sovereign AI,” but there is no daylight between the Japanese and American sovereigns.
What matters is that Nvidia chose to work with the Japanese on what they hope will be the world’s top-performing supercomputer, and that the Japanese are becoming more deeply involved in the US-led semiconductor ecosystem, regardless of political vicissitudes.
Follow this writer on X: @ScottFo83517667